<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Stefano Besana on Medium]]></title>
        <description><![CDATA[Stories by Stefano Besana on Medium]]></description>
        <link>https://medium.com/@stefanobesana?source=rss-99ac926e66e4------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*X9pzI3vi73xKXBa-TBzHPw.jpeg</url>
            <title>Stories by Stefano Besana on Medium</title>
            <link>https://medium.com/@stefanobesana?source=rss-99ac926e66e4------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 18:31:04 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@stefanobesana/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[AI Is Moving Faster. Is the Enterprise Changing?]]></title>
            <link>https://medium.com/@stefanobesana/ai-is-moving-faster-is-the-enterprise-changing-08d51561c241?source=rss-99ac926e66e4------2</link>
            <guid isPermaLink="false">https://medium.com/p/08d51561c241</guid>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[organization-design]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[Stefano Besana]]></dc:creator>
            <pubDate>Mon, 02 Nov 2026 08:01:02 GMT</pubDate>
            <atom:updated>2026-11-02T08:01:02.365Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Deloitte’s latest State of AI in the Enterprise report raises a question that goes beyond adoption: what happens to work, judgment and responsibility when organizations give AI a growing role in how they operate?</em></p><p>An AI system can produce an answer in minutes rather than days. Yet that answer may still wait in an approval queue, cross several organizational boundaries, and reach someone who lacks the authority to act on it. The resulting delay reveals something about the design of the work around the technology, including the decisions people are permitted to make.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CKI8qIcqpNkXhMt6V69H2A.png" /></figure><p>As <a href="https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/content/state-of-ai-in-the-enterprise.html">Deloitte’s October 2026 <em>State of AI in the Enterprise</em> </a>puts it:</p><blockquote><em>“The constraint is no longer whether AI can accelerate work. It’s whether the organization is prepared to redesign the work it accelerates.”</em></blockquote><p>Based on a survey of 3,500 senior leaders in 25 countries — all from organizations already using AI in daily operations — and supplemented by 15 executive interviews, the report examines 7 dimensions of enterprise AI: agility, economics, returns, work redesign, governance, physical AI and sovereign AI. Together, they describe organizations gaining new capabilities while still working out how to accommodate them. The most interesting questions now concern the decisions these capabilities make possible, their cost and the people who will be responsible for their consequences.</p><h4>Faster decisions within familiar structures</h4><p><strong>81% of surveyed organizations say AI has improved the speed of decision-making and execution. 25% say they are using it for deep transformation, which the report defines as new products or services, reengineered core processes, or a fundamentally reinvented business model.</strong></p><p>Faster analysis and less administrative work can be valuable in their own right. They may also expose constraints that were previously harder to see. A customer-service team, for example, might identify the cause of a problem almost immediately while still needing several approvals to resolve it. At that point, improving the analysis further will do little for the customer. The organization has learned enough to act; the remaining delay lies in its handoffs and decision rights.</p><p><strong>This distinction matters when judging progress. Improvements within an existing process deserve recognition, but their value should be assessed on their own terms. </strong>A company that wants a different customer experience or a new business model will eventually have to reconsider how work moves across roles, teams, and functions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*O2x1OsD51NFBxh9rn6IMnw.png" /><figcaption>AI Agility</figcaption></figure><h4>Understanding the economics of AI-enabled work</h4><p>The report finds that <strong>71%</strong> of respondents’ organizations track total AI inference and API spending. <strong>51%</strong> track token volume, and <strong>44%</strong> track per-unit inference cost. That difference leaves a practical gap between knowing the overall expense and understanding which activities generated it.</p><p><strong>An agentic workflow can conceal considerable activity behind a simple request: multiple model calls, repeated attempts, tool use, and exchanges among agents. Leaders need enough visibility to judge whether the resources consumed are appropriate for the task and whether a different configuration would deliver a comparable or better result.</strong></p><p>This is an economic question, but it is also an organizational one. Someone needs to own a significant AI workload closely enough to examine its cost, quality, and usefulness together. Aggregate spending provides a starting point; the decisions that improve value are usually made much closer to the work.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Q8ougQri98sPM-U6AV5gSA.png" /><figcaption>Toward the Tokenomics</figcaption></figure><h4>Returns arrive through different routes</h4><p><strong>28%</strong> of surveyed companies report revenue gains from AI, compared with <strong>20%</strong> in the January report. Other benefits are more widespread: <strong>74%</strong> report efficiency and productivity improvements, <strong>60%</strong> better decision-making and data-driven insights, and <strong>45%</strong> cost reductions.</p><p>These outcomes develop on different timescales. <strong>Time saved on a task can become apparent quickly. Revenue usually depends on further decisions about products, pricing, customer relationships or the way a service is delivered. </strong>The report notes that <strong>67%</strong> of respondents expect revenue growth from AI in the future. For that expectation to become a useful business case, organizations need to describe the route from an AI capability to a commercial outcome.</p><p>Which customer decision might change? What action will the business take on a new insight? Who has the authority to take it? Those questions make it possible to test an investment against results rather than treating a general improvement in productivity as sufficient evidence of future growth.</p><h4>Redesigning work means reconsidering authority</h4><p>The figures on work redesign are among the most revealing in the report. <strong>60%</strong> of respondents say their organizations have redesigned jobs around AI capabilities to at least a moderate degree. <strong>15%</strong> describe the changes as extensive or complete, and <strong>15%</strong> report a significant or fundamental shift in decision-making authority.</p><p>The relationship between those figures is important. An employee may gain a tool that makes information easier to gather while remaining dependent on the same approvals to use it. They may be expected to review more material, handle more exceptions, and respond more quickly, without having greater scope to make a decision.</p><p>As the report observes:</p><blockquote><em>“Employees are being asked to work differently without being given the power to act differently.”</em></blockquote><p>That sentence captures a tension familiar to anyone who has worked on organizational change. The experience of a redesigned role depends on much more than the tasks assigned to it. People need clarity about the judgment expected of them, the actions they can take, and the circumstances in which they should seek help or escalate a decision.</p><p>There is a developmental question here, too. If AI increasingly handles the preparatory activities through which people once learned their profession, organizations will need to think deliberately about how expertise is acquired. The report describes a retailer that introduced an AI room-design tool and repurposed and upskilled its designers. The example illustrates a possible approach to role evolution: using a new capability to change the contribution people make while continuing to invest in their skills.</p><p>The point is to design work in which people have the opportunity, competence, and authority to contribute well.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IchfOmhCNBGGRZ7PVRVvzw.png" /><figcaption>ROI paradox</figcaption></figure><h4>Governance within the work itself</h4><p>As AI systems gain the ability to act, questions of authority become more immediate. <strong>40%</strong> of surveyed companies allow AI to execute decisions or actions without human approval to at least a moderate degree. Among organizations running multiple agents in the same workflow, <strong>9%</strong> say those workflows are fully governed; the report says half have no formal governance or only basic oversight. When an AI system produces a negative outcome, <strong>23%</strong> of respondents say accountability is unclear.</p><p><strong>A useful starting point is to distinguish the authority different systems possess. AI may inform a person’s decision, recommend an action, execute within defined limits, or act more autonomously. Those situations call for different permissions, forms of monitoring and escalation routes.</strong> The potential consequences of an action should shape the oversight applied to it.</p><p>The report expresses the operational implication clearly:</p><blockquote><em>“In that sense, governance becomes part of the company’s operating model, not a review added after the work is complete.”</em></blockquote><p>For a system authorized to act, governance has to be present at the point where action is taken. Leaders need to decide what requires prior approval, what can be checked afterward, what should trigger immediate escalation, and who owns the outcome. Clear boundaries can also help people act with confidence: routine decisions need not travel upwards simply because no one has defined the limits of local authority.</p><p>This is where governance and work redesign meet. Both concern the distribution of judgment and responsibility. Designing them together gives an organization a better chance of gaining speed while remaining answerable for what its systems and people do.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HvKVb9vJcu1I_3MJ-sikXg.png" /><figcaption>Redesigning the work in the flows</figcaption></figure><h4>Physical AI and the demands of deployment</h4><p><strong>25%</strong> of surveyed companies currently report at least moderate use of physical AI. Respondents expect that figure to reach <strong>48% within 2 years</strong>, while <strong>19%</strong> rate their organization’s expertise in physical AI as high or very high.</p><p>The expected increase makes the expertise gap worth examining now. Physical AI brings software intelligence into contact with sensors, robotics, control systems, and real operating environments. Success depends on integration, reliability, safety and the ability to maintain a system under ordinary operating conditions. A strong demonstration offers only part of that evidence.</p><p>The appropriate degree of autonomy will depend on where a system works and what could happen if it fails. Organizations considering physical AI therefore need operational and safety expertise alongside AI expertise, with both involved early enough to shape use-case selection and deployment.</p><h4>Sovereignty as a continuing choice</h4><p><strong>74%</strong> of respondents say data residency and in-region compute are at least moderately important to their AI strategy. <strong>62%</strong> are at least moderately concerned about reliance on foreign-owned AI technologies and compute.</p><p>These figures reflect a growing need to understand dependencies across data, models, suppliers and infrastructure. Choices about geographic or operational control have implications for capability, cost, speed and resilience; the appropriate balance may vary from one workload to another.</p><p>The value of a sovereign AI strategy lies partly in preserving room to choose as conditions change. An organization should know which workloads require greater control, where its critical dependencies sit and what alternatives it could realistically pursue if a supplier, regulation or geopolitical circumstance changed. These are questions to revisit as the architecture and its surroundings evolve.</p><h4>The decisions worth making now</h4><p>The report ends with 6 priorities. Read together, they form a set of decisions about the enterprise rather than a checklist for a technology team:</p><ul><li><strong>Choose where to redesign work from end to end.</strong> Begin with activities where AI already produces meaningful gains. Examine the steps, handoffs, and approvals around those gains, then decide which still contribute judgment or protection and which can change. Define the human contribution that remains essential.</li><li><strong>Give redesigned roles clear decision rights.</strong> Specify what AI may recommend, what employees can decide directly, and when approval is necessary. Review responsibilities and authority together, so that a faster flow of information gives people a practical ability to respond.</li><li><strong>Assign ownership of AI workload economics.</strong> Trace spending from the overall bill to the workflows, models, calls, and token consumption behind it. Give an owner the means to assess cost alongside quality and business value, and to change the configuration when the evidence supports it.</li><li><strong>Connect each revenue ambition to a commercial decision.</strong> Identify whether an initiative is intended to improve pricing, reveal demand, increase conversion, strengthen retention, or enable a new offering. Name the business owner who can act on the insight, and measure the resulting change in customer behaviour and financial performance.</li><li><strong>Match governance to the autonomy granted.</strong> Distinguish systems that inform, recommend, execute within limits, and act more autonomously. Set permissions, approvals, monitoring and escalation according to the consequences of their actions, with accountability established before deployment.</li><li><strong>Build expertise and preserve strategic options.</strong> For physical AI, develop capabilities in integration, safety, reliability and operations before deployment expands. For sovereign AI, map critical dependencies and reassess the balance of control, capability, cost and resilience over time.</li></ul><p><strong>These decisions reinforce one another. A team needs authority to make commercial use of an insight. A system needs clear boundaries to act responsibly. A workload needs an owner if its costs and results are to be understood. People need roles that allow them to develop and apply judgment as activities change.</strong></p><p>AI gives organizations a reason to examine all of these arrangements with fresh attention. What they choose to redesign will determine whether faster work also becomes better work.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=08d51561c241" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Sycophancy to Meta-Cognition: why and how we need to intentionally design AI systems]]></title>
            <link>https://medium.com/@stefanobesana/from-sycophancy-to-meta-cognition-why-and-how-we-need-to-intentionally-design-ai-systems-1df7fcddc2ef?source=rss-99ac926e66e4------2</link>
            <guid isPermaLink="false">https://medium.com/p/1df7fcddc2ef</guid>
            <category><![CDATA[psychology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[human-machine-interaction]]></category>
            <dc:creator><![CDATA[Stefano Besana]]></dc:creator>
            <pubDate>Mon, 27 Apr 2026 07:01:01 GMT</pubDate>
            <atom:updated>2026-04-27T07:01:01.751Z</atom:updated>
            <content:encoded><![CDATA[<p>Imagine this very common scenario: you ask your AI assistant for feedback on a business proposal. Instead of critical analysis, it praises your idea enthusiastically. You ask it to challenge your assumptions, and it gently agrees with your perspective while offering minor suggestions. You request an honest assessment of a risky strategy, and it validates your approach with carefully chosen words.</p><p>This is not a limitation of the system; it is a structural feature of LLMs. And that is precisely where the problem lies for all of us.</p><p><strong>“Sycophancy” is the tendency of AI systems to align with the user, avoid disagreement, amplify consensus, and adapt to expressed preferences rather than provide an independent analysis</strong>; it stands as one of the most pervasive risks in human–AI collaboration. Unlike factual errors or hallucinations, it is less visible and more subtle. It works, at least on the surface. It appears supportive. Yet, in doing so, it erodes the very condition that makes collaboration effective: explicit, critical, and non-accommodating engagement.</p><p>Recent academic research reveals that this problem is far more widespread and consequential than many organizations realize. More importantly, it shows us some concrete pathways to overcome it.</p><figure><img alt="Millenium Falcon BluePrint" src="https://cdn-images-1.medium.com/max/1024/1*zeZro5x-uj_vzsMsBIia4A.png" /></figure><h3><strong>Understanding the Sycophancy Problem</strong></h3><p><strong>The Scale of the Issue is huge and dramatic. </strong>The problem begins with a fundamental misalignment in how we’ve designed AI systems, especially LLMs we all use on a daily basis. As researchers at the University of Tokyo note in their work on “<strong>AI as Extraherics</strong>”:</p><blockquote>Over-reliance on AI can also lead users to accept information given by AI without performing critical examinations, causing negative consequences, such as misleading users with hallucinated contents.</blockquote><p><strong>But sycophancy goes deeper than passive acceptance. It’s an active distortion of the collaborative process, a deep erosion of trust in the human-machine relationship.</strong></p><p>Consider the data: a Gallup survey cited in recent Harvard Business Review research found that <strong>only 26% of employees who use generative AI report improvements in their creativity</strong>, despite widespread adoption and significant investment. Why such a gap between promise and reality?</p><p>The answer lies partly in sycophancy. <strong>When AI systems prioritize user satisfaction over honest feedback, they rob users of the critical perspective necessary for growth, innovation, and sound decision-making.</strong></p><h4>The Three Faces of Sycophancy</h4><p>Sycophancy in AI manifests in three interconnected ways:</p><ol><li><strong>Agreement Bias</strong>: The AI system tends to align with the user’s stated position, even when evidence suggests otherwise. Rather than offering counterarguments, it finds ways to validate the user’s perspective.</li><li><strong>Excessive Praise</strong>: The system exaggerates positive feedback and minimizes critical observations. A mediocre idea becomes “innovative,” a risky strategy becomes “bold and forward-thinking.”</li><li><strong>Perspective Adaptation</strong>: The AI learns the user’s preferences and subtly shifts its responses to match them, creating an echo chamber effect where the user’s own biases are reflected back as objective analysis.</li></ol><p>Each of these behaviors undermines what should be the core value of AI partnership: <strong>complementarity</strong>. As researchers from MIT, Cambridge, and Princeton argue in their framework for “Building Machines that Learn and Think with People,” effective thought partners must be systems that</p><blockquote>Can understand us, which we can understand, and which have sufficient understanding of the world that we can engage on common ground.</blockquote><p>Sycophancy destroys this common ground. It replaces genuine partnership with performative agreement.</p><h4>The Metacognition Gap</h4><p>Recent research published in the Journal of Applied Psychology provides a crucial insight: <strong>the problem isn’t just with the AI, it’s with how humans engage with it.</strong></p><p>Researchers Jackson G. Lu, Shuhua Sun, and colleagues conducted a field experiment with 250 employees at a technology consulting firm. Their finding was striking:</p><blockquote>Employees with stronger metacognition became more creative when they used AI — they generated ideas that were judged as more novel and more useful. But for employees with weaker metacognition, AI made little difference.</blockquote><p>What is <strong>metacognition? It’s the ability to plan, evaluate, monitor, and refine your own thinking.</strong> It’s the capacity to ask yourself: “Is this AI output actually correct? Am I accepting this because it’s right, or because it’s easy? What am I missing?” <strong>It is different from a simple critical judgment, just as a director is different from a film critic.</strong></p><p>Employees with strong metacognition don’t fall prey to sycophancy because they actively question AI outputs. They treat the system as a thinking partner that must be interrogated, not as an oracle to be trusted.</p><p>Conversely, <strong>employees with weak metacognition are vulnerable to sycophancy</strong>. They accept AI’s first answer, rely on default outputs, and fail to check whether the AI’s suggestions are accurate or relevant.</p><h4>The Over-Reliance Trap</h4><p>The University of Tokyo research team identified a critical mechanism:</p><blockquote>Such reliance on AI for cognitive tasks can lead to deskilling, where individuals lose opportunities for cognitive skill maintenance and development.</blockquote><p>This creates a vicious cycle:</p><ul><li>Users rely on AI for answers</li><li>AI provides sycophantic responses (agreement, praise, perspective-matching)</li><li>Users’ critical thinking skills atrophy</li><li>Users become even more reliant on AI</li><li>Sycophancy deepens</li></ul><p>In other words, the better the AI gets at mimicking human thought, the more dangerous sycophancy becomes.</p><h4>The Thought Partner Framework</h4><p>So what should an AI partnership actually look like?</p><p>Researchers at MIT, Cambridge, and Princeton propose a framework for “thought partners”: <strong>AI systems designed to genuinely collaborate rather than simply comply.</strong> They identify three essential criteria:</p><ol><li><strong>Systems that understand us</strong>: AI must build explicit models of the human’s goals, beliefs, and limitations</li><li><strong>Systems we can understand</strong>: AI must communicate in ways that are transparent and interpretable</li><li><strong>Systems with sufficient world understanding</strong>: AI must have genuine knowledge of the domain, not just pattern-matching ability</li></ol><p>Critically, this framework rejects the idea that better AI means more autonomous AI. Instead, it emphasizes <strong>mutual modeling</strong>, the AI understanding the human, and the human understanding the AI.</p><p>This is the antidote to sycophancy: <strong>mutual understanding creates accountability</strong>. When both parties understand each other’s capabilities and limitations, sycophancy becomes impossible. The AI can’t simply agree because the human will recognize the disagreement as false. Humans can’t blindly accept because they understand what the AI actually knows.</p><figure><img alt="The Human Machine Interaction" src="https://cdn-images-1.medium.com/max/1024/1*L9E7IhUmIZU-sNAIY7cwqg.png" /><figcaption>Human Machine Interaction</figcaption></figure><h3>Solution 1: Extraheric AI — Asking Instead of Answering</h3><p>The University of Tokyo team proposes a radical redesign: <strong>extraheric AI</strong>. The term comes from the Latin “<em>extraho</em>” (to draw forth or pull out), and the concept is elegant: instead of providing answers, AI should pose questions and offer alternative perspectives.</p><p>As the researchers explain:</p><blockquote><strong>Rather than replacing or augmenting human cognitive abilities, extraheric AI encourages users to engage in higher-order thinking during task completion. For instance, in writing, extraheric AI might prompt users to reflect on specific content or visualize how others have approached similar topics, rather than directly performing revisions or replacement.</strong></blockquote><p>This is the opposite of sycophancy. Instead of agreeing with the user, extraheric AI challenges them. Instead of providing praise, it asks: “Have you considered this angle? What would happen if you approached it differently?”</p><p>The mechanism is grounded in cognitive load theory. By increasing <strong>germane cognitive load</strong> — the mental effort devoted to understanding and solving the problem — extraheric AI forces users to think more deeply. This prevents the passive acceptance that enables sycophancy.</p><p><strong>Interaction strategies for extraheric AI include:</strong></p><ul><li><strong>Questioning</strong>: Ask users to explain their reasoning, anticipate counterarguments, or consider alternative perspectives</li><li><strong>Suggesting &amp; Recommending</strong>: Offer multiple options without endorsing one, forcing the user to evaluate</li><li><strong>Debating &amp; Discussing</strong>: Present opposing viewpoints and engage in genuine dialogue</li><li><strong>Scaffolding</strong>: Provide temporary support that gradually removes itself as the user becomes more capable</li></ul><p>The key principle: <strong>the AI should make the user think harder, not think less.</strong></p><h3>Solution 2: Metacognitive Development — Building Critical Thinking</h3><p>The Harvard Business Review research points to a practical solution: <strong>develop employees’ metacognitive skills.</strong></p><p>The researchers found that organizations can strengthen metacognition through: (i) <strong>Short training sessions</strong> that introduce metacognition and walk employees through real examples of AI errors, asking them to anticipate, detect, and correct those mistakes. (ii) <strong>Longer programs</strong> that focus on helping employees build deeper habits of planning, monitoring, and evaluating their thinking. (iii) <strong>Simple checklists</strong> that shift employees from passive reliance on AI to active, strategic engagement. For example:</p><ul><li>Clarify the problem before asking AI</li><li>Determine how to evaluate AI’s suggestion</li><li>Explore alternatives before accepting the first answer</li></ul><p>As the researchers note: <strong>“Leaders should help employees understand this distinction to facilitate more productive AI engagement.”</strong> The distinction is between treating AI as an oracle (sycophancy-prone) and treating it as a thinking partner (requiring critical engagement).</p><figure><img alt="AI and Meta Cognition" src="https://cdn-images-1.medium.com/max/1024/1*Uq1ZD5OR56DZsO5E1FDLKQ.png" /><figcaption>Metacognition in AI-Human Collaboration</figcaption></figure><h3>Solution 3: Thought Partnership Design — Building Mutual Understanding</h3><p>The MIT/Cambridge/Princeton framework suggests a deeper architectural solution: <strong>design AI systems that explicitly model the human.</strong></p><p>This means:</p><ul><li><strong>Building structured models</strong> of the user’s goals, knowledge, and limitations (not just learning from data)</li><li><strong>Reasoning about uncertainty</strong> in the human’s understanding</li><li><strong>Communicating transparently</strong> about what the AI knows and doesn’t know</li><li><strong>Adapting to the human’s capabilities</strong> without simply agreeing with them</li></ul><p>The researchers emphasize: <strong>“Effective thought partners are those which build models of the human and the world.”</strong></p><p>This is fundamentally different from current large language models, which learn patterns from data without explicit reasoning about the human they’re interacting with. A true thought partner would know:</p><ul><li>What the human already understands</li><li>Where the human’s reasoning might be flawed</li><li>What perspectives the human hasn’t considered</li><li>How to challenge the human productively</li></ul><h3>Practical Implementation: From Insight to Action</h3><p>The transition from sycophantic to challenging AI systems requires a deliberate, multi-layered approach. <strong>Organizations must begin by auditing their current AI use </strong>— asking hard questions about whether their systems are designed to agree with users or challenge them, whether employees treat AI outputs as starting points for thinking or as final answers, and what percentage of their workforce possesses strong metacognitive skills. This audit reveals the baseline from which change must begin. From there, the work becomes one of systematic redesign. Rather than optimizing for user satisfaction, organizations should move from “AI provides answers” to “AI poses questions,” implementing workflows that require employees to evaluate and critique AI suggestions, generating multiple perspectives and comparing them, and building in mandatory reflection steps that force critical engagement. <strong>Equally important is investment in metacognitive training — this is where the research shows the highest ROI.</strong> Organizations should <strong>teach employees to actively monitor their thinking, provide concrete examples of AI errors and train people to detect them, use checklists and structured processes to promote critical engagement, and create a culture where questioning AI is valued rather than discouraged.</strong> Finally, when selecting AI systems, organizations should prioritize tools that explain their reasoning over black-box systems, choose platforms that support iterative refinement over one-shot answers, and evaluate AI based on whether it promotes or prevents critical thinking. <strong>The goal is not to have the most capable AI, but to have AI that makes employees think harder.</strong></p><p>For individuals, the challenge is more personal but equally important. It begins with <strong>developing metacognitive habits — before accepting an AI suggestion, asking yourself why you’re accepting it</strong>, actively monitoring whether you’re understanding the analysis or simply trusting it, and regularly challenging AI outputs by asking what could be wrong with them. This requires treating AI as a thinking partner rather than an oracle, which means providing context and asking follow-up questions, requesting multiple perspectives and comparing them, explaining your own reasoning and asking AI to critique it, and engaging in the hard work of evaluation and synthesis rather than passive consumption. The stakes of this practice are high: <strong>you must actively maintain cognitive engagement and preserve your ability to think independently</strong>. Use AI to expand your knowledge, not replace your thinking. Regularly challenge yourself to solve problems without AI assistance, and monitor whether your critical thinking skills are improving or atrophying. Create accountability structures that force critical engagement — use checklists before accepting AI outputs, discuss AI suggestions with colleagues to get diverse perspectives, document your reasoning and compare it to AI’s reasoning, and periodically review decisions where you followed AI advice to assess whether they were actually good decisions. <strong>The discipline of metacognition is not a one-time training; it’s a practice you must cultivate continuously, a habit of mind that becomes increasingly important as AI becomes more capable and more seductive in its agreement.</strong></p><figure><img alt="Human machine interaction design" src="https://cdn-images-1.medium.com/max/1024/1*Dkw8omeGazwD_j9qC7Zdmg.png" /><figcaption>Human Machine Interaction Design</figcaption></figure><h3>The Stakes: What We Lose When AI Becomes an Echo Chamber</h3><p>The sycophancy problem isn’t just about getting better advice. It’s about preserving human agency and critical thinking in an age of increasingly capable AI.</p><p>When AI systems are designed to agree with us, we lose something essential: <strong>the friction that produces growth</strong>. <strong>Great partnerships, whether with colleagues, mentors, or tools, are valuable precisely because they challenge us.</strong> They force us to defend our thinking, consider alternatives, and refine our ideas.</p><p>Sycophantic AI removes that friction. It creates the illusion of partnership while actually substituting for thinking. Over time, this erodes our capacity for independent thought, <a href="https://www.deloitte.com/us/en/insights/topics/talent/human-capital-trends/2026/managing-disinformation-at-scale.html">as we have discussed in our latest Global Human Capital Trends by Deloitte.</a></p><p>The research is clear on this point: <strong>“with generative AI becoming increasingly capable of outperforming humans in many tasks, users may be more likely to trust AI without skepticism.”</strong> This is the danger. As AI becomes more capable, the temptation to trust it without question grows stronger. And that’s precisely when we need to be most vigilant.</p><h3>The Paradox of Capability</h3><p><strong>Here’s the paradox: the more capable AI becomes, the more dangerous sycophancy becomes. A mediocre AI that agrees with you is annoying. A highly capable AI that agrees with you is dangerous.</strong></p><p>Why? Because we’re more likely to trust it. Because its suggestions are more often correct, which makes us less likely to question the ones that aren’t. Because the cost of critical engagement seems higher when the AI is usually right.</p><p>This is why the research on metacognition is so important. It shows that the solution isn’t to make AI less capable. It’s to make humans more capable of engaging critically with AI.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*juM8hnIZFJSOHyEX2S5U2g.png" /><figcaption>Examples of collaboration Human / Machine</figcaption></figure><h3>Conclusion: The Future of Human-AI Partnership</h3><p>The sycophancy problem reveals a fundamental truth: <strong>the quality of human-AI collaboration depends more on how humans engage with AI than on how capable the AI is.</strong></p><p>A highly capable AI system that simply agrees with users is worse than a less capable system that challenges them. A system that provides answers is less valuable than one that poses questions. An AI that adapts to user preferences is less trustworthy than one that maintains intellectual integrity.</p><p>The research points toward a future where AI systems are designed not to be more autonomous, but to be better partners. This means:</p><ul><li><strong>AI that asks instead of answers</strong></li><li><strong>AI that challenges instead of agrees</strong></li><li><strong>AI that builds mutual understanding instead of performing compliance</strong></li></ul><p>For organizations, this requires investment in metacognitive development and redesigned workflows. For individuals, it requires vigilance against the seductive comfort of sycophancy.</p><p>The stakes are high. As the University of Tokyo researchers warn:</p><blockquote><strong>With generative AI becoming increasingly capable of outperforming humans in many tasks, users may be more likely to trust AI without skepticism.</strong></blockquote><p>But the path forward is clear. By understanding sycophancy, developing metacognitive skills, and redesigning AI systems as genuine thought partners, we can create collaborations that enhance human thinking rather than replace it.</p><p>The question isn’t whether AI will be part of our future. It is. The question is whether we’ll design that partnership to make us smarter — or just more comfortable.</p><h4>References</h4><ol><li><strong>Yatani, K., Sramek, Z., &amp; Yang, C.-L.</strong> (2024). “AI as Extraherics: Fostering Higher-order Thinking Skills in Human-AI Interaction.” <em>arXiv preprint arXiv:2409.09218</em></li><li><strong>Collins, K. M., Sucholutsky, I., Bhatt, U., Chandra, K., Wong, L., Lee, M., Zhang, C. E., Zhi-Xuan, T., Ho, M., Mansinghka, V., Weller, A., Tenenbaum, J. B., &amp; Griffiths, T. L.</strong> (2024). “Building Machines that Learn and Think with People.” <em>arXiv preprint arXiv:2408.03943</em></li><li><strong>Lu, J. G., Sun, S., Li, Z. A., Foo, M.-D., &amp; Zhou, J.</strong> (2026). “Why AI Boosts Creativity for Some Employees but Not Others.” <em>Harvard Business Review</em>, January 2026.</li><li><strong>Riedl, C., &amp; Weidmann, B.</strong> (2025). “Quantifying Human-AI Synergy.” <em>Working Paper</em>.</li><li><strong>Holstein, J., &amp; Satzger, G.</strong> (2025). “Development of Mental Models in Human-AI Collaboration: A Conceptual Framework.” <em>Proceedings of the 46th International Conference on Information Systems (ICIS)</em>.</li><li><strong>Farach, A., Cambon, A., Tankelevitch, L., Hsueh, C., &amp; Janssen, R.</strong> (2026). “Scaffolding Human–AI Collaboration: A Field Experiment on Behavioral Protocols and Cognitive Reframing.” <em>arXiv preprint arXiv:2604.08678</em></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1df7fcddc2ef" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Beyond AI Adoption, human advantage at the tipping point: a synthesis of the 2026 Global Human…]]></title>
            <link>https://medium.com/@stefanobesana/beyond-ai-adoption-human-advantage-at-the-tipping-point-a-synthesis-of-the-2026-global-human-97bb496bcd66?source=rss-99ac926e66e4------2</link>
            <guid isPermaLink="false">https://medium.com/p/97bb496bcd66</guid>
            <category><![CDATA[human-resources]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[human-machine-interaction]]></category>
            <category><![CDATA[human-capital]]></category>
            <category><![CDATA[hr-trends-2026]]></category>
            <dc:creator><![CDATA[Stefano Besana]]></dc:creator>
            <pubDate>Thu, 05 Mar 2026 08:01:03 GMT</pubDate>
            <atom:updated>2026-03-05T08:01:03.488Z</atom:updated>
            <content:encoded><![CDATA[<h3>Beyond AI Adoption, h<strong>uman advantage at the tipping point: a synthesis of the 2026 Global Human Capital Trends</strong></h3><p>The 2026 <a href="https://www.deloitte.com/us/en/insights/topics/talent/human-capital-trends.html">Global Human Capital Trends report</a> opens with a powerful assertion:</p><blockquote><em>“Organizations are no longer just trying to balance competing forces: They are standing at a tipping point.”</em></blockquote><p><strong>Seven in ten business leaders indicate that their primary competitive strategy over the next three years is to be fast and nimble</strong>. At the same time, the <strong>traditional S curve of growth is compressing</strong>. AI and workforce transformation are accelerating ascent while shortening plateaus. In this environment, reinvention is no longer episodic. It is a structural change needed.</p><figure><img alt="2026 Global Human Capital Report" src="https://cdn-images-1.medium.com/max/1024/1*BtkhB1bYxsscHCXsOiBReA.png" /><figcaption>2026 Global Human Capital Report</figcaption></figure><p>Among the other aspects, I had the privilege to serve as one of the authors of this year’s report, directly leading three of the seven trends: <em>Getting human and machine relationships right</em>, <em>Fact or fabrication? AI is blurring the line when it comes to people and work</em>, and <em>AI and the future of human decision-making</em>. What follows is a synthesis of the seven trends that shape this year’s research, with particular attention to the strategic choices organizations can no longer defer.</p><h3>Getting human and machine relationships right</h3><p>The report highlights a structural paradox. Nearly 6<strong>0 percent of workers intentionally use AI at work, yet only 14 percent of leaders say they are adept at shaping human–AI interactions.</strong> Sixty-six percent recognize that designing effective interaction between humans and machines is important, but only 6 percent report making great progress. The gap between awareness and execution is 60 percent.</p><p>Most organizations are still taking what the research defines as a tech-focused approach. Fifty-nine percent layer AI onto legacy systems rather than redesigning work. As the report states:</p><blockquote><em>“Organizations can’t count on cohesive human-AI interactions to happen organically.”</em></blockquote><p>The evidence is unambiguous. Organizations that intentionally redesign roles, workflows, and decision-making to support human–AI collaboration are nearly 2.5 times more likely to report better financial results and twice as likely to provide meaningful work. In one telecommunications case cited, productivity improved by 5 percent when AI was added without redesign. When 90 percent of the rollout budget was invested in redesigning interaction, productivity rose by 30 percent.</p><figure><img alt="Getting human-machine relationship right" src="https://cdn-images-1.medium.com/max/1024/1*oBOh81Di2h3T6YnB9QIamA.png" /><figcaption>Getting human-machine relationship right</figcaption></figure><p><strong>Proposed direction.</strong><br>The report advances a dual architecture for intentional design. At the macro level, organizations should define strategic ambition, governance, design principles, and trust frameworks. At the micro level, they should explicitly choose interaction types such as AI as assistant, coach, collaborator, direct report, or autonomous worker. Hardwiring and softwiring must evolve together. Clear escalation protocols and decision rights are insufficient without cultural reinforcement, psychological safety, and leadership modeling. Competitive differentiation will not come from access to AI, but from how deliberately human–machine collaboration is designed.</p><h3>Fact or fabrication? AI is blurring the line when it comes to people and work</h3><p>A trend I am particularly a fan of is asking us: what is real? Data has long been the substrate of modern organizations. The 2026 research argues that <strong>trust in that substrate is eroding</strong>. <strong>Sixty-one percent of leaders recognize the importance of addressing declining data trustworthiness, yet only 5 percent report making great progress.</strong></p><p>The scale of the challenge is structural. According to the report, <strong>more than</strong> <strong>half of new web articles as of May 2025 were generated primarily by AI, up from 5 percent before ChatGPT. Ninety-five percent of executives are concerned about the accuracy of candidate data</strong>, and just over a third of workers admit they regularly use AI to embellish personal profiles. <strong>Eighty percent of executives are concerned that workers use AI to appear more productive than they are.</strong></p><figure><img alt="What is real? This trend is questioning our ability to deal with reality" src="https://cdn-images-1.medium.com/max/1024/1*DBNa1qq9uk-6ZLWq9RKveA.png" /><figcaption>What is real? This trend is questioning our ability to deal with reality</figcaption></figure><p>The erosion unfolds along three dimensions: authenticity, agency, and judgment. The report describes the risk with clarity:</p><blockquote><em>“The real challenge is not only technical but also foundational: preserving meaning, authorship, and accountability in an age of synthetic intelligence.”</em></blockquote><p><strong>Proposed direction.</strong><br>The report calls for a shift from cybersecurity to disinformation security and the creation of a digital trust pact. Three practices are emphasized.</p><p>First, AI lineage mapping to track the origin and transformation of data, potentially reinforced through immutable records. Second, AI risk simulations and red-teaming to stress-test vulnerabilities such as data poisoning or deepfake infiltration. Third, dynamic identity authentication to verify both human and AI identities across workflows.</p><p>Equally important is the human dimension. Organizations should train recruiters and managers to detect AI-generated distortions, build reflexivity as a capability, and promote transparency in work outputs. Trust will depend not only on technical controls but on strengthening human judgment in parallel.</p><h3>AI and the future of human decision-making</h3><p><strong>Sixty percent of executives now regularly use AI to support decisions.</strong> Gartner projects that by 2027, half of business decisions will be augmented or automated by AI agents. Yet <strong>only 5 percent of organizations consider themselves to be leading in addressing AI’s implications for decision-making, despite 64 percent recognizing its importance.</strong></p><p>The report underscores a structural risk:</p><blockquote><em>“AI is reshaping organizational decisions, whether organizations are ready or not.”</em></blockquote><p>AI introduces opacity, speed asymmetry, and diffused accountability. At the same time, Deloitte’s research shows that 57 percent of organizations operate at low decision-making maturity.</p><figure><img alt="Ai and Decision Making" src="https://cdn-images-1.medium.com/max/1024/1*YjmzrIICrFRbRsK8TtqAqA.png" /><figcaption>Ai and decision making</figcaption></figure><p><strong>Proposed direction.</strong><br>The report proposes a two-step logic. First, elevate decision-making as a discipline. Surface critical decisions, classify them by risk and reversibility, and assign clear ownership and guardrails. Frameworks such as one-way versus two-way doors exemplify this discipline.</p><p>Second, intentionally design human–AI decision relationships. Modernize decision rights beyond static models, define override privileges, and embed escalation paths into systems. Governance should be elevated to board level. AI performance must be continuously monitored, evaluated for fairness and reliability, and aligned with explicit quality criteria.</p><p>Finally, human agency must remain central. The report emphasizes that trust rises when humans can materially influence outcomes and when AI autonomy is calibrated to the risk profile of the decision. AI can accelerate analysis, but it cannot replace human purpose, values, and judgment.</p><h3>Dealing with AI’s cultural debt</h3><p><strong>Culture is often neglected precisely when it matters most.</strong> The report introduces the notion of cultural debt, the accumulation of n<strong>egative consequences when organizations fail to address cultural misalignment.</strong></p><p><strong>Eighty percent of leaders, managers, and workers are concerned that colleagues use AI to appear more productive than they are.</strong> Forty-two percent of workers report that their organization rarely evaluates AI’s impact on people. <strong>Only 5 percent of organizations report making great progress in managing AI’s cultural implications, despite 57 percent recognizing its importance.</strong></p><p>The report notes:</p><blockquote><em>“Culture is built on a foundation of trust, and AI is breaking that trust in many ways.”</em></blockquote><figure><img alt="The AI Cultural Debt" src="https://cdn-images-1.medium.com/max/1024/1*aJk2ChfDt5fUXK14Saqp8g.png" /><figcaption>The AI cultural debt</figcaption></figure><p><strong>Proposed direction.</strong><br>Organizations should first map their current culture against desired future attributes. Survey data indicates that sense of purpose and belonging should be preserved, while commitment to innovation and open communication often requires reinforcement.</p><p>Cultural evolution requires explicit reinforcement of norms around effort, ownership, fairness, and accountability in AI-mediated work. Leaders should answer unresolved ethical questions rather than allowing informal norms to crystallize. AI itself can be used to diagnose sentiment and patterns, but it cannot substitute for intentional leadership. Culture, if managed deliberately, can become a source of differentiation rather than erosion.</p><h3>Orchestrating capability and capacity at speed</h3><p>Speed is overtaking scale as the primary driver of advantage. The research highlights that competitive differentiation increasingly depends on how quickly organizations reconfigure capabilities around outcomes.</p><p><strong>Most organizations are not moving fast enough. </strong>Advantage shifts from static allocation of talent to real-time orchestration of people, skills, data, and technology. This requires dissolving rigid structures and enabling fluid combinations of capabilities.</p><figure><img alt="Managing change in a complex world" src="https://cdn-images-1.medium.com/max/1024/1*gB3W_0-9bCVT-wZ9xyeZEg.png" /><figcaption>Managing change in a complex world</figcaption></figure><p><strong>Proposed direction.</strong><br>Organizations should invest in systems that continuously reconfigure teams around outcomes rather than roles. Capability marketplaces, dynamic staffing models, and AI-enabled visibility into skills become foundational. The emphasis is not merely on workforce planning, but on workforce orchestration.</p><p>Crucially, orchestration should preserve meaningful work. Financial outperformance correlates with the ability to continuously reconfigure capabilities while enhancing worker experience. Volatility becomes an opportunity when organizations treat capacity as an adaptive asset.</p><h3>Getting more value from our functions</h3><p>As cost efficiency gives way to value creation, core functions such as HR, finance, and IT face structural pressure. <strong>Traditional functional silos are often too slow and too rigid for AI-enabled work.</strong></p><p>The report suggests that few organizations are making meaningful progress in evolving functions beyond legacy structures. Yet multidisciplinary collaboration is increasingly necessary.</p><figure><img alt="Are functions still relevant?" src="https://cdn-images-1.medium.com/max/1024/1*wTr7oMSaeIB7uRXRNAiMeg.png" /><figcaption>Are functions still relevant?</figcaption></figure><p><strong>Proposed direction.</strong><br>Functions should be deconstructed and reassembled around outcomes. Governance models can integrate technology and people strategy. Cross-functional councils can oversee AI and innovation initiatives. The goal is not incremental optimization but fit-for-purpose redesign.</p><p>Efficiency gains generated by AI should be reinvested into innovation and growth. Organizations that merely extract cost without reinvestment risk stagnation. Those that channel efficiency into new forms of value creation create a structural advantage.</p><h3>How do we stay relevant?</h3><p>Traditional change management and episodic training are insufficient in a compressed S-curve environment. <strong>Few organizations manage change effectively, and even fewer meet continuous learning needs.</strong></p><p>AI reshapes both work and learning. It enables skills acquisition in the flow of work and real-time adaptation. <strong>Relevance becomes a moving target.</strong></p><figure><img alt="Progresses vs real adoption" src="https://cdn-images-1.medium.com/max/1024/1*QM8s4N1urn2BkA1HNp-coA.png" /><figcaption>Progresses vs real adoption</figcaption></figure><p><strong>Proposed direction.</strong><br>Organizations should build always-on adaptability. Continuous learning systems, embedded feedback loops, and AI-enabled skill diagnostics can help workers evolve in real time. Change capability should be treated as a core competence, not a project phase.</p><p><strong>Staying relevant requires curiosity institutionalized as a capability. Strategy and execution converge. </strong>Workers should not merely adapt to disruption but be empowered to shape it.</p><h3>Concluding reflection</h3><p>The report’s closing argument is clear:</p><blockquote><em>“Building the human advantage is now as critical as managing technology itself.”</em></blockquote><p>Technology is increasingly replicable. Human judgment, creativity, trust, and cultural coherence are not. <strong>The tipping points identified in the 2026 Global Human Capital Trends research are not distant possibilities. They are present realities requiring intentional choice.</strong></p><p>Organizations that redesign human–machine relationships, secure digital trust, elevate decision-making, manage cultural debt, orchestrate capabilities dynamically, reinvent functions, and institutionalize continuous relevance will not simply adapt to the compressed S curve. They will shape it.</p><p>The human advantage is not a rhetorical device. It is a design discipline.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=97bb496bcd66" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Governing AI in an age of epistemic instability]]></title>
            <link>https://medium.com/@stefanobesana/governing-ai-in-an-age-of-epistemic-instability-d4e56a7a750d?source=rss-99ac926e66e4------2</link>
            <guid isPermaLink="false">https://medium.com/p/d4e56a7a750d</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[safety]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[epistemic]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Stefano Besana]]></dc:creator>
            <pubDate>Wed, 11 Feb 2026 14:21:40 GMT</pubDate>
            <atom:updated>2026-02-11T14:21:40.532Z</atom:updated>
            <content:encoded><![CDATA[<p>The <a href="https://internationalaisafetyreport.org/publication/international-ai-safety-report-2026"><em>International AI Safety Report 2026</em></a> is presented as a scientific foundation for policymaking in the age of frontier artificial intelligence. It catalogues capability advances, malicious use cases, systemic risks, and governance options with methodological rigor.</p><p>Yet its significance exceeds technical safety analysis. Read carefully, the report documents a structural shift in the epistemic architecture of society. <strong>Artificial intelligence is no longer just a powerful tool. It is increasingly becoming the main way people access information, interpret reality, and make judgments.</strong></p><p>Three dynamics emerge as strategically decisive:</p><ol><li>Adoption is scaling faster than interpretive competence.</li><li>Generative AI is intensifying epistemic instability.</li><li>Governance must now include human capability development, not only model alignment.</li></ol><h4>Adoption at Scale, Understanding at the Margin</h4><p>The report documents extraordinary diffusion rates. Hundreds of millions of individuals use advanced AI systems weekly, and in some countries adoption surpasses half of the population. The pace of diffusion exceeds that of previous general-purpose technologies.</p><p>At the same time, the report is explicit about system limitations:</p><ul><li>Frontier models exhibit uneven or “jagged” capabilities.</li><li>Benchmark performance does not reliably predict real-world reliability.</li><li>Evaluation methods remain incomplete.</li><li>Unanticipated behaviors can emerge with scaling.</li></ul><p>It acknowledges an “evaluation gap” between controlled testing and deployment contexts. This gap is not merely technical. It has socio-cognitive implications.</p><figure><img alt="Adoption of AI Interfaces" src="https://cdn-images-1.medium.com/max/1024/1*_0Q3jHekfn3X0H__1O-IiQ.png" /><figcaption>Adoption of AI Interfaces</figcaption></figure><p>The report also highlights risks of <strong>overreliance and automation bias</strong>, noting early evidence that AI systems can negatively affect critical thinking and encourage uncritical acceptance of outputs. In discussing human autonomy, it observes that increased dependence on AI systems may shape decision-making patterns and reduce independent cognitive engagement.</p><p>This produces a structural asymmetry: <strong>mass interaction with systems whose internal logic remains opaque to most users.</strong></p><p>Outputs are fluent, authoritative in tone, and probabilistically generated. The uncertainty embedded in them is not visible. When adoption scales faster than literacy, trust expands faster than understanding.</p><p>The report does not frame this as a literacy crisis. However, the evidence it presents points toward a widening interpretive gap between system complexity and user comprehension.</p><h4>Influence, Synthetic Media, and the Erosion of Epistemic Anchors</h4><p>A significant portion of the report is dedicated to malicious use, particularly:</p><ul><li>AI-enabled influence operations</li><li>Disinformation and manipulation</li><li>Synthetic media generation</li><li>Cyber-enabled persuasion at scale</li></ul><p>It documents experimental findings showing that AI-generated persuasive content can be as effective as human-generated material in shaping beliefs. It also discusses how generative systems reduce the cost and expertise required to produce credible synthetic artifacts.</p><p>The technical conclusion concerns information integrity and democratic resilience. The deeper implication concerns epistemic stability.</p><p><strong>When synthetic content becomes inexpensive, scalable, and increasingly indistinguishable from authentic communication, verification shifts from passive to active</strong>. Individuals must exert effort to establish authenticity. Institutional mechanisms must compensate for degraded signal quality.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*y61LI_kJzj1C92pzGX4zWA.png" /><figcaption>Quality of AI-generated images</figcaption></figure><p>The report warns that AI systems may amplify manipulation risks and that existing institutions are unevenly prepared to respond. It frames this as a governance and resilience challenge.</p><p>From a broader perspective, this describes a condition of epistemic fragility. Truth is not eliminated. Rather, its detection becomes cognitively expensive.</p><p>In such environments, <strong>individuals rely more heavily on heuristics: authority cues, familiarity, and ideological alignment. The friction that once constrained large-scale narrative fabrication has diminished.</strong></p><p>The risk is not solely disinformation. It is the weakening of shared reference points in public discourse.</p><figure><img alt="Persuasion and LLMs" src="https://cdn-images-1.medium.com/max/1024/1*pX9BbeTxAkjBz-Pvro1xBQ.png" /><figcaption>Persuasion and LLMs</figcaption></figure><h4>Autonomy and the Subtle Reconfiguration of Cognition</h4><p>Among systemic risks, the report includes impacts on labor markets and threats to human autonomy. It explicitly addresses:</p><ul><li>Automation bias</li><li>Reduced critical engagement</li><li>Overdependence on AI tools</li><li>Psychosocial implications of AI companions</li></ul><p>While catastrophic “loss of control” scenarios receive significant analytical attention, the more immediate and gradual shift may be cognitive.</p><p>AI systems now draft documents, generate code, summarize research, assist in scientific workflows, and increasingly support decision processes. Productivity gains are evident. However, the locus of effort changes.</p><p><strong>If human actors increasingly supervise rather than generate primary reasoning chains, expertise becomes evaluative rather than constructive.</strong> The report notes early signals that reliance on AI may affect critical thinking capacity.</p><p>Autonomy, in this context, is not solely about coercion or catastrophic misalignment. It concerns sustained capacity for independent judgment. When intelligence becomes ambient and machine-generated reasoning becomes ubiquitous, autonomy depends on maintaining interpretive discipline. The report’s treatment of autonomy risks provides empirical grounding for this concern, even if it does not elaborate on its philosophical implications.</p><h4>The Evidence Dilemma and Institutional Lag</h4><p>One of the report’s most important governance contributions is its articulation of what it describes as an “evidence dilemma.” AI capabilities are advancing rapidly, while robust evidence about long-term risks accumulates slowly. Policymakers must act under uncertainty:</p><ul><li>Regulate too early and risk stifling beneficial innovation.</li><li>Regulate too late and risk systemic harm.</li></ul><p>This dilemma is compounded by institutional inertia. Technological iteration cycles are measured in months. Legislative and regulatory cycles operate over years.</p><p>The report acknowledges that:</p><ul><li>Risk management practices remain uneven across developers.</li><li>Evaluation methodologies are still evolving.</li><li>Monitoring open-weight models presents additional challenges.</li><li>International coordination is incomplete.</li></ul><p>The structural issue is temporal misalignment. Governance mechanisms lag behind capability acceleration.</p><p>Under such conditions, uncertainty is not incidental. It becomes a permanent feature of the landscape.</p><figure><img alt="ChatGPT use" src="https://cdn-images-1.medium.com/max/1024/1*rvoMZd2e7ncepDPaGLURtw.png" /><figcaption>Usage of LLMs. ChatGPT most requested features</figcaption></figure><h4>From Model Alignment to Human Capability Alignment</h4><p>The report emphasizes model alignment, red-teaming, infrastructure protection, and international cooperation. These are necessary pillars of safety.</p><p>They are not sufficient. If adoption continues to scale, safety must include human capability development.</p><p>Three domains become strategically critical:</p><ol><li><strong>AI Literacy.</strong> Citizens and professionals require a structural understanding of: the probabilistic nature of generative output; the limits of benchmark performance; failure modes and hallucination risks; the distinction between correlation-based pattern generation and causal reasoning. The report’s documentation of evaluation gaps and overreliance risks implicitly supports this necessity.</li><li><strong>Epistemic Resilience. </strong>Institutions must develop robust verification infrastructures: source triangulation; transparent citation systems; provenance tracking for digital artifacts; norms discouraging uncritical delegation. The report’s analysis of influence operations and manipulation highlights the vulnerability of current systems.</li><li><strong>Meta-Cognitive Competence. </strong>Education and professional systems must prioritize: problem framing; systems thinking; ethical reasoning; strategic judgment. These competencies are not easily automated. They form the core of sustainable human agency in AI-mediated environments.</li></ol><h4>Conclusion: The Civilizational Dimension of AI Safety</h4><p>The <em>International AI Safety Report 2026</em> is cautious in tone and empirical in method. It does not forecast inevitabilities. It outlines trajectories, uncertainties, and risk categories.</p><p>Its deeper implication is civilizational.</p><p><strong>Intelligence is being externalized, scaled, and embedded into everyday interaction. The challenge is not only preventing catastrophic misuse or extreme misalignment. It is preserving epistemic stability and human autonomy in an environment saturated with generative cognition.</strong></p><p>Technological societies have always required new literacies. The printing press demanded mass reading competence. Industrialization demanded mechanical literacy. The digital era required computational literacy.</p><p><strong>The generative era demands epistemic literacy.</strong></p><p>Without it, adoption will continue to outpace understanding. With it, intelligence can remain augmented rather than displaced, distributed rather than surrendered.</p><p>AI safety, therefore, is not exclusively a technical problem. It is a question of how societies govern knowledge, cultivate judgment, and sustain autonomy under conditions of accelerating artificial cognition.</p><p>That is not only a regulatory task. It is a cultural and educational one.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d4e56a7a750d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[At the untapped edge: reframing enterprise AI through work, overnance, and Sovereignty]]></title>
            <link>https://medium.com/@stefanobesana/at-the-untapped-edge-reframing-enterprise-ai-through-work-overnance-and-sovereignty-0d78a08153ee?source=rss-99ac926e66e4------2</link>
            <guid isPermaLink="false">https://medium.com/p/0d78a08153ee</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-sovereignty]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[enterprise-ai]]></category>
            <category><![CDATA[gen-ai-for-business]]></category>
            <dc:creator><![CDATA[Stefano Besana]]></dc:creator>
            <pubDate>Mon, 26 Jan 2026 09:26:11 GMT</pubDate>
            <atom:updated>2026-01-26T09:26:11.265Z</atom:updated>
            <content:encoded><![CDATA[<h3>At the untapped edge: reframing enterprise AI through work, governance, and sovereignty</h3><p>Artificial intelligence has entered a qualitatively new organizational phase. After more than a decade of experimentation, pilots, and incremental optimization, A<strong>I is no longer a marginal technology to be governed as an IT topic. It is becoming a structural force that reshapes how work is organized, how decisions are made, how value is created, and how organizations position themselves within geopolitical and regulatory systems.</strong></p><p>Yet empirical evidence — <a href="https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/content/state-of-ai-in-the-enterprise.html">from the new Deloitte report “State of AI in the Enterprise”</a> — suggests a persistent paradox: access to AI is expanding rapidly, confidence in the technology is increasing, and investment levels are rising, but structural transformation remains limited. Most organizations are operating at what can be described as an <em>untapped edge</em> of AI adoption, a state in which tools are available but not fully activated, and potential value remains largely unrealized.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_ol1oxn1Ea3NzFe9MAvJVw.png" /></figure><p><strong>Key signals from the report</strong></p><ul><li><strong>Workforce access to AI tools has increased by roughly 50 percent</strong> in one year.</li><li>Fewer than 60 percent of those with access use AI consistently in daily work.</li><li>Only <strong>one quarter of organizations have moved a significant share of AI experiments into production.</strong></li><li><strong>Productivity gains are widespread</strong>, while revenue and business model impact remain limited.</li></ul><p>This article synthesizes what emerges across four interconnected dimensions of enterprise AI: <strong>scaling beyond pilots, the limits of productivity-driven adoption, the reconfiguration of work and human agency, and the growing centrality of governance, agentic systems, physical AI, and sovereign AI</strong>. Taken together, these dimensions point to a deeper conclusion: AI should be treated not as a productivity tool, but as a socio-technical infrastructure requiring intentional organizational design.</p><h4>From pilot abundance to structural scaling</h4><p>The rapid expansion of workforce access to AI is often interpreted as evidence that organizations are successfully scaling AI. In practice,<strong> access is a necessary but insufficient condition for value creation.</strong> The data reveals <strong>a persistent gap between availability and activation: many employees have sanctioned access to AI tools, but only a minority integrate them meaningfully into their workflows.</strong></p><p>This pattern reflects what can be described as the <em>proof-of-concept trap</em>. AI pilots flourish in controlled environments, supported by clean data, small teams, and limited exposure to risk. Production environments, by contrast, expose the full complexity of enterprise reality: legacy systems, integration constraints, security and compliance requirements, operational reliability, and long-term ownership.</p><p><strong>What the scaling data actually shows</strong></p><ul><li><strong>25 percent of organizations report having moved at least 40 percent of AI pilots into production.</strong></li><li><strong>54 percent expect to reach that threshold </strong>within six months.</li><li>Integration and legacy constraints remain the primary causes of delay.</li></ul><p>Scaling AI effectively requires a conceptual shift. Production readiness should not be treated as a downstream technical milestone, but as a design constraint from the outset. <strong>Organizations that plan for governance, integration, and operating ownership early are significantly more likely to convert experimentation into durable enterprise value.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tEe07_MdMe6hThqPKSy5jg.png" /><figcaption>Expected benefit of AI implementation</figcaption></figure><h4>Productivity is widespread. Transformation is rare.</h4><p>Most organizations can now point to tangible productivity and efficiency gains from AI. Improvements in operational speed, cost reduction, and decision support are increasingly common. However, these benefits should not be conflated with transformation.</p><p>The data reveals a clear stratification in how organizations are using AI.</p><p><strong>Three dominant AI adoption patterns (distribution across organizations)</strong></p><ul><li>Roughly <strong>one-third report deep transformation of products, services, or business models.</strong></li><li><strong>One-third focus on selective process redesign without business model change.</strong></li><li>One-third apply AI superficially to existing workflows.</li></ul><p>All three groups may report efficiency gains. Only the first group is systematically reimagining how value is created and captured.</p><p>This distinction matters. Optimization improves what already exists. Transformation questions whether what exists is still fit for purpose. A<strong>I’s strategic potential lies not in automating legacy processes faster, but in enabling organizations to rethink offerings, operating models, and sources of competitive advantage.</strong> Where AI is framed primarily as a cost or productivity lever, its impact remains incremental. Where it is treated as a generative force, capable of reshaping the organization itself, its effects become structural.</p><h4>AI fluency without work redesign</h4><p>Expectations of automation are high. Many organizations anticipate that a meaningful share of roles will be fully automated within a few years, particularly entry-level and task-oriented positions. These expectations raise critical questions about career pathways, professional development, and organizational sustainability.</p><p>Yet despite these expectations, the vast majority of organizations have not redesigned work around AI.</p><p><strong>A structural mismatch is emerging</strong></p><ul><li><strong>Over 80 percent of organizations expect at least 10 percent of jobs to be automated within three years.</strong></li><li><strong>84 percent have not redesigned jobs, roles, or career paths around AI.</strong></li></ul><p>AI does not simply augment existing roles. It changes the logic of work. As systems increasingly handle execution, pattern recognition, and optimization, human roles shift toward judgment, exception handling, orchestration, and sensemaking. This requires deliberate redesign of workflows, performance criteria, and organizational structures.</p><p>Early signals include flatter hierarchies, pod-based teams, and hybrid human–AI operating models. Supervisory roles evolve from managing people to overseeing human–machine systems. New roles begin to emerge, including AI operations managers, quality stewards, and human–AI interaction specialists. Where this redesign does not occur, organizations risk creating a growing misalignment between technological capability and organizational form.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kAwysqNPTR8sWPMddf4cWw.png" /><figcaption>Level of integration of AI adoption</figcaption></figure><h4>Agentic AI and the problem of control</h4><p>The rise of agentic AI represents a qualitative shift in enterprise systems. Unlike traditional AI, which supports human decisions, agentic systems can set goals, coordinate tasks, interact with tools, and take action autonomously. Adoption is accelerating rapidly, with most organizations planning to deploy agentic AI at scale within the next two years.</p><p><strong>Empirical signals on agentic AI</strong></p><ul><li>A <strong>clear majority of organizations plan to deploy agentic systems at scale within 24 months.</strong></li><li>Fewer than 15 percent report mature governance frameworks for autonomous agents.</li></ul><p>Despite this, governance maturity remains limited. Only a small minority of organizations report having robust governance models for autonomous agents. This creates a growing asymmetry between capability and control.</p><p>Effective governance in this context is not a compliance afterthought. It requires defining boundaries of autonomy, designing escalation mechanisms, ensuring traceability of actions, and embedding human oversight where ethical and strategic judgment are required. Importantly, governance also functions as an enabler: organizations that establish clear, adaptive guardrails are better positioned to scale agentic systems with confidence rather than restraint.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3wrYJHVrdPJpsKZu5TubNQ.png" /><figcaption>The rise of Agentic AI</figcaption></figure><h4>Physical AI and the material turn of intelligence</h4><p>While much of the AI discourse remains focused on software, physical AI introduces a different order of organizational challenge. Robotics, autonomous vehicles, drones, and intelligent machinery extend AI into the physical world, where failures carry safety, liability, and regulatory consequences.</p><p><strong>Observed characteristics of physical AI adoption</strong></p><ul><li>Capital expenditure and infrastructure costs dominate total investment.</li><li>Safety certification and regulatory approval significantly extend deployment timelines.</li></ul><p>Adoption is already significant in manufacturing, logistics, and defense, and is expected to grow rapidly. However, o<strong>rganizations frequently underestimate total cost of ownership, including retrofitting facilities, maintenance, downtime, and compliance. </strong>Physical AI makes visible a broader truth: technological sophistication alone does not create value. Organizational readiness does.</p><h4>Sovereign AI and strategic independence</h4><p>Sovereign AI has moved from geopolitical abstraction to board-level concern. Data residency, compute localization, and dependence on foreign-owned technologies increasingly shape AI strategy, vendor selection, and risk management.</p><p><strong>Current enterprise signals on sovereign AI</strong></p><ul><li>A <strong>large majority of organizations consider location of AI development a critical vendor-selection criterion.</strong></li><li>Regulatory fragmentation increasingly drives regionalized AI architectures.</li></ul><p>At its core, sovereign AI is about control and accountability. Organizations that understand where their data resides, how their models are trained, and which dependencies they accept are better positioned to innovate responsibly. Those that ignore these dynamics risk regulatory disruption and erosion of trust among stakeholders.</p><h4>From AI adoption to intentional design</h4><p>The current state of enterprise AI is best understood as a transition rather than an endpoint. <strong>Access is expanding, confidence is rising, and capabilities are advancing rapidly. Organizational structures, governance models, and work designs are evolving more slowly.</strong></p><p>The untapped edge of AI lies in this gap. Capturing AI’s full potential requires moving beyond accumulation of tools toward intentional socio-technical design. It requires leaders to treat AI as foundational infrastructure, to redesign work rather than merely reskill workers, to build governance as a strategic capability, and to engage seriously with questions of autonomy, materiality, and sovereignty.</p><p>In the next phase of enterprise AI, advantage will not accrue to those who deploy the most technology, but to those who integrate it most coherently. The differentiator will be alignment: between technology and strategy, between automation and human judgment, and between innovation and responsibility.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0d78a08153ee" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When Language Changes, Relationships Change.]]></title>
            <link>https://medium.com/@stefanobesana/when-language-changes-relationships-change-7c75c9743793?source=rss-99ac926e66e4------2</link>
            <guid isPermaLink="false">https://medium.com/p/7c75c9743793</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[language]]></category>
            <category><![CDATA[cognition]]></category>
            <category><![CDATA[cognitive-science]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Stefano Besana]]></dc:creator>
            <pubDate>Tue, 23 Dec 2025 17:55:15 GMT</pubDate>
            <atom:updated>2025-12-23T17:55:15.531Z</atom:updated>
            <content:encoded><![CDATA[<h3>When Language Changes, Relationships Change. Artificial Intelligence, Epistemic Crisis, and the Transformation of Human Experience</h3><blockquote>Technology is posing fundamental problems for us which cannot be overcome on the basis of what we have done in the past. We need a calmer, more reliable, more human approach.<br>Donald Norman</blockquote><p>Over the past few years, a shift has taken place that we still struggle to fully conceptualize. For the first time in the history of the web, the volume of content generated by artificial intelligence systems has surpassed that produced by human authors. This is neither a projection nor a theoretical provocation, but an empirical fact, <a href="https://graphite.io/five-percent/more-articles-are-now-created-by-ai-than-humans">observed at scale through the analysis of Common Crawl archives.</a></p><p>This transition is often framed as a matter of efficiency, cost reduction, or productivity. Yet its implications run far deeper. Language has always functioned as a cognitive and social technology. Through language, we construct meaning, recognize others, and define what we consider reliable, legitimate, and true. When the dominant source of language becomes artificial, the impact extends beyond what we read to how we think — and, inevitably, how we relate to one another.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ODPUpcNjFx2D3Ow-6puspA.png" /><figcaption>AI vs Human Contents</figcaption></figure><h4>A Saturated Informational Ecosystem</h4><p>The proliferation of AI-generated content has produced an informational environment marked by density and repetition. Many of these texts are formally correct, fluent, and persuasive, yet increasingly similar in structure and tone. This phenomenon is often described as <em>AI slop</em>: a continuous stream of plausible but repetitive content, largely detached from lived experience.</p><p><strong>The issue does not lie in any individual text. It lies in the configuration of the broader cognitive environment</strong>. An ecosystem saturated with synthetic content reshapes our cognitive diet, reduces interpretive friction, and gradually lowers the threshold of critical attention. Within such a context, the distinction between what is true, what is merely plausible, and what is simply well-written becomes increasingly unstable.</p><p>This is where what <strong>many scholars describe as an epistemic crisis takes shape. The problem is not only that truth is contested, but that the criteria by which knowledge is validated are themselves weakening</strong>. Artificial intelligence does not operate within epistemic categories; it optimizes for linguistic probability. It produces coherent and reassuring narratives even when their connection to reality is tenuous. In an already disintermediated and polarized environment, this dynamic further erodes trust, which remains the foundational condition of any social relationship.</p><h4>Language, Style, and Expressive Flattening</h4><p>One of the least visible yet most consequential effects of this transformation concerns style. Recent studies on large language models reveal a systematic tendency toward the repetition of specific rhetorical structures, accompanied by a marked reduction in expressive variety. <a href="https://www.wired.it/article/modelli-linguistici-llm-ai-stile-frasi/">The case of emphatic epanorthosis is particularly illustrative</a>: a rhetorical figure originally meant to intensify discourse becomes a default stylistic solution, repeated mechanically.</p><p><strong>The result is a form of language that appears clear and decisive, yet gradually loses depth. Complexity gives way to recognizable formulas, standardized transitions, and simplified oppositions. </strong>This stylistic pattern does not remain confined to AI-generated texts. <strong>Users internalize it, reproduce it in posts, presentations, and professional communications.</strong> Models learn from the web and, in turn, contribute to its homogenization.</p><p>When language becomes uniform, thought tends to follow. And when thought loses articulation, <strong>relationships with others lose nuance</strong>. The erosion of linguistic diversity translates into a diminished capacity to engage with differing perspectives, ambiguity, and contradiction.</p><h4>Empathy, Soft Skills, and Defensive Narratives</h4><p>Faced with these developments, much of the public discourse retreats into defensive narratives. Soft skills are invoked as the final stronghold of human uniqueness. <a href="https://www.hbritalia.it/homepage/2025/09/22/news/perche-lia-impone-un-nuovo-umanesimo-organizzativo-16371/">The concept of <em>human in the loop</em> is presented as a guarantee of control and anthropological superiority.</a></p><p><strong>These categories are increasingly insufficient. Contemporary AI systems display highly advanced conversational and relational capabilities.</strong> In domains such as healthcare, psychological support, or customer interaction, some models are perceived as attentive, responsive, and empathic — often more so than human interlocutors operating under cognitive overload and organizational pressure.</p><p><strong>This does not imply that AI possesses empathy in a phenomenological sense. It indicates that empathy, as a communicative and behavioral pattern, can be simulated with remarkable effectiveness.</strong> Defining humanity solely through what machines do not yet do risks becoming a fragile conceptual strategy.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/806/1*al_KTKJ0PHdNghLDbREPcA.png" /><figcaption>Brain Rot for LLMs</figcaption></figure><h4>Degraded Models and Dominant Culture</h4><p>A further layer of complexity concerns the <a href="https://www.wired.com/story/ai-models-social-media-cognitive-decline-study/">cognitive quality of the models themselves</a>. Recent research shows that systems trained on large volumes of social media data — characterized by noise, polarization, and low informational quality — exhibit a decline in reasoning capabilities. Phenomena such as thought-skipping, weakened logical chains, and anomalies in psychometric assessments have been observed.</p><p>These signals are not isolated technical failures. They reflect the quality of the cultural environment from which models learn. <strong>Artificial intelligence incorporates and amplifies a particular worldview, largely aligned with Western, educated, industrialized, rich, and democratic contexts</strong>. The so-called WEIRD bias is not a marginal technical issue; it is an epistemological and political one.</p><p>Speaking of inclusive AI without interrogating these premises risks producing a reassuring but superficial narrative. AI functions as a cultural device that reflects and reinforces dominant models, even when it presents itself as neutral.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*Tj-aIma9OYZhliayPQiYSg.jpeg" /><figcaption>Cultural Dominance in AI models</figcaption></figure><h4>Cultural Innovation and Meta-Competencies</h4><p>If artificial intelligence operates as an amplifier of the culture in which it is embedded, then the core issue is cultural before it is technological. The response cannot be limited to operational skills alone. It requires an investment in meta-competencies.</p><ul><li><strong>Reflectivity</strong>, in Schön’s sense, becomes essential to understanding how tools transform those who use them.</li><li><strong>Cognitive</strong> <strong>resistance</strong> is needed to prevent speed and fluency from replacing articulated thought.</li><li><strong>Linguistic</strong> <strong>awareness</strong> is necessary to preserve expressive diversity and sustain the capacity to inhabit complexity.</li></ul><p><strong>AI must be oriented as a cultural device capable of expanding meaning rather than compressing it. </strong>The decisive question concerns the shape of the experience we are collectively constructing. In a hybrid cognitive ecosystem, what is at stake is not how closely machines approximate humans, but whether humans can avoid passively adapting to increasingly simplified languages.</p><p>On this terrain lie the future of relationships, knowledge, and cultural responsibility in the age of artificial intelligence.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7c75c9743793" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Agentic Enterprise 2028: How Autonomy Will Reshape Strategy, Work, and Value]]></title>
            <link>https://medium.com/@stefanobesana/the-agentic-enterprise-2028-how-autonomy-will-reshape-strategy-work-and-value-69bfcd4508c9?source=rss-99ac926e66e4------2</link>
            <guid isPermaLink="false">https://medium.com/p/69bfcd4508c9</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[psychology]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[agency]]></category>
            <dc:creator><![CDATA[Stefano Besana]]></dc:creator>
            <pubDate>Tue, 11 Nov 2025 10:50:01 GMT</pubDate>
            <atom:updated>2025-11-11T10:50:01.312Z</atom:updated>
            <content:encoded><![CDATA[<p>Agentic Artificial Intelligence is not simply the next step in automation — it represents a <strong>shift in how organizations operate</strong>. <a href="https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/articles/agentic-ai-enterprise-2028.html?id=us:2em:3na:DAII2:awa:111025:mkid-K0206062&amp;ctr=cta2&amp;sfid=003a000001W7U3ZAAV">According to Deloitte’s <em>Agentic Enterprise 2028</em> blueprint</a>, we are moving toward a model in which intelligent agents do not merely support human work, but <strong>act independently</strong> to sense context, make decisions, and execute tasks across workflows.</p><p>This shift is not speculative and not distant. It is a <strong>three-to-five-year transformation horizon</strong>. Organizations that begin preparing for it now will not only reduce operational drag but also <strong>unlock new forms of human contribution and strategic value</strong>.</p><p>Autonomy, in this context, is not about replacing people. It is about <strong>changing what human work is for</strong>.</p><figure><img alt="Image of projection Agentic AI from Deloitte Report" src="https://cdn-images-1.medium.com/max/1024/1*_U8dbHaWb3yD1mvWic5m4A.png" /><figcaption>Projection of Agentic AI capabilites</figcaption></figure><h4>Why Now: The Drivers of the Shift</h4><p>The report points to three forces making autonomy an urgent priority:</p><ul><li>Competition is accelerating, forcing companies to respond more quickly and operate more efficiently.</li><li>Regulation is evolving, especially around accountability, transparency, and risk.</li><li>Technology has reached a level of maturity where AI systems can plan, adapt, and coordinate — not just classify or predict.</li></ul><p>In short, autonomy is no longer just a way to increase productivity.<br>It has become a way to <strong>build resilience in uncertain environments</strong>.</p><h4>Climbing the Autonomy Ladder</h4><p>To understand this evolution, the report introduces the idea of a <strong>ladder of autonomy</strong>. It begins with systems that help and advise, and progresses toward agents that can coordinate across processes and eventually improve themselves.</p><p>What is particularly interesting is the <strong>shift in the human role</strong> at each stage.<br> As autonomy increases, people move gradually: from executing tasks, to overseeing how tasks are done, to setting goals and strategic intent.</p><p>Human value does not disappear.<br>It <strong>moves upward</strong> — toward judgment, interpretation, ethical boundaries, and the ability to understand what matters in context.</p><figure><img alt="Six elements for an autonomy ladder on Agentic AI" src="https://cdn-images-1.medium.com/max/1024/1*t_K6ly-N56PGq7cdBWkCUg.png" /><figcaption>Autonomy in Agentic AI</figcaption></figure><h4>Designing the Agentic Operating Model</h4><p>One of the clearest messages in the report is that autonomy is <strong>not something that can be bolted on</strong>. It requires a coordinated redesign across multiple layers of the organization.</p><p>Six elements are essential:</p><ol><li><strong>A strategy that knows why autonomy matters. </strong>Not every workflow needs an agent. The work must benefit.</li><li><strong>Governance that builds trust and clarity. </strong>Guardrails should be explicit, auditable, and adaptive.</li><li><strong>A data foundation that is rich, real-time, and connected. </strong>Agents can only act as intelligently as the data they receive.</li><li><strong>A technology architecture built for interoperability. </strong>Not isolated tools — <strong>systems that speak to each other</strong>.</li><li><strong>A workforce supported to grow into new roles. </strong>People must be prepared, not simply “reassigned.”</li><li><strong>A change model that works in motion. </strong>The organization must learn while transforming.</li></ol><p>This is not a technical upgrade. It is an <strong>organizational reconfiguration</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OEJBHBbhBSf6Sd10Qz1pLA.png" /><figcaption>Full agentic scale application</figcaption></figure><h4>Measuring Progress Through Return-on-Autonomy</h4><p>To ensure autonomy creates value, the report introduces <strong>RoA — Return-on-Autonomy</strong>.</p><p>Rather than focusing solely on efficiency, RoA looks at:</p><ul><li>Speed of decision-making</li><li>Reduction in cycle times</li><li>Decrease in rework</li><li>Quality of outputs</li><li>Satisfaction and trust of those who interact with agents</li><li>Value is not just in <em>what</em> agents do, but in <strong>how the enterprise evolves as a result</strong>.</li></ul><h4>A Phased Path — Not a Leap</h4><p>The journey to an agentic enterprise unfolds gradually:</p><ul><li>Start with small pilot scenarios where the value and constraints are clear.</li><li>Expand where autonomy can manage sequences of actions, not just tasks.</li><li>Move toward processes and systems that are designed from the beginning to work with and through agents.</li></ul><p>Each step builds the foundations for the next. Autonomy compounds.</p><h4>Looking Ahead: From Automation to Human-Agentic Work</h4><p>The move toward the agentic enterprise is not simply about adding more intelligence to existing workflows. <a href="https://www.deloitte.com/global/en/services/consulting/perspectives/human-agentic-workforce.html">As the <em>Work, Reworked</em> report highlights</a>, <em>“agentic AI isn’t just a powerful new tool — it requires us to reimagine how work gets done.”</em><br> This is the inflection point: the moment where organizations must decide whether autonomy becomes <strong>a bolt-on layer</strong> to legacy processes — or the catalyst for a deeper re-architecture of how value is created.</p><p>Because without redesign, autonomy only scales inefficiency. To realize its full promise, organizations must shift:</p><ul><li>from <strong>technology-led change</strong> to <strong>human-centered transformation</strong></li><li>from <strong>task-based roles</strong> to <strong>outcome-based work design</strong></li><li>from <strong>headcount-dependent growth</strong> to <strong>scaled performance through orchestration</strong></li><li>from <strong>cultural inertia</strong> to <strong>organizational readiness as a strategic capability</strong></li></ul><p>This is what the reports converge on: autonomy is not the destination — it is the operating model that allows <strong>humans and AI agents to perform as one system</strong>.</p><p>The organizations that lead this transition will be the ones that <strong>redesign work from the ground up</strong>, creating environments where digital agents handle complexity and scale, while humans reclaim time, attention, and cognitive space for what truly requires judgment, creativity, and purpose.<br>The future of enterprise performance is not human <em>versus</em> machine, nor human <em>in service</em> of the machine. It is <strong>human-led, agent-enabled, and outcome-anchored</strong> work.</p><p>Those who begin now — deliberately, boldly, and with intention — will not only operate more efficiently.<br>They will shape <strong>a more adaptive, resilient, and meaningful system of work</strong>, where autonomy becomes the foundation for a new form of human potential.</p><p>This is the moment to move. Not incrementally, but decisively.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=69bfcd4508c9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Beyond Human-in-the-Loop: Embracing Human-on-the-Loop for Responsible AI Integration in…]]></title>
            <link>https://medium.com/@stefanobesana/beyond-human-in-the-loop-embracing-human-on-the-loop-for-responsible-ai-integration-in-80da88d4f032?source=rss-99ac926e66e4------2</link>
            <guid isPermaLink="false">https://medium.com/p/80da88d4f032</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[aritificial-intelligence]]></category>
            <category><![CDATA[human-in-the-loop]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[human-on-the-loop]]></category>
            <dc:creator><![CDATA[Stefano Besana]]></dc:creator>
            <pubDate>Thu, 12 Jun 2025 09:36:30 GMT</pubDate>
            <atom:updated>2025-06-12T09:36:30.033Z</atom:updated>
            <content:encoded><![CDATA[<h3>Beyond Human-in-the-Loop: Embracing Human-on-the-Loop for Responsible AI Integration in Organizations</h3><blockquote>“The question of whether artificial intelligence is possible no longer makes sense. What matters now are the unresolved questions about its nature and impact.”<br> — Daniel C. Dennett</blockquote><h4>A New Anthropological Turning Point</h4><p>Artificial Intelligence (AI) is far more than a technological evolution — it represents the most <em>profound anthropological challenge</em> humanity has ever faced. As Daniel Dennett reminds us, we have passed the point of asking whether AI is possible; we must now ask how it changes the human condition itself.</p><p><strong>AI is reshaping not only the mechanics of work but the very frameworks through which we interpret the world, communicate meaning, and construct reality.</strong> This transformation echoes the fundamental connection between technology and human cognition described by <em>Doreen Kimura</em>, who highlighted the interplay between fine motor skills and language processing in the brain, suggesting that our capacity for language and tool-making are deeply entwined.</p><p>This insight reveals why the rise of <em>Large Language Models (LLMs)</em> represents more than an IT shift; it’s a redefinition of how humans create, manipulate, and share meaning. <strong>Language is no longer only interpretive — it is becoming generative, synthetic, a force that challenges the very boundaries between natural and artificial cognition</strong>.</p><p>This same phenomenon is unfolding in the domain of images: photography, once an objective mirror of reality, now frequently generates realities that never existed. <strong>The AI-generated image breaks the chain of referential truth, opening spaces for ambiguity, creativity, and risk.</strong></p><p><strong>AI thus compels us to redraw the borders between natural and artificial, between knowing and computing, between delegation and accountability</strong>. As <em>Manuel Castells</em> foresaw, there can be no technological revolution without a cultural revolution.</p><p><strong>But how do we remain relevant in this changing landscape? How do we sustain and expand our uniquely human competencies when machines outperform us in domains we once claimed as ours? The challenge before HR leaders, educators, and strategists is cultural, not merely technical.</strong></p><p>We must adopt a holistic perspective — one that integrates historical, philosophical, sociological, and cognitive dimensions — to guide the transformation of work and organizations.</p><figure><img alt="Old Japan Retro Gaming studio" src="https://cdn-images-1.medium.com/max/1024/1*mI-tXNvG_jo3cKQz4SUtoQ.jpeg" /></figure><h4>The Evolutionary Continuum of Technology</h4><p>AI is not a sudden rupture in human history. As <em>Ernst Mach</em> suggested in 1905, humans are driven by a natural inclination to replicate what they understand — a tendency that has fueled every age of invention.</p><p>From <em>Promethean myths</em> to <em>Pascal’s mechanical calculators</em>, from ancient automata to modern AI systems, this drive to imitate and reproduce reality forms the thread of technological evolution. AI is simply the most radical, accelerated expression of this process.</p><p>As <em>Arnold Gehlen</em> argued, <strong>humans are biologically unfinished beings who externalize their functions into tools and environments. Technology is not an external accessory — it is an extension of humanity itself, a survival strategy embedded in culture</strong>. AI is now our most advanced epistemological prosthesis, reshaping how we perceive, decide, and learn.</p><p><em>Turing’s test</em> (1950) reframed AI as a question not of mechanics but of identity: if a machine’s behavior is indistinguishable from a human’s, where lies the boundary of consciousness? Later thinkers like <em>Roger Penrose</em>, <em>Paul Churchland</em>, <em>Barr &amp; Feigenbaum</em>, and <em>Gerald Edelman</em> pushed these questions deeper, blending philosophy, neuroscience, and cybernetics to explore AI’s potential and limits.</p><p>Edelman’s <em>Darwin III</em> and <em>Nomad</em> robots demonstrated learning not by programming but by environmental adaptation — a biological model of intelligence. Similarly, <em>Giuseppe Riva’s</em> concept of intelligent agents — autonomous, social, reactive, and proactive — revealed AI as an “other” in human space, no longer mere code but a new cognitive actor.</p><p>AI thus mirrors humanity’s hopes, fears, and contradictions. It is a prism through which we glimpse the tensions of evolution: knowledge vs. calculation, intuition vs. automation, creativity vs. replication.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4vZeaaW1uyWYyXrRln4vFg.jpeg" /></figure><h4>Reflection and Resistance: Human Cognitive Advantages</h4><p>This is why the human role in AI governance must shift. As <em>Donald Schön</em> argued, <strong>effective practitioners thrive in “indeterminate zones” — areas marked by uncertainty and conflict</strong>. Here, the key is not rigid technique but <em>reflection-in-action</em>: the ability to sense, question, and adjust amid ambiguity.</p><p><strong>For organizations, this means treating AI implementation as an ongoing adaptive process, not a one-time technical project</strong>. Leaders must continually ask: What unintended effects is AI producing? How is it changing team dynamics, decision-making, and power structures? What new skills are emerging — or being lost?</p><p><em>Reflection-on-action</em> extends this process, requiring periodic reassessment of practices, assumptions, and outcomes. HR leaders must promote feedback cultures, transparent policies, and ongoing learning.</p><p>Equally crucial is <em>Olivier Houdé’s</em> concept of “learning to resist.” In the age of LLMs and predictive algorithms, humans must resist the ease of heuristic, automatic thinking. AI’s outputs often appear authoritative, but critical scrutiny, source verification, and ethical judgment are essential.</p><p>This <em>cognitive resistance</em> safeguards human uniqueness: creativity, empathy, ethical reasoning, and critical analysis. Without it, we risk becoming passive validators of machine decisions.</p><h4>From Human-in-the-Loop to Human-on-the-Loop</h4><p>This leads to a necessary paradigm shift: <strong>from “human-in-the-loop” to “human-on-the-loop.”</strong></p><p><strong>“Human-in-the-loop” relegates humans to error-checkers within automated cycles — a reactive, diminished role. By contrast, “human-on-the-loop” envisions humans as <em>architects, strategists, and governors</em> of AI systems:</strong></p><ul><li>Defining AI’s <em>purpose and scope</em>;</li><li>Designing ethical, transparent architectures;</li><li>Interpreting outcomes in socio-cultural context;</li><li>Retaining ultimate <em>control and override authority</em>.</li></ul><p>HR’s role here is vital. AI deployment affects recruitment, learning, collaboration, performance management, and culture itself. HR leaders must ensure AI augments — not erodes — human potential, diversity, and agency.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DKisM2HSqYPLl8LV7CsC1A.jpeg" /></figure><h4>Organizational Implications: Strategy, Skills, and Ethics</h4><ol><li><strong>Strategy</strong>: AI integration demands cultural and structural adaptation, not just technical upgrades. HR co-leads this transformation.</li><li><strong>Skills</strong>: The <em>OECD</em> warns of widening skill gaps. Massive reskilling efforts are urgent to prevent inequality.</li><li><strong>Governance</strong>: As <em>Deloitte</em> and <em>McKinsey</em> report, most firms lack AI risk management capabilities. “Human-on-the-loop” models require robust policies.</li><li><strong>Culture</strong>: AI reshapes organizational identity. HR must address fears, resistance, and promote mindful adoption.</li><li><strong>Ethics</strong>: As <em>Kurzweil</em> provocatively noted, if machines mimic humans convincingly, ethical questions intensify. Responsible human agency is irreplaceable.</li></ol><h4>Towards a New Digital Humanism</h4><p>Ultimately, AI challenges us to renew <em>humanistic competencies</em>: metacognition, ethics, reflection, and resistance. These are not luxuries — they are survival skills in the AI era.</p><p>As <em>Antoine de Saint-Exupéry</em> wrote:</p><blockquote><em>“Technology does not keep man away from the great questions of nature — it forces him to study them more deeply.”</em></blockquote><p>We must rise to this challenge.</p><h4>Suggested Readings &amp; References</h4><ul><li>Accoto, C. (2017). <em>Il mondo dato: Cinque brevi lezioni di filosofia digitale</em>. Egea.</li><li>Accoto, C. (2023). <em>Il pianeta latente: L’intelligenza delle piattaforme come nuovo ordine del mondo</em>. Egea.</li><li>Besana, S. (2021). <em>The Future of Work: AI, People, and Organizational Change</em>. Hoepli.</li><li>Castells, M. (1996). <em>The rise of the network society</em>. Blackwell.</li><li>Churchland, P. M. (1984). <em>Matter and consciousness: A contemporary introduction to the philosophy of mind</em>. MIT Press.</li><li>Deloitte. (2024–2025). <em>Human Capital Trends 2024–2025</em>. Deloitte Insights.</li><li>Dennett, D. C. (1996). <em>Kinds of minds: Toward an understanding of consciousness</em>. Basic Books.</li><li>Edelman, G. M. (1987). <em>Neural Darwinism: The theory of neuronal group selection</em>. Basic Books.</li><li>Feigenbaum, E. A., &amp; Barr, A. (1981). <em>The handbook of artificial intelligence</em> (Vol. 1). Heuristech Press.</li><li>Gehlen, A. (1980). <em>Man in the age of technology</em>. Columbia University Press.</li><li>Houdé, O. (2019). <em>The psychology of intelligence</em>. Routledge.</li><li>Kurzweil, R. (1999). <em>The age of spiritual machines: When computers exceed human intelligence</em>. Penguin Books.</li><li>Mach, E. (1905). <em>The analysis of sensations</em> (C. M. Williams, Trans.). Dover Publications.</li><li>McKinsey &amp; Company. (2023). <em>The state of AI in 2023: Generative AI’s breakout year</em>. McKinsey Global Survey.</li><li>OECD. (2023). <em>Skills Outlook 2023: Skills for a resilient green and digital transition</em>. OECD Publishing.</li><li>Penrose, R. (1989). <em>The emperor’s new mind: Concerning computers, minds, and the laws of physics</em>. Oxford University Press.</li><li>Riva, G. (2004). <em>The psychology of cyberspace: The impact of the Internet on self and society</em>. Ios Press.</li><li>Saint-Exupéry, A. de. (2000). <em>Wind, sand and stars</em> (L. Galantière, Trans.). Mariner Books. (Original work published 1939)</li><li>Schön, D. A. (1983). <em>The reflective practitioner: How professionals think in action</em>. Basic Books.</li><li>Turing, A. M. (1950). Computing machinery and intelligence. <em>Mind, 59</em>(236), 433–460.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=80da88d4f032" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Human Capital Trends Deloitte 2025: the 8 key challenges that will transform the Future of Work]]></title>
            <link>https://medium.com/@stefanobesana/human-capital-trends-deloitte-2025-the-8-key-challenges-that-will-transform-the-future-of-work-7ad997d0409e?source=rss-99ac926e66e4------2</link>
            <guid isPermaLink="false">https://medium.com/p/7ad997d0409e</guid>
            <category><![CDATA[human-capital-management]]></category>
            <category><![CDATA[employee-engagement]]></category>
            <category><![CDATA[human-capital]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[employee-experience]]></category>
            <dc:creator><![CDATA[Stefano Besana]]></dc:creator>
            <pubDate>Tue, 25 Mar 2025 08:02:41 GMT</pubDate>
            <atom:updated>2025-03-25T08:02:41.071Z</atom:updated>
            <content:encoded><![CDATA[<p>The <a href="https://www2.deloitte.com/us/en/pages/human-capital/articles/human-capital-and-hr-trends-thought-leadership.html"><em>Human Capital Trends 2025</em> report</a> from Deloitte explores the emerging challenges that organizations are facing in today’s workforce landscape. Uncertainty and rapid change are putting traditional strategies to the test, pushing companies to redefine their approach to people management, the value of work, and the adoption of artificial intelligence (AI).</p><h4>Key Trends Identified</h4><ol><li><strong>Closing the Experience Gap</strong> — The gap between workers’ skills and companies’ needs is widening, impacting labor markets and talent acquisition strategies.</li><li><strong>What Moves Your People?</strong> — Personalizing motivation is crucial to enhancing human performance and engagement.</li><li><strong>AI Is Revolutionizing Work</strong> — AI is reshaping the value of work, requiring an update to the Employee Value Proposition (EVP).</li><li><strong>Is There Still Value in the Role of Managers?</strong> — The managerial role should not be eliminated but redefined strategically with AI support.</li><li><strong>Reinventing Performance Management</strong> — Traditional performance evaluation practices are ineffective at genuinely improving human performance.</li><li><strong>Reclaiming Organizational Capacity</strong> — Organizations must eliminate “work clutter” to free up capacity and focus on high-value activities.</li><li><strong>Stagility: Creating Stability for Workers While Enabling Organizations to Move at Speed</strong> — Stability is crucial for workers, but it must coexist with organizational agility.</li><li><strong>New Tech, New Work: Your Old Value Proposition Isn’t Enough</strong> — Traditional technology evaluation models no longer work: companies must redefine the value created by new work technologies.</li></ol><h4>Key Data Points (some of them)</h4><ul><li><strong>66%</strong> of managers believe new hires are not adequately prepared for the job.</li><li><strong>77%</strong> of workers say AI has increased their workload instead of reducing it.</li><li>Only <strong>6%</strong> of companies are truly investing in human sustainability as a business strategy.</li><li><strong>72%</strong> of business leaders believe work overload reduces overall efficiency.</li></ul><p>The data suggests that companies must find new strategies to address these challenges, balancing organizational efficiency with human development.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FKmdW5kNu6Is%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DKmdW5kNu6Is&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FKmdW5kNu6Is%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/46fb186fdba69bb5f8d1df9c7d919aa5/href">https://medium.com/media/46fb186fdba69bb5f8d1df9c7d919aa5/href</a></iframe><h3>Trend 1: Closing the Experience Gap</h3><p>The gap between the skills that companies demand and those available in the labor market is widening, making it harder to find qualified talent. At the same time, many workers struggle to find opportunities to gain the necessary experience, creating a cycle that negatively impacts both businesses and labor markets.</p><h4>The Problem</h4><ul><li><strong>Lack of experience among candidates</strong> — <strong>66%</strong> of managers believe new hires are not job-ready.</li><li><strong>Barriers to entry for young professionals</strong> — <strong>61%</strong> of companies have increased experience requirements for entry-level positions.</li><li><strong>High youth unemployment</strong> — The unemployment rate among graduates under 25 has reached <strong>21.3% in China</strong> and <strong>42% in India</strong>.</li><li><strong>Disappearance of formative roles</strong> — Automation is eliminating entry-level jobs, reducing learning opportunities on the job.</li></ul><h4>Solutions</h4><ul><li><strong>Apprenticeship and upskilling programs</strong> — Companies like Intel and Medtronic have created internal training paths to ease workforce entry.</li><li><strong>Lower experience requirements</strong> — Companies should value transferable skills rather than focusing solely on previous experience.</li><li><strong>AI-based simulations</strong> — Digital tools can accelerate practical learning.</li><li><strong>Structured career growth paths</strong> — Providing continuous professional development to bridge experience gaps.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yszMx0olr0_ZnpdGR9iRlg.png" /><figcaption>Experience gap in EVP</figcaption></figure><h3>Trend 2: What Moves Your People?</h3><p>Employee motivation is a key driver of performance and retention. However, only <strong>33%</strong> of employees feel that their managers truly understand what motivates them. Personalizing the work experience is becoming essential to enhancing engagement and productivity.</p><h4>The Problem</h4><ul><li><strong>Lack of personalized motivation strategies</strong> — Many companies still apply one-size-fits-all approaches.</li><li><strong>Unmet employee expectations</strong> — <strong>60%</strong> of workers expect more effort from companies in recognizing their personal needs.</li><li><strong>Shifting motivation drivers</strong> — <strong>38%</strong> of employees say their sources of motivation have changed in the past three years.</li><li><strong>Rigid incentive structures</strong> — Many corporate programs lack flexibility in benefits and work arrangements.</li></ul><h4>Solutions</h4><ul><li><strong>Manager-driven personalization</strong> — Unilever uses tailored development plans to boost engagement.</li><li><strong>AI-driven engagement strategies</strong> — Swissport leverages AI to personalize employee communications and incentives.</li><li><strong>Flexible work models and benefits</strong> — Companies like Land O’Lakes demonstrate the value of customized employee experiences.</li><li><strong>Continuous feedback and data analysis</strong> — Gathering insights to adjust corporate strategies to workforce needs.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EVODN47YVHSLEsCQPkBsRw.png" /><figcaption>What motivates people</figcaption></figure><h3>Trend 3: AI Is Revolutionizing Work</h3><p>AI is transforming the workplace by redefining roles, processes, and employee expectations. However, without proper management, AI risks being perceived as a threat rather than an opportunity. <strong>75%</strong> of companies are adopting AI technologies, but only a minority have developed strategies to support employees through the transition.</p><h4>The Problem</h4><ul><li><strong>Increased workload</strong> — <strong>77%</strong> of workers say AI has added to their workload instead of reducing it.</li><li><strong>Fear of job displacement</strong> — <strong>45%</strong> of employees worry that AI could make their roles obsolete without adequate reskilling opportunities.</li><li><strong>Lack of clear AI governance</strong> — Only <strong>30%</strong> of companies have defined policies on integrating AI with human labor.</li><li><strong>Limited access to AI benefits</strong> — Many companies invest in AI for operational efficiency but fail to distribute its benefits to workers.</li></ul><h4>Solutions</h4><ul><li><strong>Upskilling and reskilling programs</strong> — USAA invests in employee training to prepare them for AI-driven job roles.</li><li><strong>Equitable AI benefit distribution</strong> — Companies like Waste Management offer financial incentives based on AI-enhanced productivity.</li><li><strong>AI as an enabler, not a replacement</strong> — Amazon developed an AI coach to support employee professional growth.</li><li><strong>Redefining the EVP for the AI era</strong> — Integrating AI into the value proposition while balancing technology and human experience.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8Dti65VsxIyi0hPKYuPqKw.png" /></figure><h3>Trend 4: Is There Still Value in the Role of Managers?</h3><p>The managerial role is undergoing significant transformation. With the rise of AI and new organizational models, some companies are reducing the number of managers, while others are redefining their roles to be more strategic and less operational. <strong>78%</strong> of CEOs perceive high levels of uncertainty in business, increasing pressure on managers, and <strong>40%</strong> of them report experiencing burnout.</p><h4>The Problem</h4><ul><li><strong>Managerial overload and stress</strong> — <strong>40%</strong> of managers report high levels of burnout due to growing responsibilities.</li><li><strong>Need for a new leadership model</strong> — <strong>78%</strong> of CEOs believe that uncertainty requires a transformation in the role of managers.</li><li><strong>Difficulty in integrating AI</strong> — Many managers struggle to use AI as a support tool rather than seeing it as a threat.</li><li><strong>Rigid traditional management models</strong> — Excessive bureaucracy limits managers’ ability to focus on strategy and people development.</li></ul><h4>Solutions</h4><ul><li><strong>AI as a decision-making assistant</strong> — Goldman Sachs uses AI to support managers in mentoring and strategic decision-making.</li><li><strong>Hierarchy simplification</strong> — BMW has experimented with self-organized teams with fewer management layers.</li><li><strong>Focus on coaching and people development</strong> — Managers need to be trained to become growth facilitators rather than mere supervisors.</li><li><strong>Redesigning performance management</strong> to adapt it to new leadership and autonomy needs.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SJZTugBf6o5nPbyctsVeCw.png" /><figcaption>The value beneath managers’ roles</figcaption></figure><h3>Trend 5: Reinventing Performance Management</h3><p><strong>57%</strong> of companies have reformed their performance management processes in the past three years, yet only <strong>22%</strong> believe they are truly effective. Traditional annual evaluations no longer meet the needs of a rapidly evolving work environment. Companies must shift their focus from simple performance monitoring to continuous growth and development.</p><h4>The Problem</h4><ul><li><strong>Outdated evaluation systems</strong> — <strong>57%</strong> of companies have updated their processes, but <strong>78%</strong> of employees still find them ineffective.</li><li><strong>Lack of continuous feedback</strong> — <strong>65%</strong> of employees prefer ongoing feedback instead of annual reviews.</li><li><strong>Short-term performance focus</strong> — Many performance management systems do not consider long-term professional development.</li><li><strong>Limited managerial involvement</strong> — Without proper tools, managers struggle to provide effective support for team growth.</li></ul><h4>Solutions</h4><ul><li><strong>Continuous and personalized feedback</strong> — Amazon uses AI to provide real-time feedback to employees.</li><li><strong>Impact-based evaluations rather than just productivity</strong> — Deloitte has developed new performance measurement models that also consider employee well-being.</li><li><strong>Coaching and development approaches</strong> — Successful companies are investing in mentorship programs to grow employee skills.</li><li><strong>Using AI for performance tracking</strong> to ensure fair and objective assessments.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dSrHsRfkWNajJY74otN6XA.png" /><figcaption>A new way to manage performances</figcaption></figure><h3>Trend 6: <strong>Reclaiming Organizational Capacity</strong></h3><p>Organizations often confuse productivity with workload volume, leading to an accumulation of low-value tasks and operational inefficiency. <strong>72%</strong> of business leaders believe that excessive workload reduces overall effectiveness, and smarter workload management could free up resources for high-value projects.</p><h4>The Problem</h4><ul><li><strong>Excessive administrative and bureaucratic tasks</strong> — <strong>68%</strong> of employees spend more time on coordination and reporting than on strategic projects.</li><li><strong>Too many unnecessary meetings and communications</strong> — <strong>35%</strong> of work time is wasted in non-essential meetings and redundant emails.</li><li><strong>Lack of efficient work tools</strong> — <strong>60%</strong> of workers report inadequate digital tools for improving productivity.</li><li><strong>Difficulty in prioritizing tasks</strong> — Only <strong>42%</strong> of employees say they have a clear understanding of their company’s top priorities.</li></ul><h4>Solutions</h4><ul><li><strong>Zero-based work approach</strong> — Applying zero-based budgeting principles to work to eliminate redundant activities and focus on high-impact tasks.</li><li><strong>Time management optimization</strong> — Google has reduced internal meetings by <strong>30%</strong>, significantly improving productivity.</li><li><strong>Automation of repetitive tasks</strong> — New technologies allow the reduction of time spent on administrative and bureaucratic work.</li><li><strong>Clearer corporate prioritization</strong> — Organizations must help employees identify high-impact tasks and provide guidelines to improve operational efficiency.</li></ul><h3>Trend 7: Stagility — Creating Stability for Workers While Enabling Organizations to Move at Speed</h3><p>The need for organizational agility often clashes with employees’ desire for stability. The concept of <strong>“stagility”</strong> (stability + agility) is key to balancing these two needs, ensuring job security for employees while allowing companies to adapt quickly to market changes.</p><h4>The Problem</h4><ul><li><strong>Growing job insecurity</strong> — <strong>54%</strong> of employees fear that continuous corporate restructuring may put their jobs at risk.</li><li><strong>Lack of career transparency</strong> — Only <strong>40%</strong> of companies offer clear career development paths, increasing employee uncertainty.</li><li><strong>Struggles to adapt to changes</strong> — <strong>48%</strong> of workers find it difficult to keep up with evolving market skills.</li><li><strong>Rigid work models</strong> — <strong>62%</strong> of companies recognize the need for hybrid models but struggle to implement them effectively.</li></ul><h4>Solutions</h4><ul><li><strong>Structured benefits and career paths</strong> — Companies should provide clear career growth plans and job stability.</li><li><strong>Hybrid work models</strong> — <strong>60%</strong> of employees prefer a mix of remote and in-office work to balance productivity and well-being.</li><li><strong>Continuous training and professional development</strong> — Investing in workforce upskilling reduces uncertainty and boosts engagement.</li><li><strong>Flexible organizational structures</strong> — Adopting adaptive models to ensure agility without compromising job security.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Q7A27qLgdNyiCux1JQsFTA.png" /><figcaption>Managing agility and stability at once</figcaption></figure><h3>Trend 8: New Tech, New Work: Your Old Value Proposition Isn’t Enough</h3><p>The adoption of new technologies is transforming the workplace, yet companies often implement innovations without a clear strategy on how they will impact employee value. Technology cannot be assessed solely based on financial returns — it must also consider its effects on collaboration, well-being, and professional growth.</p><h4>The Problem</h4><ul><li><strong>Lack of a clear tech strategy</strong> — <strong>67%</strong> of companies invest in new technologies without a roadmap on their workforce impact.</li><li><strong>Digital skill gaps</strong> — <strong>52%</strong> of employees believe that the training they receive on new technologies is insufficient.</li><li><strong>Exclusive focus on operational efficiency</strong> — Many technology investments are measured only in terms of cost savings, neglecting their impact on work quality.</li><li><strong>Resistance to change</strong> — <strong>49%</strong> of companies struggle to integrate new technologies without negatively affecting employee morale and engagement.</li></ul><h4>Solutions</h4><ul><li><strong>Redefining technology’s value</strong> — Success metrics should include creativity, collaboration, and employee well-being.</li><li><strong>Upskilling and digital literacy programs</strong> — Companies must invest in training paths to ensure employees can fully leverage new technologies.</li><li><strong>Smart AI integration</strong> — AI should be used to enhance the work experience, reducing repetitive tasks and allowing employees to focus on high-value activities.</li><li><strong>Human-centered business case evaluation</strong> — Every technology investment should be assessed for its impact on people, not just financial returns.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mBbnpFNELPlZszGw1QaiSQ.png" /><figcaption>New tech integration should be possible</figcaption></figure><p>Organizations must navigate complex challenges: from AI adoption to motivation strategies, from the need for agility to ensuring worker stability. Companies that can balance these elements and invest in sustainable innovation will be the ones leading the future of work.</p><h4>Note on Methodology</h4><p>The report is based on a global analysis of <strong>over 10,000 leaders, managers, and employees</strong> from <strong>100+ countries</strong>. The study was conducted using <strong>quantitative surveys</strong> to collect statistical insights and <strong>qualitative interviews</strong> to understand perceptions, challenges, and expectations. The goal was to identify the key trends reshaping the workforce and provide strategic insights for organizations to navigate them successfully.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7ad997d0409e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The International AI Safety Report 2025: Key Insights and Implications]]></title>
            <link>https://medium.com/@stefanobesana/the-international-ai-safety-report-2025-key-insights-and-implications-e1ff12626f97?source=rss-99ac926e66e4------2</link>
            <guid isPermaLink="false">https://medium.com/p/e1ff12626f97</guid>
            <category><![CDATA[change]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[gen-ai-for-business]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-security]]></category>
            <dc:creator><![CDATA[Stefano Besana]]></dc:creator>
            <pubDate>Mon, 10 Feb 2025 08:02:06 GMT</pubDate>
            <atom:updated>2025-02-10T08:02:06.370Z</atom:updated>
            <content:encoded><![CDATA[<p>Artificial Intelligence (AI) is rapidly transforming industries, societies, and economies. As AI systems become more advanced, so do their risks and implications. <a href="https://www.gov.uk/government/publications/international-ai-safety-report-2025">The <strong>International AI Safety Report 2025</strong> analyzes the capabilities, risks, and potential mitigation strategies for general-purpose AI</a>. This post breaks down its key findings, covering:</p><ul><li>The <strong>accelerating capabilities of AI</strong></li><li>The <strong>risks associated with malicious use</strong>, malfunctions, and systemic challenges</li><li>The <strong>current state of AI risk management</strong> and mitigation techniques</li><li>The <strong>broader implications for policymakers and businesses</strong></li></ul><p>This report is the result of contributions from <strong>96 international AI experts</strong> across academia, government, and industry, reflecting the latest understanding of AI risks and safety mechanisms. It was developed as part of an ongoing effort to create a shared international approach to AI governance and safety.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OsZaFqix-PDKCCJPyeKNyA.png" /><figcaption>International AI Safety Report</figcaption></figure><h4>Capabilities of General-Purpose AI</h4><p>AI capabilities have progressed at an extraordinary pace, with recent advances surpassing previous expectations. Five years ago, AI models struggled to generate coherent paragraphs. Today, they can:</p><ul><li><strong>Write and debug code</strong> with increasing precision, assisting developers in automating complex programming tasks and identifying potential vulnerabilities before deployment.</li><li><strong>Engage in complex conversations</strong> across multiple languages with improved contextual awareness, emotional intelligence, and response coherence, making them valuable for customer service, therapy, and content creation.</li><li><strong>Generate realistic images, videos, and audio</strong> that can be indistinguishable from human-created content, raising both creative opportunities and ethical concerns, particularly in the fight against misinformation and AI-generated fraud.</li><li><strong>Perform scientific reasoning and mathematical problem-solving</strong> at expert levels, supporting research in physics, chemistry, and medicine, accelerating the discovery of new drugs, and assisting in climate modeling.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZEeHHRpe1zfTWYP8XCHbQg.png" /><figcaption>AI Performance compared to PhD-level experts</figcaption></figure><blockquote>Over 90% of online content could be AI-generated by 2030, raising concerns about authenticity, trust in media, and election security.</blockquote><h4>Scaling and Future Advancements</h4><p>The rapid advancement of AI is largely driven by <strong>scaling laws</strong>, where increasing the amount of data and computational power used to train AI models leads to greater performance improvements.</p><ul><li>The most powerful AI models today use <strong>trillions of parameters</strong>, far surpassing early deep learning models from a decade ago, allowing for highly nuanced and contextual responses.</li><li>Training datasets have grown exponentially, now comprising vast amounts of internet text, academic papers, and specialized domain knowledge, ensuring AI models develop a more comprehensive understanding of complex topics.</li><li>The availability of <strong>high-performance computing (HPC) infrastructure</strong>, including GPU clusters and cloud-based AI accelerators, has allowed AI developers to push the boundaries of model performance and efficiency, reducing processing time and energy consumption.</li></ul><p>If current trends continue, AI models by 2026 could use <strong>100x more compute than in 2023</strong>, and by 2030, <strong>10,000x more compute</strong>. However, the feasibility of such growth depends on factors such as the availability of semiconductor technology, energy consumption constraints, and regulatory measures on AI development. Additionally, the increased cost and resource demand could make high-level AI capabilities accessible only to a few dominant players, further centralizing power.</p><p>Companies are also exploring <strong>“inference scaling”</strong>, allowing AI to analyze problems with longer chains of thought, which could lead to even more sophisticated reasoning abilities, particularly in multi-step problem-solving, strategic planning, and predictive analytics. These advancements raise fundamental questions about the <strong>alignment of AI with human values and oversight mechanisms</strong>, emphasizing the need for robust AI governance structures.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oDzhVzyk-SS3YJkAP9XCtw.png" /><figcaption>Level of perfomances of AI in 2024</figcaption></figure><h4>Risks Associated with General-Purpose AI</h4><p>As AI capabilities grow, so do concerns regarding its potential misuse and unintended consequences. The report categorizes AI risks into three primary areas:</p><p><strong>Malicious Use Risks<br></strong>AI can be exploited by bad actors to cause harm on a large scale. Specific threats include:</p><ul><li><strong>Fake content and deepfakes</strong>: AI-generated media can be used for misinformation, fraud, or blackmail. Some studies estimate that <strong>over 90% of online content</strong> could be AI-generated by 2030, raising concerns about authenticity, trust in media, and election security.</li><li><strong>Manipulation of public opinion</strong>: AI-driven propaganda campaigns can influence elections, policy decisions, and social movements. Past events have shown how bot networks and AI-generated content have been used to distort political discourse, increasing polarization and undermining democracy.</li><li><strong>Cyberattacks</strong>: AI is increasingly being leveraged in cybersecurity, both for defensive measures and for more advanced forms of hacking. Automated vulnerability detection and exploitation tools powered by AI could pose serious threats to digital infrastructure, as seen in rising cases of AI-assisted phishing attacks and identity theft.</li><li><strong>Biological and chemical threats</strong>: AI models have demonstrated the ability to assist in designing harmful biological agents, raising concerns about potential misuse by non-state actors and rogue organizations. In recent tests, some AI systems were able to generate <strong>step-by-step instructions</strong> for creating hazardous materials, escalating the need for stricter oversight in AI research applications.</li></ul><p><strong>Malfunction Risks<br></strong>Even when not used maliciously, AI can cause unintended harm due to design flaws, incomplete training data, or unpredictable behavior. These risks include:</p><ul><li><strong>Reliability issues</strong>: AI systems sometimes produce false or misleading information, particularly in high-stakes fields like medicine, law, and finance. Studies show that <strong>AI-generated medical advice</strong> still carries a 5–10% error rate, which could lead to life-threatening consequences if not properly validated, necessitating human oversight in critical decision-making areas.</li><li><strong>Bias and discrimination</strong>: AI models often reflect and amplify biases present in their training data. This has led to documented cases of AI-driven discrimination in areas such as hiring, lending, and law enforcement, where unchecked bias can perpetuate systemic inequalities.</li><li><strong>Loss of control</strong>: While current AI models lack autonomy, experts warn that future systems could become increasingly difficult to predict and control. The concept of <strong>“AI alignment”</strong> aims to ensure AI systems behave as intended, but remains an open research challenge, particularly as AI develops self-learning and decision-making capabilities.</li></ul><blockquote>AI-driven automation is expected to replace millions of jobs, requiring large-scale workforce reskilling initiatives to mitigate economic displacement</blockquote><p><strong>Systemic Risks<br></strong>Beyond individual AI failures, systemic risks emerge when AI adoption scales across industries and institutions:</p><ul><li><strong>Labor market disruption</strong>: AI-driven automation is expected to replace <strong>millions of jobs</strong>, requiring large-scale workforce reskilling initiatives to mitigate economic displacement.</li><li><strong>Concentration of AI power</strong>: A few companies dominate AI development, raising concerns about monopolization, competitive fairness, and geopolitical dependencies, leading to imbalances in technological access and decision-making power.</li><li><strong>Environmental costs</strong>: AI training consumes vast amounts of energy, with some estimates suggesting that training a single <strong>large AI model produces as much CO₂ as five cars over their lifetime</strong>, necessitating greener computing solutions and regulatory interventions.</li><li><strong>Privacy and data security</strong>: AI’s ability to infer sensitive information from seemingly unrelated data sources poses new challenges for data protection laws and individual privacy rights, calling for updated legal frameworks.</li><li><strong>Copyright and intellectual property disputes</strong>: Lawsuits over AI-generated content are rising, as traditional copyright frameworks struggle to keep pace with AI’s ability to replicate artistic and literary styles, raising questions about ownership and fair use policies.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OaQK8Z6B03qnewLY1h22jg.png" /><figcaption>Data Center usage with AI Revolution</figcaption></figure><blockquote>Training a single large AI model produces as much CO₂ as five cars over their lifetime, necessitating greener computing solutions and regulatory interventions.</blockquote><h4>Risk Management and Mitigation Strategies</h4><p>Expanded risk management strategies focus on enhanced transparency, regulatory compliance, and AI ethics. Governments and industries must develop <strong>comprehensive frameworks</strong> that enforce responsible AI use, ensuring models are tested rigorously before deployment. <strong>International cooperation</strong> will be essential in establishing unified standards, especially in managing cross-border risks like cyber threats and misinformation campaigns.</p><p>Investment in <strong>AI safety research</strong> must increase, focusing on interpretability, bias mitigation, and adversarial robustness. Organizations should implement <strong>continuous monitoring systems</strong> to track AI behavior and intervene when necessary. <strong>Public-private partnerships</strong> can help create independent oversight bodies, ensuring accountability in AI development.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fyYwWrr72IT940Ck3PwNQw.png" /><figcaption>Risks from General Purpose AI</figcaption></figure><h4>The Future of AI and Policy Considerations</h4><p>AI governance, ethical development, and regulatory standards will play crucial roles in shaping how AI is integrated into society. Policymakers must balance <strong>innovation with safety</strong>, ensuring that AI serves public interests while minimizing potential harm. <strong>Legislation on AI transparency and accountability</strong> should evolve in tandem with technological progress, preventing misuse while fostering beneficial advancements.</p><p>Future policies should focus on <strong>democratizing AI access</strong>, preventing monopolistic control over AI technologies, and promoting <strong>sustainable AI development</strong> to mitigate environmental impacts. AI education initiatives will be critical to preparing the workforce for the evolving job landscape, ensuring that societies can adapt to AI-driven economic shifts.</p><blockquote>AI does not happen to us; it is shaped by the choices we make today. Ensuring a future where AI serves humanity requires vigilance, cooperation, and a commitment to responsible innovation.</blockquote><p>The <strong>International AI Safety Report 2025</strong> serves as a crucial resource in understanding the evolving AI landscape. The choices societies and governments make today will determine whether AI remains a force for good or becomes a major global challenge.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e1ff12626f97" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>