The debate on the adoption of artificial intelligence within organisations moves, with few exceptions, along two lines that are by now well established. The first is automation: which activities the machine can take on, and with what savings. The second is cognitive augmentation: how AI broadens people’s capacity to understand, anticipate, create and decide. Most adoption frameworks, productivity metrics, surveys on tool usage and, to a large extent, the managerial imaginary surrounding AI are organised around these two lines.

Reshuffle. Who Wins When AI Restacks the Knowledge Economy by Sangeet Paul Choudary has the merit of forcefully introducing a third dimension, which those working in AI Adoption should not, in my view, treat as a marginal note: coordination. The book’s thesis is that the deepest value of next-generation AI lies not in the acceleration of tasks, but in reducing the costs with which an organisation aligns fragmented activities, people, data and systems. For Choudary, coordination is the dimension that explains why AI transforms organisations and not merely job roles.

Having read this book, which I found highly stimulating even though I do not share some of its positions, I should like to put forward in this article three points that strike me as significant and that may enrich the discussion devoted to AI Strategy and AI Ethics. First, to give an account of Choudary’s thesis in its most robust part: the way AI enables coordination within an organisation, which I regard as the book’s most useful contribution for those governing adoption processes. Second, to propose a more nuanced position than his: Choudary is so trenchant as to appear, at times, indifferent to automation and cognitive augmentation, whereas the available empirical evidence suggests that the three dimensions operate jointly, and that coordination may be the condition under which the other two, grasped at a systemic level, cease to dissipate value, rather than being an alternative to them. Third, to show that the coordination dimension carries with it a specific ethical question, distinct from those raised by automation, one that invokes the notion of biopolitical power and that lends itself to being identified in certain precise articles of the European AI Act.

A preamble on scope. Reshuffle develops the coordination thesis also on an economic-systemic scale (the shift of value from ownership to the orchestration of what one does not own, the struggle over control points in the AI value chain) and on the transformations of the form of work, from the crisis of the cognitive augmentation paradigm to judgement as the residual human advantage. These are analytical planes that would merit dedicated treatment and on which, especially when the author considers the transformations of the professions and of the labour market, I would have more than a few critical reservations; here, however, they will remain in the background. The focus of this piece is AI Adoption: what changes, for those designing the adoption of AI within an organisation, when coordination is added to the dimensions of automation and cognitive augmentation.

Three dimensions, two axes

It is worth first bringing some conceptual order, because the three dimensions do not stand epistemologically on the same plane. Automation and cognitive augmentation are symmetrical categories: both describe how the execution of an activity changes, and they differ in the allocation of action between human and machine. It is the most well-established opposition in the literature on human-AI collaboration, substitution versus augmentation, which we find in Stanford HAI’s definitions of human-centered AI, in NIST’s spectrum of human-machine configurations, and in the OECD’s distinction between AI that substitutes for human activities and AI that improves their performance. On this topic, see also my earlier contributions bridging the theory of the extended mind and System 0. Coordination belongs to a different order: it does not describe the performance of an agent or of a human-machine pair, but the structure of the interdependencies among multiple activities, datasets, people, systems and organisations. It is a systemic and organisational concept, not a performance-related one.

The more rigorous formulation I propose is therefore not a taxonomy of three parallel categories, but a matrix with two axes. On the first axis lies the transformation of execution, which breaks down into automation (AI takes on a share of activity or decision-making) and cognitive augmentation (AI broadens the human capacity to understand, anticipate, create and decide). On the second axis lies the transformation of interdependencies: coordination, which in turn unfolds on a local scale (the handover between adjacent activities), an organisational one (the linking of functions, data and workflows) and an ecosystemic one (interoperability among firms, platforms and markets).

An example makes the distinction concrete. Faced with the assessment of a credit application, an organisation can automate the assessment, proposing or taking the decision without substantial human intervention; it can cognitively augment the analyst, summarising documents, identifying anomalies, generating scenarios; or it can coordinate the flow of information and decisions among the analyst, compliance, risk management, external data sources and operational systems. The first two options answer the question: who performs the activity, and with what capabilities? The third answers a question of a different nature: how are activities that are distributed among different actors made compatible and chained together?

This arrangement has two consequences that we will draw on throughout the article. The first: one and the same implementation can be simultaneously automating, augmenting and coordinating, with the three dimensions taking shape not as alternative outcomes, but as value-creation mechanisms operating at different levels. The second: automation and cognitive augmentation explain how AI alters individual performances; coordination explains how those alterations become organisational transformation. This is exactly the point at which Choudary’s thesis is to be embraced and, as we shall see, also the point at which, in my view, it should be tempered.

The coordination tax

The organisational diagnosis from which Reshuffle sets out is what Choudary calls the coordination tax: the hidden cost that every organisation pays to codify knowledge, organise it and put it to work. It is a cost that grows non-linearly with the number of actors to be aligned, and that no previous technology managed to bring down for a precise reason: none could handle unstructured information. The technological waves of the past improved the external coordination of work, such as logistics, transactions and communication, while leaving the internal architecture of knowledge management substantially intact: information locked in silos, accessible to those who know where to look, embedded in the roles and teams that generated it.

Fragmentation, after all, is not a pathology of contemporary work: it is its norm. Tools, teams and flows proliferate for legitimate reasons such as specialisation, autonomy and local adaptation, but every historical attempt to unify them from above, by imposing common platforms or shared standards, has run up against prohibitive migration costs and the rational resistance of teams to relinquishing the tools on which they have built their own practices. The result is one that every manager recognises: alignment overload, the manual patchwork of stitching together disconnected systems, knowledge that exists somewhere in the organisation but does not reach where it would be needed.

It is on this historical asymmetry that generative AI intervenes, and here Choudary’s reading is convincing. The distinctive capability of next-generation models lies not so much in intelligence as measured against human benchmarks as in the handling of unstructured information: converting it into structured knowledge, making the tacit explicit by rendering it searchable and reusable, decoupling organisational knowledge from the roles and teams that produced it and, the decisive step, proactively injecting it into the relevant workflows. Adopted in this light, AI does not optimise existing knowledge management but changes its nature, transforming knowledge from an archive to be queried into a resource that reaches whoever needs it.

Coordination without consensus

The book’s theoretical core, however, lies one level deeper, in the relationship between coordination and consensus. Historically, coordinating fragmented systems required one of two conditions: a dominant actor able to impose structure from above by controlling access to demand; or a prior agreement on standards and protocols, as in the containerisation of freight transport. In both cases, consensus, always either hierarchical or negotiated, preceded coordination.

AI introduces a third way, which Choudary considers the true discontinuity: it learns to understand each actor in its native format, extracts structure from unstructured sources and builds a unified representation ex post, dispensing with the constraint of the initial agreement. The composition of heterogeneous systems ceases to be a preliminary infrastructural requirement and becomes an emergent capability of the algorithmic translation layer. Consensus does not disappear; it shifts in logical position. From a precondition of coordination it becomes its deferred outcome, because actors converge towards shared practices by attraction, in response to the value generated, not by prior obligation. This is what Choudary calls AI-driven rebundling: functional integration dissociates from formal unification, and ecosystems function as if they were integrated even when they are not.

Effective coordination, in the reconstruction offered by Reshuffle, is articulated across five operating levers: unified representation, which establishes a common way of perceiving the environment; aligned decision-making, which turns conflicting signals into operational judgement under conditions of imperfect information; execution, whether assistive or agentic depending on where agency resides; composition, which connects heterogeneous actors into shared flows; and governance, which with AI ceases to be an ex ante norm and becomes an ex post outcome, the product of the recursion among observation, feedback and re-orientation. This fifth lever is well worth keeping in mind: it is the one on which, in closing, the ethical question will centre.

Inside the organisation: how AI enables coordination

What does it mean, concretely, to adopt AI as an infrastructure of coordination rather than as an exclusive tool of automation? The difference, viewed through Choudary’s lens, can be observed on at least four planes.

The first is the overcoming of the recurrent failure of unification projects. AI-driven rebundling unifies fragmented information, decisions and execution without requiring shared standards or platform migrations. Each team retains its own tools, and AI acts as a bridge, translating between formats to keep everyone aligned to the same evolving reality. It is a way out of the alternative that has condemned generations of transformation programmes: either the top-down imposition of a single platform, or resignation to fragmentation.

The second plane is the dissolution of a trade-off that organisational theory has always treated as structural: that between autonomy and coordination. If coordination passes through the system rather than over the teams, in the form of hierarchy, meetings and reporting, teams can be simultaneously autonomous and aligned. Coordination ceases to be a function that consumes autonomy and becomes a property of the infrastructure.

The third plane is the inversion of the flow of knowledge work. In the inherited model, it is the knowledge user who has to move: to search the repositories, query colleagues, manually stitch together disconnected tools. In the coordinative model the work reaches the user: the relevant knowledge is injected at the point in the flow where it is needed, in the format in which it is needed.

The fourth plane is the temporal depth of the effects, which Choudary arranges across three orders. First-order effects are those of automation: tasks executed more quickly and at lower cost. Second-order effects are the coordinative cascade: the gains propagate along the interdependencies among actors and workflows. Third-order effects are recursive learning: the organisation does not merely evolve under the effect of AI, but co-evolves with the system that coordinates it, in a cycle in which each interaction improves the shared representation and each improvement of the representation improves the interactions. The full value of adoption resides in the third order, and is accessible only to those who have invested in the second.

Corresponding to these four planes, conversely, is the risk that Choudary calls the coordination paradox. When AI is implemented to automate isolated activities in the functions with the most resources and the most urgency, it introduces a disparity of capabilities among functions that generates new friction precisely at the points of interface: teams that decide with sophisticated predictive tools have to synchronise with teams working on the previous quarter’s reports. Fragmented adoption does not merely fail to improve coordination, but breaks down the coordination that already exists, and the increase in the coordination tax can outweigh any local efficiency gain. This produces hyper-efficient teams within organisations that are, overall, more fragile. It is the difference, on which the author insists, between task-centric firms, which graft AI onto isolated activities and encounter diminishing returns because the productivity from shared tools becomes commoditised, and system-centric firms, which reconsider the architecture of the way work is coordinated and decisions are taken.

What the data say

If the coordination thesis is correct, we should observe two things empirically: that individual gains from automation and augmentation, on their own, do not necessarily convert into stable organisational value; and that the organisations which redesign their flows by working on the coordinative dimension systematically capture more value. The most recent surveys from Boston Consulting Group offer a remarkably precise corroboration on both fronts.

The fourth edition of the AI at Work survey (June 2026, 11,749 respondents across 14 markets) certifies first of all that the phase of access to tools is complete: 74% of frontline employees describe themselves as regular AI users, an increase of 23 percentage points in a year. But the most instructive figure concerns the strategic deficit: among regular frontline users, 42% save at least a full working day per week, and yet 66% report receiving limited or no guidance on what to do with the time saved. The time freed up by automation and individual augmentation, in other words, evaporates at the boundaries of the task. It is neither redirected into other activities, nor into training and career-advancement projects, nor even into work-life balance projects. This is time given back to people and organisations that no one, at present, is choosing to put to use.

The mirror-image confirmation comes from the tripartition with which BCG classifies adoption strategies: Deploy (rolling out tools and supporting individual productivity), Reshape (redesigning end-to-end workflows and processes), Invent (building new business models). The organisations engaged in Reshape or Invent initiatives pull ahead of those stuck at Deploy on every metric: +22 percentage points of employees who save at least one day per week (53% versus 31%), +24 points in the perception of measurable business impact (67% versus 43%), +20 points in employee job satisfaction (68% versus 48%). And the deciding factor is not the tools: among those with strategic clarity but limited access to tools, 80% report measurable impact; among those with full access to tools but weak strategic direction, 60%. A significant number of percentage points thus separate strategy from technological provision, hence the title of the survey, Strategy Matters More Than Tools.

The picture is completed by the data on AI agents, the technological form in which the coordinative dimension is materialising: the integration of agents into workflows has more than doubled in a year (from 13% to 30%), and 61% of respondents believe that agents could carry out half of their own work within three years. But half of the employees state that their company has neither clear governance for managing human-AI teams nor a defined strategy for developing such governance.

A second BCG contribution, Agentic AI Turns Every Team into Its Own Transformation Engine (June 2026), adds an element concerning the very form of the transformation: with agentic AI the adoption bottleneck has shifted from technology to organisational capability, and change no longer scales through a few large central programmes but through many local reinventions conducted in parallel by the teams that hold the domain expertise. The AWS case documented there is instructive: engineers’ access to GenAI tools did not in itself alter either the way of working or the results; the value came when the transformation was tied to business outcomes and conducted by the teams themselves across the entire software life cycle, with a measured result of 27% more features released.

A more nuanced position

These data confirm the coordination thesis, but they confirm it, on closer inspection, in a less trenchant version than Choudary’s. And it is here that I should like to strike a more nuanced critical note.

In Reshuffle automation is systematically treated as the surface dimension of the phenomenon: the “first-order effects” on which public debate dwells, the naive promise as opposed to the authentic one. Cognitive augmentation fares even worse: its paradigm is declared to be in crisis, because tasks may persist while the logic that aggregated them into professions vanishes. There is, in this hierarchy, an understandable rhetorical move (the book has to win space for a neglected dimension!) but also an analytical risk: replacing one reduction with another, reading the whole phenomenon through coordination just as one previously read it all through automation.

The empirical evidence suggests a different relationship among the three dimensions, and the two-axis matrix I proposed at the outset makes it possible to formulate it precisely. The BCG data do not say that automation and cognitive augmentation produce no value: they say that they produce a great deal of it and that this value dissipates in the absence of strategy and of the coordinative dimension. The 42% of frontline workers who save a day a week is a fully real joint effect of automation and cognitive augmentation; the 66% who do not know what to do with it is a deficit of strategy and coordination. The Reshape organisations have not given up on automating: they have embedded automation within a redesign of interdependencies. In the AWS case too the order of factors is clear: the tools for augmenting the individual engineer were already there; it was the coordinative redesign of the flow that converted them into output.

The formulation that seems to me most defensible is therefore this: the organisational value of AI manifests itself through at least three mechanisms which are not, however, isomorphic: the automation of activities, the cognitive augmentation of performance, the reduction of coordination costs. The first two transform local performance and answer the question of who performs the activity and with what capabilities; the third transforms the architecture of interdependencies and answers the question of how distributed activities become a system. Automation and cognitive augmentation explain how AI alters individual performances, while coordination explains how those alterations become organisational transformation or, in its absence, how they fail to become it.

For those governing AI Adoption, the practical consequence is clear-cut but is not the maximalist one of subordinating every initiative to systemic redesign. It is rather a criterion for combining the options: every automation or cognitive augmentation initiative should be assessed on the second axis as well, asking which interdependencies it touches, which disparities of capability it introduces at the interface points, and within which flow redesign its local gains will become system gains. The mistake is to automate as if the task were an island.

The biopolitical underside of coordination

One final step remains, which I regard as indispensable precisely for those who read coordination as an organisational possibility. The coordinative dimension carries with it a specific ethical question, distinct from the serious ones already raised by automation and cognitive augmentation. Coordination poses the problem of who defines the categories through which the organisation knows and governs.

Choudary himself developed this side of the thesis in a piece written after the book (Reshuffle and biopolitical power, September 2025), which functions as an internal self-critique of Reshuffle’s most celebrated formula. Coordination without consensus, he observes, holds in two senses that are worth keeping distinct. The first is technical-organisational: AI coordinates systems that have not agreed standards and semantics in advance; it is interoperability without standardisation, the source of value we have discussed up to this point. The second is political: AI can coordinate people and institutions by autonomously inferring equivalences, risks, intentions and affiliations from ordinary and heterogeneous data, without the subjects concerned ever having consented to the inferences through which they will be known, classified and treated. The two phenomena are connected but do not coincide: a system can translate between two databases without significant biopolitical consequences, and an automated decision can exercise power even in perfectly standardised environments. The critical shift occurs when informational integration becomes the capacity to take or to shape decisions about people.

The reference to biopolitics is not ornamental. In the Foucauldian notion, biopower operates not primarily through explicit prohibitions or coercions, but through the knowledge and administration of populations: classifying, differentiating treatments, distributing opportunities, configuring the environment of possibilities rather than commanding behaviours. It is a surprisingly precise description of what a mature algorithmic coordination layer does. The power resides not in the individual data collected. The time a service is used, the speed with which a task is completed, the metadata of daily work: each is innocuous in isolation. It resides, rather, in what the system (which is, moreover, never perfectly free of bias) is able to deduce by combining them, and in the fact that those inferences become operative, propagating through the workflows that the coordination layer, by construction, connects.

For organisational AI Adoption this means one precise thing: a coordination system does not merely carry out tasks, it institutionalises categories. A workforce analytics system that uses ordinary operational data to infer potential, attrition risk or productivity does not merely improve an HR decision: it contributes to defining what “talent”, “risk” and “potential” mean in that organisation. And the relevant opacity is not only that of the model, the classic black box, but that of the chain: the person may not know which attribute has been inferred about them, from which data, in which decisions it is used, and may have no means of verifying or contesting it. Even a technically explainable algorithm can produce opaque governance. Nor is the informed consent of the privacy tradition sufficient: one may have consented to providing a piece of data for one purpose without having consented, because it is not realistically possible, to the open space of future inferences that can be constructed by combining it with unknown sources.

In sum, the reduction of coordination costs is not a neutral process. The very infrastructure that reduces the costs of translation between systems increases the capacity to classify and govern people. If automation transfers execution and cognitive augmentation redistributes cognitive capacity, coordination redistributes the power to define relationships, categories and access.

This is exactly the terrain on which the European Regulation on Artificial Intelligence intervenes. The AI Act already oversees the coordinative dimension of people at several points. Article 5, among the prohibited practices already applicable since February 2025, includes emotion recognition in the workplace (Art. 5(1)(f)) and social scoring that produces detrimental treatment in contexts unrelated to those in which the data were generated (Art. 5(1)©): two cases that strike precisely at the passage from inference to differential treatment. Annex III, point 4, classifies as high-risk the AI systems used in employment and worker management, such as recruitment and selection, decisions on promotions and terminations, allocation of tasks based on individual behaviour or traits, monitoring and evaluation of performance and behaviour, that is, to a good approximation, a significant part of the scope of application of the internal coordination layer. For these systems the Regulation prescribes obligations that respond almost point by point to the opacities listed above: transparency to deployers (Art. 13), effective human oversight (Art. 14), the obligation on employers who deploy high-risk AI systems to inform workers and their representatives before putting the system into service (Art. 26(7)), and the person’s right to obtain an explanation of individual decisions taken with the contribution of such systems (Art. 86).

The coordination layer that I have considered in this article in its organisational dimension therefore presents important possibilities and responsibilities above all when it also directly involves people, and not only information, data, automated processes and systems. Those who design AI adoption and governance in the coordination dimension should ask themselves, from now on, questions such as: which attributes does the system infer that were not present in the original data, into which workflows do those inferences propagate, who decides that an inference is reliable enough to become operative, does the person know that this attribute has been assigned to them, which groups bear the errors and which benefit from the efficiency?

Conclusion

The lesson of Reshuffle, brought back to the scale of AI Adoption, after the critical treatment I have proposed, can be condensed as follows: coordination is a dimension that was missing from the grammar of adoption; without it, automation and cognitive augmentation produce local gains that risk dissipating at the boundaries of the task, and may even break down existing coordination. But coordination does not replace the other two dimensions; on the contrary, it can recompose them. The relevant matrix has two axes: the transformation of execution and the transformation of interdependencies. And organisational value is generated at their intersection.

One final caveat. Coordination is not simply efficiency under another name: the layer that coordinates is also the layer that classifies, and to adopt it means deciding, together with the architecture of the flows, the categories through which the organisation will know the people who inhabit it. This is the material organisation of the organisation itself. Those who design adoption along this dimension also choose a mode of governing. And it is for this reason that, in the coordination dimension of AI, strategy and ethics should not meet downstream but should proceed from the very first steps as a single chapter to be worked on.

References

• Sangeet Paul Choudary, Reshuffle. Who Wins When AI Restacks the Knowledge Economy, 2025.
• Sangeet Paul Choudary, Reshuffle and biopolitical power, Substack, 24 September 2025.
• Boston Consulting Group, AI at Work. Strategy Matters More Than Tools, fourth edition, June 2026.
• Gavin Parker, Caryn Katsikogianis, Alan Wong, Simon Murphy, Nicole Sibilio, Agentic AI Turns Every Team into Its Own Transformation Engine, Boston Consulting Group, 19 June 2026.
• Regulation (EU) 2024/1689 (AI Act): Art. 5(1)© and (f); Annex III, point 4; Arts. 13, 14, 26(7), 86; as amended by the “Digital Omnibus AI” simplification package (2026) as regards the application timelines of the obligations for high-risk AI systems.

Some books describe what is already happening; others help us understand it. Beyond Vibe Coding, by Addy Osmani, engineering leader at Google Chrome, belongs to the second category. Built as an expansion of a series of the author’s blog posts, it is anything but purely conceptual: it includes substantial operational sections that are highly useful today, even if they risk becoming dated as tools and practices evolve. What lasts is present, and it warrants careful reading: a conceptual map for navigating AI-assisted development, and an honest treatment of the challenges this transition brings with it. A senior developer and a junior one, a product manager and a vibe coder, will each find a useful reference here for different reasons: to deepen their thinking, absorb best practices, reorganize processes, or draw inspiration.

The Spectrum and the 70%

In my view, the book’s most durable contribution is conceptual. Osmani introduces a spectrum running from vibe coding to AI-assisted engineering, and this mapping offers an effective framework for thinking strategically about one’s positioning as a developer. Vibe coding, a term coined by Andrej Karpathy, describes an exploratory approach in which the developer delegates nearly all implementation to the AI, focusing on intent rather than code. Osmani describes it as “improvisational jazz, minimal structure, lots of room for creative riffs.” AI-assisted engineering, at the opposite end, is methodical and iterative, with the developer firmly in the driver’s seat: “Treat the AI as your intern, not your replacement.”

Osmani avoids simplistically ranking the two poles. Vibe coding is appropriate for validating an idea or building a prototype; AI-assisted engineering becomes necessary when you are building for the long term. Learning to move along the spectrum, and deciding where to position oneself, becomes a matter of analysis and judgment. Weighing context, objectives, risks, and potential, the developer must make the most effective choice. It is a choice that demands judgment, not an abstract rule. And this contribution alone carries significant value: it gives teams a shared vocabulary for aligning around the critical use of these tools.

The analytical heart of the book is the so-called 70% problem: AI is effective at covering the accidental complexity of software (boilerplate, scaffolding, recurring patterns), handling roughly 70% of a typical project. The remaining 30% is where the machine reveals its structural limitations: architecture, edge cases, security, deep domain understanding. This “essential complexity,” to borrow Fred Brooks’s terminology, remains on the human developer’s desk. It represents a redistribution of responsibilities, one that has occurred many times in the history of computing, and it demands a rethinking of what it means, today, to be a good software engineer.

Data and the Reality of Change

On this point, the book offers operational proposals worth discussing within a team, while also opening a broader conversation that it does not exhaust. It is a conversation to be held with the wider community, and one to which it is worth contributing by developing the empirical evidence now available.

The DORA State of AI-Assisted Software Development 2025, conducted across nearly 5,000 industry professionals, finds that 90% of respondents use AI at work, a 14.1% increase over the previous year. 80% perceive an increase in their individual productivity. These figures point to a transition that has already become the operational default. Yet the report’s most telling finding is not adoption rates, but AI’s nature as an amplifier: “AI magnifies the strengths of high-performing organizations and the dysfunctions of struggling ones” (p. 3). Adoption alone, without deliberate design and governance, does not guarantee results. The decisive variable is the organizational system into which AI is introduced: where engineering practices are solid, AI strengthens them; where they are fragile, it destabilizes them.

The Opsera AI Coding Impact 2026 Benchmark Report, built on behavioral data from over 250,000 developers in production, adds a dimension that sentiment surveys tend to obscure. AI reduces Time-to-PR by 48–58% (p. 6), but pull requests generated by AI wait on average 4.6 times longer to enter review than manually written code. The acceleration in the inner loop produces a bottleneck in the outer loop, neutralizing a significant portion of the gain. The report’s conclusion is unequivocal: “How AI is governed matters more than whether it is present” (p. 13). And further: “AI-assisted development has crossed the point of experimentation and entered the realm of enterprise operating risk” (p. 18).

Salvatore Sanfilippo, antirez, creator of Redis, has addressed the question with the same candor that characterizes his technical work. In his essay Don’t Fall Into the Anti-AI Hype, he writes: “for most projects, writing the code yourself is no longer sensible, if not to have fun.” This is an acknowledgment of fact, not an ideological posture, and it carries particular weight given the quality of outputs the most recent models have achieved. His advice is direct: “Whatever you believe about what the Right Thing should be, you can’t control it by refusing what is happening right now. Skipping AI is not going to help you or your career.”

Linus Torvalds has also put his assessment of AI-generated code on the record, using Google’s Antigravity on one of his own projects and commenting in the commit message with characteristic economy: “Is this much better than I could do by hand? Sure is.”

Responsibility in the Vibe

Adoption, then, is a given. But mature adoption requires an ethical and professional infrastructure that Osmani addresses with seriousness, proposing orientations and guidance that will need to be developed further, both in international debate and in the day-to-day operations of teams.

The principle that runs through the book like a backbone is the developer’s responsibility for the code they commit, regardless of who, or what, wrote it. “Never commit code you don’t fully understand” is presented as one of the golden rules of AI-assisted engineering. A developer who accepts an AI output without understanding it is not merely risking a production bug: they are abdicating their role as critical evaluator, which is precisely the irreducible competence required in the remaining 30%. Redefining professional responsibility in this context is the point from which an updated code of practice should be built, alongside the best practices, tools, and processes that can support it.

This position is also empirically necessary. Opsera’s data shows that AI-generated code contains 15–18% more security vulnerabilities than manually written code, with flaws that are often subtle and logic-based, difficult to catch through automated testing. A large-scale analysis of GitHub Copilot cited by Osmani indicates that between 25% and 33% of generated code contained potential security vulnerabilities. The speed of generation expands the attack surface precisely at the moment when attention tends to relax.

Osmani engages honestly with questions of transparency in AI use, bias management, and the need for explicit governance structures. His proposed “responsible AI checklist” is useful as an operational aide-mémoire and can serve as a starting point for developing procedures suited to a team’s specific needs, provided it does not reduce to mechanical routine what is, in reality, a culture shift that requires time, training, and leadership. DORA puts it plainly: “The value of AI is not going to be unlocked by the technology itself, but by reimagining the system of work it inhabits.” This should not be read as an argument that minimizes AI’s impact, but as an argument against the naivety with which it is so often adopted as a solution to pre-existing organizational problems that no tool can resolve on its own.

Beyond Vibe Coding leaves the reader with a reasoned map and a clear-eyed view of a transition already underway. Osmani makes no claim to resolve what the field is still learning to articulate. By sharing his professional experience, he sustains those who want to move from experimentation to more structured practice, and those who, after initial attempts to integrate AI into their workflow, want to engage with the thinking of someone who has already taken these practices into industrial-scale production. What remains open, and worth bringing into the daily conversations of teams, organizations, and professional communities, is the construction of an AI-assisted engineering culture equal to what is at stake. Not merely a matter of tools or checklists, but of collective maturity: the capacity to integrate speed and responsibility, tool autonomy and professional judgment, without allowing either to erode the other. It is slow, deliberate work. It clearly takes more than one book to set it in motion, but books like this one help us understand where to begin.

In previous articles, I began exploring artificial intelligence not only as an automation technology but also as a cognitive infrastructure capable of extending the human mind and reconfiguring the relationship among mind, environment, and action.

In the first article, From Extended Mind to Product Design: Rethinking AI as Cognitive Extension, I proposed reading AI, following Andy Clark’s extended mind theoretical hypothesis, as part of the cognitive circuit connecting brain, body, and world. Through extension mapping and memory policies, I hypothesized that designing intelligent systems now means designing portions of our own mind: identifying where and how AI intervenes in processes of perception, understanding, and decision-making.

In the second article, Predictive Processing for AI Products, I shifted attention to the dynamics of this integration. Predictive processing, the capacity to anticipate, correct, and learn from errors, showed how AI does not act neutrally: it can amplify our cognitive patterns but also constrain them, depending on how prediction and feedback cycles are designed. The quality of interaction and effectiveness depend not only on model power but on how the model fits into users’ cognitive rhythms through design choices, such as balancing uncertainty and control, action and reflection.

With this third article, I want to shift the focus even further upstream. Having explored the what (the extended mind) and the how (predictive dynamics), I now consider the why and with what responsibilities. Once we’ve addressed how AI integrates into our mental processes, the question becomes how it pre-configures them and how what can or cannot enter our field of attention is defined a priori. Following this line of inquiry takes us into a region where experience is filtered, ordered, and made available to thought by artificial intelligence systems, a region we might call System 0, as proposed in a paper by Chiriatti, Ganapini, Panai, and Riva (2024).

System 0 represents the algorithmic preprocessing layer that today mediates much of our interaction with the digital world: search engines, recommendation systems, generative models, and contextual assistants. Before we perceive, evaluate, or decide, these systems have already selected, classified, and made certain contextual elements salient. In this sense, it’s a cognitive extension that operates as epistemic infrastructure, acting upstream of experience and contributing to structuring its content and form. Reasoning about System 0 therefore means addressing the question of cognitive responsibility beyond simply correcting AI’s effects, focusing instead on understanding its premises.

The paper System 0: Transforming Artificial Intelligence into a Cognitive Extension, one of the most interesting contributions published in 2025, focuses precisely on the cognitive contexts where AI grafts itself as a preprocessor, providing valuable guidance through solid frameworks for realizing an ethically grounded cognitive ecology oriented toward understanding AI as enabling enhancement rather than replacing human capabilities.

In what follows, I’ll explore some aspects of System 0, relating it to the hypotheses I discussed in the two previous articles, before considering it in its ethical dimension.

Dual Process Model and the Emergence of System 0

The expression System 0 was chosen by the authors to position itself directly in relation to the Dual Process Model, the theory developed by Daniel Kahneman who, in Thinking, Fast and Slow, describes human thought as the asymmetric interaction of two functional systems: System 1 and System 2 (Kahneman, 2011). The first is the autopilot of daily intuitions: fast, associative, and immediate, it builds coherent representations from partial cues, privileges cognitive fluency, and operates according to the What You See Is All There Is (WYSIATI) principle, with well-known outcomes in terms of biases, heuristics, and illusions of truth. The second is the slow gear of reasoning: deliberative, sequential, and capable of control, but costly in terms of attentional resources; it inhibits, verifies, and plans, activating under surprise or difficulty while remaining vulnerable to ego depletion (fatigue from prolonged efforts of will and self-control). In this framework, rationality isn’t a unitary agent but the unstable outcome of continuous negotiation between associative automatisms (System 1) and cognitive vigilance (System 2).

In the contemporary digital context, before System 1 generates intuitions or System 2 initiates analysis, as the study’s authors note, a distributed algorithmic infrastructure has already reorganized the field of experience by:

• filtering what reaches us (recommendation systems)
• classifying importance and relevance (search engines)
• structuring presentation (generative interfaces, ranking, formats)
• orienting interpretive schemas (conversational assistants and agents)

To name this new level where informational preprocessing occurs, modulating what becomes salient upstream of our cognitive processes, Chiriatti, Ganapini, Panai, and Riva propose the term System 0: an algorithmic layer that acts before System 1 and System 2, selecting, reordering, and making inputs and contexts relevant (Chiriatti et al., 2024; 2025).

System 0 is characterized by four operational properties:

• algorithmic (statistical patterns without intrinsic semantic anchoring)
• distributed (ecosystems of interconnected platforms)
• autonomous (learns and adapts with minimal supervision)
• pervasive (persistent and influential across multiple domains)

Unlike biological systems, this layer doesn’t attribute meaning: it enables, modulates, and reorders, leaving meaning-making to human processes. At the same time, it shows a high degree of functional integration with our cognitive processes.

Using criteria proposed by Heersmink (information flow, reliability, duration, trust, procedural and informational transparency, individualization, transformation) the authors evaluate how System 0 can be considered part of our cognitive scaffolding (Heersmink, 2015; Clark, 2008).

If with System 0 “algorithmic relevance” tends to precede, and sometimes replace, our criteria of relevance and verification, a shift in knowledge practices occurs in line with what I discussed in previous articles: from the Extended Mind perspective, System 0 redesigns segments of the cognitive circuit by distributing competencies and memory (Clark, 2008; 2023); from the predictive processing perspective, it affects precision weights and the predictive budget before System 1 and System 2 engage, predisposing attentional patterns, hypotheses, and action trajectories (Clark, 2016; Friston et al., 2006); from the ethical-epistemic perspective, it invites interrogation of preprocessing quality and its boundaries: what it makes visible, how contestable it is, how much remains reversible.

Paradoxes and Tensions of System 0: Between Extension and Constraints

Applying the System 0 lens to AI product design means immediately encountering intrinsic tensions and paradoxes, and requires building multifunctional teams where ethics, cognitive science, engineering, and design converge to realize an effective, reliable, high-quality HAX (Human–AI Experience).

A first paradox concerns an AI system’s capacity to reduce uncertainty, complexity, and decisional latency, producing cognitive comfort that enables action but, over time, may contract users’ margin for learning and adaptation: when activities requiring exploration, working memory, or judgment are stably delegated, some competencies tend to weaken. In predictive processing terms, this corresponds to a recalibration of precision weights in favor of stable external signals, reinforcing conservative expectations (Clark, 2016; Friston et al., 2006). The design objective then becomes keeping the “growth” side open without renouncing the “comfort” benefit: scaffolds that enable, avoiding prostheses that rigidify.

A second tension concerns the distinction between epistemic and executive modes. Many systems don’t thematize it: they offer the same support profile regardless of intention and context, risking functional drift that confuses expanding understanding with operational efficiency. From a practical standpoint, this drift is recognizable when decisional latency concentrates almost entirely on understanding, signaling insufficient rationales, or when most user modifications are “corrective” rather than “constructive,” indicating misalignment between objective and support.

Making explicit or carefully inferring the engagement modality and modulating accordingly the format, breadth, and tone of AI intervention is good practice for making interaction more effective and attentive.

Finally, if System 0 pre-structures what becomes visible and relevant, even small asymmetries can produce large cognitive drifts: compliant convergence (sycophancy), impoverishment of alternatives (echo chamber), miscalibrated trust relative to real uncertainty level (over-trust). Here the problem isn’t only output accuracy but preprocessing quality: what has been excluded and why, with what trade-offs and what reversibility (Chiriatti et al., 2024; 2025). The challenge is open both for large foundation model providers and for organizations developing AI products: the latter can, for example, make structural a synthetic adversarial process, few but genuinely diverse alternatives, accompanied by “why not X”, guarantee criteria visibility and low-friction contestability, and verify map fidelity so that steps reserved for human judgment remain such. In parallel, experimental measures like segmenting latency into understanding/validation, qualified override rate, and reversibility friction help monitor loop health and maintain user agency as an operational reference.

As we’ve seen, System 0’s tensions and paradoxes are no small challenges and impact crucial areas like scaffolding organization and users’ cognitive hygiene. Without yielding to the moral panic that often accompanies new technologies, these challenges represent genuine fields of innovation and research, where each solution is an opportunity to consolidate the trust between organization and user that forms the basis for obtaining long-term value from adopting responsible AI systems.

Designing Within System 0: Ethics and Practice

But what does it concretely mean to build a responsible AI product in light of System 0? In the debate on responsible artificial intelligence, especially in the European context, a paradigm emerges that defines trustworthiness through three inseparable dimensions:

1. Legal compliance (conformity with regulations on safety, data protection (GDPR), and sectoral regulations)
2. Technical robustness (reliability, security, and resilience under variable conditions)
3. Ethical soundness (coherence with ethical principles such as equity, transparency, autonomy, and well-being)

The System 0 hypothesis highlights the necessity of this triple attention: if one dimension is missing, the entire cognitive circuit extended to AI can prove fragile or harmful. A technically robust but opaque system undermines autonomy; one ethically inspired but technically fragile generates harm despite good intentions; one compliant but designed without considering cognitive well-being respects the letter but betrays the spirit of responsibility.

A System 0 ethics that engages with compliance and technical robustness requirements is therefore not ornamental or an external constraint but part of the product process and a resource for ensuring identification of risks to mitigate and opportunities to seize. In this sense, the seven frameworks proposed in the paper translate ethical values into very concrete product choices, operationalizable through tools the team decides are most pertinent to the project, in our case, to continue the experimentation discussed in this article series, these can be extension mapping, memory policies, and cognitive metrics:

• Enhanced Cognitive Scaffolding responds to the value of autonomy: AI must enhance user capabilities, not replace them.
• Symbiotic Division of Labor ensures AI doesn’t monopolize decisions requiring contextual judgment or moral responsibility.
• Dialectical Cognitive Enhancement counters sycophancy by designing contextual AI agents that also function as intellectual sparring partners.
• Agentic Transparency and Control operationalizes the right to understanding and control. Beyond technical transparency, it makes visible the pathways of information and actions in a system, revealing when AI enters and when humans do.
• Expertise Democratization realizes equity in access to knowledge through cross-domain translation layers.
• Social-Emotional Augmentation recognizes the affective dimension of interaction. Designing for positive emotion cultivation and psychological safety is ethical responsibility, but requires avoiding dark patterns that exploit vulnerabilities.
• Duration-Optimized Integration manages the relationship with an AI system over time. Effectiveness isn’t constant but depending on context and domain can assume varying intensity and benefits.

It may be interesting to consider how the seven frameworks we briefly analyzed invoke specific articles of the EU AI Act, showing how ethical values and legal requirements converge in the European regulatory framework:

This mapping with the AI Act reveals how a system designed for progressive scaffolding, transparent division of labor, dialectical enhancement, agentic control, expertise democratization, authentic emotional support, and temporal monitoring respects not just the letter but especially the spirit of regulation: protecting autonomy, promoting equity, guaranteeing well-being. Each element we considered communicates a vision of what it means to enhance human cognition and establishes the premises for distributing AI products capable of lasting over time, strengthened by the trust established between user and organization.

System 0 is already part of our cognitive ecosystem, and increasingly will be. Orienting ourselves now to design AI products that, instead of constraining through comfort and compliance, enable and stimulate capabilities through structured support and productive tension; that, instead of deciding in opacity, expand possibilities in transparency; that, instead of generating dependency, cultivate autonomy — this can make the difference. As Clark observes in the article from which this exploration of Extended Mind and contemporary artificial intelligence systems began, younger generations already show greater awareness of issues around privacy, transparency, and cognitive ecology (Clark, 2025). In the future, competencies and practices for critically evaluating what we incorporate into our digitally extended minds will be far more widespread and refined than current ones, and designing AI products now that align with extended cognitive hygiene offers both an advantage for the organization and an opportunity to contribute to establishing a responsible System 0 characterized by genuine cognitive growth and expansion.

References

• M. Chiriatti, M. Ganapini, E. Panai, B. K. Wiederhold & G. Riva, System 0: Transforming Artificial Intelligence into a Cognitive Extension. (2025)
• A. Clark, Extending Minds with Generative AI. Nature Communications. (2025)
• M. Chiriatti, M. Ganapini, E. Panai, M. Ubiali & G. Riva, The case for human–AI interaction as system 0 thinking. Nature Human Behaviour. (2024)
• European Union, Regulation (EU) 2024/1689 of the European Parliament and of the Council on Artificial Intelligence (AI Act). (2024)
• OECD, Recommendation of the Council on Artificial Intelligence. (2024)
• High-Level Expert Group on AI, Ethics Guidelines for Trustworthy AI. European Commission. (2019)
• R. Heersmink, Dimensions of integration in embedded and extended cognitive systems. Phenomenology and the Cognitive Sciences. (2014)
• D. Kahneman, Thinking, Fast and Slow. (2011)
• A. Clark, Supersizing the Mind: Embodiment, Action, and Cognitive Extension. (2008)
• K. Friston, J. Kilner & L. Harrison, A free energy principle for the brain. Journal of Physiology-Paris. (2006)
• A. Clark, The Experience Machine: How Our Minds Predict and Shape Reality. (2023)
• A. Clark, Surfing Uncertainty: Prediction, Action, and the Embodied Mind. (2016)

In my previous article, I introduced extended mind theory applied to an experimental framework for AI product design. In that discussion, I examined prediction-action cognitive cycles in light of AI integration, and touched on how the mind operates as a continuous prediction engine that constantly restructures itself based on experience. Andy Clark calls this characteristic predictive processing and first explored it in his Surfing Uncertainty: Prediction, Action, and the Embodied Mind (2016), then revisited the topic in a more accessible form in his recent The Experience Machine: How Our Minds Predict and Shape Reality (2023).

The idea that the mind is not simply a tabula rasa passively receiving sensory data, but functions as an active apparatus for organizing experience, is not new. In the modern era, we can trace it back to Immanuel Kant, who argued for the mind’s active role in structuring experience. From there it traversed various fields of knowledge: from psychology to physiology, from cybernetics to contemporary neuroscience. Between the 1980s and 1990s came the formalization of the Bayesian brain (Rao & Ballard, 1999): the brain doesn’t simply record what arrives from the senses, but continuously generates predictions about what it should perceive and updates its hypotheses by comparing them with reality. In the early 2000s, Friston took this further with the free energy principle (Friston et al., 2006): not only do we perceive predictively, but our actions and adaptation strategies also serve to reduce the uncertainty between what we expect and what actually happens. For readers interested in delving deeper into the free energy principle, a recent interview with Friston on the Machine Learning Street Talk podcast offers an excellent introduction.

Today, predictive processing, closely related to and often modeled through active inference, which highlights how agents not only predict but also act to minimize uncertainty, offers a rigorous though still debated theoretical foundation for understanding intelligent systems. From this perspective, an agent, whether biological or artificial, is any system that maintains an internal model of relationships between its environment and the inputs it receives. In living organisms these inputs are sensations in the proper sense; in artificial agents they are signals or data from sensors or the operational context. In both cases, through perceiving and acting, agents constantly seek to reduce uncertainty by minimizing free energy.

As Pezzulo, Parr, and Friston (2025) observe, current AI systems learn predominantly passively from datasets, whereas the future of embodied AI will require forms of interactive learning that more closely mirror how biological agents learn in relation to the physical and social environments in which they operate. This transition from passive to interactive learning is crucial for understanding the emergence of new cognitive coalitions and represents one of the most relevant research axes in both data engineering and human-computer interaction.

In my previous article, I also highlighted how the value of new AI features and products lies not only in the automation they introduce, but also in the quality of cognitive processes they extend. In this sense, a coding assistant that anticipates the next lines of code, a writing tool that suggests email tone, or an intelligent design system that proposes contextual layouts are potential participants in what Clark (2023) defines as the predictive knowledge budget: systems of predictive knowledge management that extend beyond the boundaries of the individual user’s “head.” As with the extended mind, the principles of active inference can help us, with theoretically grounded language, describe human-AI interaction loops more carefully, interpret phenomena, propose analytical categories, and identify problematic patterns. In short, they can be invaluable when thinking about requirements and designing new AI products. Below, I offer a brief introductory exploration of these principles drawing on Clark’s work and related research.

Anatomy of Cognitive Loops

Let’s start with a field observation: imagine a developer working with GitHub Copilot and the rhythm of their interaction with the assistant. Beyond the simple command-response logic of a conversational window, the flow works like this: the developer starts typing a function, the model anticipates the intention and proposes a completion; the user can accept, modify, or ignore it. Looking through the lens of extended active inference, we notice that, over time, the programmer tends to adapt their own style (naming conventions, argument order, recurring patterns) in ways that make predictions more relevant. In turn, the model leverages that context to generate better-aligned suggestions. The user predicts the system and the system predicts the user: a mutual calibration that emerges from the perception-action-correction cycle, even without persistent memory of rejections or user preferences and limited to local context. We can find this pattern in other contexts as well, for instance, an experience I suspect most readers share is the iterative process of restructuring prompts to optimize output procedurally. Or a writer using an editing assistant who modifies sentence constructions to elicit more effective predictions.

Let’s return to our developer and their AI assistant. We can say the assistant maintains an internal model of the user: not a psychological model in the strong sense, but an implicit representation that emerges both from the code context on which Copilot was trained and from what it’s currently analyzing. If the programmer adopts consistent variable naming schemes, recurring function structures, or a precise documentation style, the system tends to predict that pattern will repeat and thus proposes coherent completions. It’s in this sense that we can speak, by analogy, of an internal model of the user’s preferences and behaviors. The quality of these predictions depends on how much “weight” the system assigns to its own expectations versus new inputs, what the framework calls precision weighting. If the precision attributed to predictions is too high, the assistant risks ignoring subtle signals of change, such as when the programmer wants to deviate from their usual pattern; if it’s too low, suggestions become generic or inconsistent. When a prediction doesn’t match the user’s actual action, prediction error comes into play. In active inference models, error serves to correct and update hypotheses; in Copilot’s case, while it can’t directly modify the model’s parameters, the error still becomes informative: for the AI because it enables better realignment with the immediate context, for the user because rejecting or modifying a suggestion becomes a moment of reflection on the code they were about to write.

What makes these loops powerful is precisely this recursive nature, where the user adapts to the system and the system adapts to the user in a continuous process that can generate, beyond a “good feeling”, new forms of cognitive collaboration. What’s decisive is the quality of the loops: when well-designed, both partners, human and artificial, enhance capabilities; when dysfunctional, they can lead to dependencies, rigidities, or weakening of critical skills.

Emerging Patterns in Interaction

To design effective loops, it helps to observe the different interaction types and recurring patterns that have emerged in recent years as humans and artificial intelligences coordinate cognitively. While there’s no definitive classification yet, some distinctions appear highly significant and could prove useful during AI product discussion and design.

A first distinction concerns the user’s cognitive orientation. In exploratory mode, the person seeks to understand the possibilities within a problem space, generate alternatives, discover unexpected connections. In confirmatory mode, they already have a clear direction and mainly desire effective execution. The AI product should modulate its behavior based on these two states.

In the context of coding assistants, the difference is evident. During exploration, the developer needs suggestions that expand the space of possible solutions, what in active inference language corresponds to epistemic behavior, oriented toward reducing uncertainty. When implementing an already-decided strategy, they prefer completions that accelerate execution without introducing ambiguity: a pragmatic behavior, centered on the goal. Many current systems, however, don’t distinguish between these two cognitive states. They offer the same type of support regardless of the user’s intent, thus missing the opportunity for finer calibration and more effective collaboration.

Let’s delve deeper into this problem: epistemic actions aim to improve understanding of a situation and seek to acquire information; pragmatic actions aim instead to produce a practical, immediate result in the world. When a user queries an assistant, are they performing an epistemic or pragmatic action? The answer should change the type of support they receive. If I ask “How can I optimize this SQL query?” to learn something about databases, I’ll need a detailed explanation of underlying principles more than an immediate solution. If I pose the same question because I need to solve a production performance problem, efficiency matters more and I need a crisp, concise, precise answer to validate.

Alongside the distinction between exploratory and confirmatory orientation lie the operational modes through which AI interacts: it can propose content ex novo or help choose among already available alternatives. In the first case, the user faces generative output (a paragraph, a block of code, a mockup) that didn’t exist before. In the second case, the system acts as selection, as a filter: it sorts, prioritizes, highlights what’s already present but dispersed, reducing noise and simplifying decisions. For the user, the two regimes activate different cognitive processes. Generation requires evaluating creativity, relevance, and quality of novel proposals: a curatorial role that calls upon editorial judgment and critical capacity. Selection, conversely, solicits abilities of comparison and pattern recognition among known options, asking users to weigh trade-offs rather than validate novelty. It’s not the same thing working with a writing tool that composes entire paragraphs versus one that only suggests the next word: in the first case evaluation happens at a high level; in the second you enter a micro-temporal dynamic of rapid local choices that accompany the rhythm of writing.

Then there’s the question of support duration. Some assistants are designed as temporary scaffolding that help learn a skill and progressively withdraw, favoring growing user autonomy. Others become permanent prosthetics, stably integrated into the cognitive environment: autocomplete in IDEs is the classic example, a function that once entered into a programmer’s cognitive scaffolding would hardly be considered optional. The problem is that this distinction is rarely made explicit in design. Not infrequently, tools intended to be transitory end up stabilizing as permanent by default, triggering unintentional “dependency patterns”: dependency schemes that don’t strengthen user capabilities but risk limiting them. This is what happens, for example, with slide or graphic layout suggestion systems: born to speed up a project’s initial phase, they end up becoming the default model to rely on, progressively reducing the drive to explore more original alternatives.

These distinctions outline practical criteria that can guide the design of systems more sensitive to the user’s cognitive context. An interface capable of recognizing whether the user is in exploratory or confirmatory phase, whether they’re acting to reduce uncertainty or achieve a goal, whether they need generation or selection, can calibrate its behavior in a finer, more collaborative way. The open challenge is understanding how to detect these intentions in real time and translate them into differentiated responses without breaking the loop’s fluidity. Alternatively, during product or feature design, and if context allows, conscious decisions can be made that make cognitive loops more virtuous, placing the most pertinent functionality at the right point in sequences.

When Loops Break: Anti-Patterns

Many of us have experienced at least once a phenomenon like this: an AI system that initially provided relevant, well-calibrated suggestions gradually begins to behave less reliably. Autocorrect that “learns” from our recurring typing errors and starts suggesting them as valid corrections, or a coding assistant that always proposes the same solutions regardless of specific context. Active inference theory suggests that what’s manifesting here is precision collapse, the system loses the ability to distinguish between high- and low-confidence predictions. Everything starts being treated with the same weight, generating suggestions that oscillate between the obvious and the inappropriate without sensible modulation. This typically happens when feedback loops degrade. The system interprets continued use as confirmation of its predictions, but in reality the user might simply be in a routine phase or one of low critical attention. The absence of explicit negative feedback gets erroneously translated into approval, leading the system to rigidify its predictions around patterns that might be contingent or temporary. Indicators of this anti-pattern include suggestions that become progressively less contextual, repetition of standard solutions even in situations requiring different approaches, and loss of sensitivity to variations in user input.

These are the best-known anti-patterns, within everyone’s experience, but other more subtle mechanisms merit consideration. First among these concerns the progressive erosion of cognitive skills when AI assistance becomes too invasive. This isn’t dependency in a pathological sense, but what we might call “unintentional cognitive offloading.” The classic example is GPS navigation. Many people report having lost their sense of spatial orientation after years of relying on navigation systems. The brain, according to active inference, optimizes the use of cognitive resources, if a capability isn’t exercised because it’s externally delegated, the neural structures that support it adapt differently. Applied to cognitive work, this pattern takes more subtle forms. A highly effective writing assistant might reduce over time the ability to construct discourse autonomously. A coding assistant might erode the programmer’s familiarity with APIs and libraries.

The social history of technology and human-machine relationships reminds us, however, that this isn’t necessarily bad: cognitive evolution is always made of trade-offs and negotiations in the restructuring of cognitive scaffolding. It becomes a critical problem, however, when users lose awareness and control of these transformations, or when dependency on the system prevents adaptation to new contexts.

Excessive AI invasiveness can also manifest in another anti-pattern: excessive feedback noise. Some writing assistants flag every minimal stylistic variation; certain IDEs highlight every micro-optimization of code. The intention is support, but the result is cognitive overload. From the active inference perspective, the brain continuously filters relevant signals from background noise. When an AI system generates too frequent or poorly calibrated feedback, it disturbs this natural process of attentional filtering, forcing users to dedicate cognitive resources to ignoring irrelevant suggestions instead of focusing on the main task.

But among the signals that tax our attentional budget, perhaps most relevant is how the AI system communicates its own confidence. Many tools present outputs with apparent certainty that doesn’t reflect the real uncertainty of predictions; the “confidently delivered hallucinations” of large language models are the extreme version of this problem. In active inference terms, this introduces a distortion in user calibration: normally we use social and contextual cues to estimate how much to trust information, but if the system always “speaks” with the same confident voice, those mechanisms risk being bypassed. The result is miscalibrated trust: too high in domains familiar to the model; paradoxically too low when the system expresses genuine uncertainty. This is a particularly risky pattern where accuracy matters and verification isn’t immediate.

At the opposite extreme from hyper-inflated trust lies one last pattern worth mentioning in this quick exploration: the erosion and fragmentation of cognitive continuity that occurs when AI systems facilitate multitasking without preserving cognitive continuity. For example, an assistant that helps manage multiple projects simultaneously might inadvertently encourage context switches that disturb deep thought processes. Active inference theory emphasizes the importance of continuity in predictive models. Each interruption requires reconstruction of cognitive context, and thus significant expenditure of attentional resources. When AI facilitates too many context switches, it can fragment the user’s cognitive experience, causing feelings of mental dispersion, difficulty maintaining prolonged focus, and paradoxically a productivity that seems high in the short term but generates growing cognitive fatigue.

The anti-patterns discussed are not inevitable, but depend on the form taken by prediction and correction cycles and on the care with which precision, uncertainty, and attentional cost are governed. Predictive processing offers guidance for this care, enabling the design of systems that modulate uncertainty instead of masking it, that make their own boundaries explicit, that preserve contextual continuity and provide for intentional withdrawal from the scaffold when user competence reorganizes.

From this perspective, an AI product’s value isn’t exhausted by accuracy or time saved alone, but is also measured by the quality of transformations introduced in individual and collective cognitive patterns: the ability to sustain attention when necessary, improve judgment under uncertainty, increase the reversibility of choices. For organizations, adopting this vision in the development cycle already makes a difference: building a relationship of trust founded on a cognitive ecology that truly enriches human capabilities.

References

- G. Pezzulo, T. Parr & K. Friston, Shared worlds, shared minds: Strategies to develop physically and socially embedded AI. (2025)
- A. Clark, Extending the Predictive Mind. (2022)
- A. Constant, A. Clark, M. Kirchhoff, & K. Friston, Extended active inference: Constructing predictive cognition beyond skulls. (2019)
- K. Friston, The free-energy principle: A rough guide to the brain? (2009)
- K. Friston, J. Kilner, L. Harrison, A free energy principle for the brain (2006)
- R. P. N. Rao & D. H. Ballard, Predictive coding in the visual cortex: A functional interpretation of some extra-classical receptive-field effects. (1999)
- A. Clark, The Experience Machine: How Our Minds Predict and Shape Reality. (2023)
- A. Clark, Surfing Uncertainty: Prediction, Action, and the Embodied Mind. (2016)

We’re mistaken when we think AI “replaces” human thinking. And we’re equally mistaken when we treat notebooks, checklists, GPS devices, or LLMs as mere operational support tools: very often we are actually extending our minds. We’re not simply externalizing laziness; under certain conditions we are reconfiguring the cognitive circuit that connects brain, body, and world. The mind, in fact, doesn’t coincide with the biological boundaries of the brain: it’s a distributed system that integrates neural, bodily, and environmental resources in dynamic cycles of perception and action. This is the thesis of the Extended Mind, proposed and developed over the years by philosophers of mind and cognitive scientists Andy Clark and David Chalmers. More recently, Clark has discussed the impact of new generative intelligences in a commentary in Nature Communications, titled Extending Minds with Generative AI(2025).

Following Clark’s proposal, the error lies in locating the value of generative AI in simple text generation, when it actually resides in designing the alliance between people, artifacts, and models. Thinking about an AI product means designing the choreography of cognitive circuits: what remains with the brain, what passes to body/tools, what gets delegated to AI and with which rules of trust, memory, and autonomy.

Extended Mind Theory and Artificial Intelligence

In his Supersizing the Mind: Embodiment, Action, and Cognitive Extension (2008), Clark shows how the mind is a distributed system: cognition emerges from dynamic cycles of perception and action that interweave brain, body, and environment. The brain doesn’t do everything alone: it orchestrates an economy that exploits ecological regularities and morphological constraints (morphological computation), integrating tools when they become stable routines (from an Alzheimer’s patient’s notebook to brain–machine interfaces). The boundary between human agent and world is plastic: under the right conditions of functional integration (routine, reliability, accessibility), artifacts become actual parts of the cognitive system.

The primary extensive infrastructure is language. It’s not merely a medium but a technology that labels and structures the world, opens computational opportunities, stabilizes thoughts as reflective objects, and enables the construction of designed cognitive environments (cognitive niches). Through this scaffolding function, it not only enhances the individual but makes possible the cumulative transmission of practices, tools, and institutions, giving shape to a true collective epistemic engineering that is passed down and evolves from generation to generation. From this perspective, conversational tools like LLMs and agents can be extremely powerful elements of cognitive scaffolding.

Indeed, generative AI extends (doesn’t break or interrupt) the extended mind. LLMs and agents graft onto our prediction–action cycles, reducing uncertainty and opening previously invisible spaces. This externality can generate new forms of creativity and lead to innovative solutions, or it can rigidify cognitive monocultures or integrate biases; the difference lies in how we design the human–AI coalition and which rules social and organizational govern it, also in light of specific properties, current limitations and issues (e.g., probabilistic nature, bias, etc.). After all, if an AI product not only automates and speeds up routine operations but influences decisions, priorities, and attention, it is already actively participating in structuring the user’s cognitive scaffolding.

Extension mapping and memory

If we look at the AI product through the lens of the extended mind, artificial intelligence is no longer just a matter of backend process optimization, usability, and satisfaction: it becomes knowing how to curate effective cognitive coalitions in which people, artifacts, and models cooperate to reduce uncertainties, improve decision quality, and increase capabilities. Within the framework of the predictive mind (cf. A. Clark, The Experience Machine: How Our Minds Predict and Shape Reality), according to which the brain constantly anticipates what will happen and reduces uncertainty by comparing its expectations with the external world in a prediction–action–feedback loop, designing an AI product means designing circuits that make these cycles fast, intelligible, and controllable.

In this framework, the first design tool can be an extension map. It’s not an organizational chart of features, but a choreography that clarifies where it’s beneficial for competence to remain “in the user’s head,” where to rely on stable (templates, checklists, sensors) and deterministic tools, and where to activate probabilistic traits mediated by AI (exploration of alternatives, synthesis, plausible predictions). Workflows become the orchestration fabric between these two planes: some steps remain deterministic to guarantee traceability, compliance, and certain timing; others are AI-mediated when uncertainty is high or the variety of options requires search and generalization. A good map makes explicit when and why AI comes into play, which criteria it applies, and where human judgment re-enters: thus the flow is not an opaque chain of service calls, but a readable circuit that reduces uncertainty without sacrificing agency and responsibility.

Memory is grafted onto this framework. In an extended mind system, memory is not a neutral repository, but the way the human–AI alliance preserves continuity and meaning. In the more deterministic segments, it’s advisable to establish stable and auditable memory (glossaries, constraints, already validated decisions); in the probabilistic segments, light, versioned and reversible memory is needed instead, which records hypotheses and rationales without rigidifying, with the possibility of working in guest mode or automatically expiring preferences and contexts. Reasoning traceability is crucial: every AI proposal should carry with it the sources and steps that generated it, so that the user doesn’t just inherit the outcome but can reconstruct, correct, or reset when objectives and conditions change. In this way, memory supports the prediction–action–feedback cycle, strengthens trust and agency, and prevents the cognitive scaffolding from becoming an opaque automatism.

A quick application draft could be a hybrid customer service. An assistance tool, where the extension map reserves to the operator the interpretation of emotional tone and the decision on reparative gestures (the “in head” part), delegates to AI the probabilistic exploration of the knowledge base and the proposal of candidate responses (with readable motivations and sources), and maintains deterministic compliance and escalation workflows. Memory distinguishes a “hot” one for conversation (temporary hypotheses and preferences with automatic expiration) and a “cold” one for auditable glossaries and policies. Result: less uncertainty in the first exchange, shorter resolution times, and explainable decisions, without ever losing operator agency.

This is merely a simple example that I plan to develop further by exploring more operational tools, but it demonstrates how extended mind theory can serve as a powerful experimental framework for creating valuable products. It focuses attention not only on efficiency and usability but also on the user’s cognitive experience while supporting autonomy and growth. Interfaces don’t just capture attention, guide behaviors, and suggest content: they’re environments where we think. If AI extends our capabilities, product design choices also shape what becomes part of us. Creating well-regulated extension maps and memory systems is a concrete way of accepting this responsibility: one case at a time, with rigor and care, ensuring that human–AI alliances improve how we perceive, decide, and act.

Metrics for extension

To make these principles operational, I propose an experimental set of metrics that serve as compasses for observing and optimizing extension. These are not traditional performance KPIs, but indicators that help us understand whether we are effectively designing extension.

Cognitive effectiveness

Decision space reduction (ΔS) measures how much the AI system helps the user move from a state of high indeterminacy to one of focused choice. It’s not about achieving absolute certainties, generative AI is, after all, still probabilistic, but about transforming an open and ambiguous problem into a manageable set of concrete alternatives. If I start with infinite vague possibilities and AI helps me converge on three well-outlined options with explicit trade-offs, it’s fulfilling its role. A low ΔS suggests that AI is simply shifting uncertainty rather than structuring it usefully.

Decision latency tracks time from initial request to confirmed decision, not to AI’s first suggestion. Segmenting this time between “understanding proposals” and “modification/validation” reveals where to intervene: if users spend a lot of time understanding suggestions, the problem is explainability; if they modify them extensively, there might be a misalignment in expectations.

Qualified override rate goes beyond simply counting how many times the user modifies AI suggestions. It distinguishes between “constructive” overrides (that enrich or contextualize the proposal) and “corrective” ones (that repair errors or inadequacies). A high rate of constructive overrides signals that the user maintains creative agency; a corrective rate indicates reliability or context comprehension problems.

Transparency and control

Explanation coverage verifies what percentage of AI outputs are accompanied by verifiable sources and rationales. But providing explanations isn’t enough: effective engagement with them must also be measured (clicks on sources, consultation time). Ignored explanations are transparency theater, not real transparency.

Reversibility friction quantifies how easy it is to undo or modify the effect of an AI intervention. It’s measured in number of steps or time needed to return to the previous state. High friction means the user is trapped in AI choices; low friction preserves the possibility of reconsideration and experimentation.

Dissent coverage tracks how often the system offers alternative interpretations or divergent solutions, avoiding confirmation bias. Generating alternatives isn’t enough: whether they’re actually considered must be measured. A system that presents only confirmations creates cognitive echo chambers.

System integrity

Map fidelity verifies consistency between who should perform a task according to design (human, deterministic tool, AI) and who actually performs it. Frequent deviations signal process drift or trust issues: if AI invades spaces reserved for human judgment or vice versa, the cognitive circuit is reconfiguring in unplanned ways.

Memory balance distinguishes between leverage (how much memory improves decision quality and speed) and staleness (how often suggestions based on obsolete memory degrade the experience). Memory that doesn’t forget can be as harmful as memory that doesn’t remember: the ratio between these two values indicates whether the system is maintaining a healthy balance between continuity and adaptation.

These metrics are not definitive but starting points for new explorations. After all, each context requires specific calibrations and tools, but having a shared vocabulary for discussing cognitive extension and AI can be a step toward designing effective and high-impact products.

References

- A. Clark, Extending Minds with Generative AI (2025)
- A. Clark, Supersizing the Mind: Embodiment, Action, and Cognitive Extension (2008)
- A. Clark, Surfing Uncertainty: Prediction, Action, and the Embodied Mind (2016)
- A. Clark, The Experience Machine: How Our Minds Predict and Shape Reality (2023)