<?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 David Rovinsky on Medium]]></title>
        <description><![CDATA[Stories by David Rovinsky on Medium]]></description>
        <link>https://medium.com/@rovinsky.david?source=rss-8e72b1593664------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*GlvqMdlPywHhmVDI</url>
            <title>Stories by David Rovinsky on Medium</title>
            <link>https://medium.com/@rovinsky.david?source=rss-8e72b1593664------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 16:25:25 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@rovinsky.david/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 Tunneling]]></title>
            <link>https://medium.com/@rovinsky.david/ai-tunneling-why-smart-people-make-bad-decisions-about-ai-2a40499c0ccc?source=rss-8e72b1593664------2</link>
            <guid isPermaLink="false">https://medium.com/p/2a40499c0ccc</guid>
            <category><![CDATA[software-architecture]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[engineering-leadership]]></category>
            <category><![CDATA[decision-making]]></category>
            <category><![CDATA[executive-leadership]]></category>
            <dc:creator><![CDATA[David Rovinsky]]></dc:creator>
            <pubDate>Fri, 17 Jul 2026 20:38:03 GMT</pubDate>
            <atom:updated>2026-07-26T15:26:26.071Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Hidden Cognitive Trap Behind Decisions About AI</h3><p><em>The greatest risk isn’t that AI gives the wrong answer. It’s that people make the wrong decisions about where, when, and how AI should influence human judgment.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F1X08KB0LDkbQGQlw4vBUQ.png" /><figcaption>CEO: “Good news! AI will do 90% of the work…”</figcaption></figure><p>Artificial Intelligence is transforming software engineering at an unprecedented pace. But its greatest impact may not be the code it generates — it may be the decisions people make about how and where to use it.</p><p>At the engineering level, AI Tunneling produces plausible implementations without sufficient architectural judgment.</p><p>At the executive level, it produces plausible strategies without sufficient understanding of technology, organizations, and long-term consequences.</p><p>The difference is scale. An engineer’s mistake may affect a feature or a system. An executive’s mistake can reshape an entire organization. Decisions about adopting AI, reducing experienced teams, shifting investment, or redefining strategy can have consequences measured not in bugs, but in years and millions of dollars.</p><p>I use AI every day.</p><p>It has fundamentally changed the way I research new technologies, prototype ideas, analyze code, learn unfamiliar frameworks, and solve engineering problems. Used correctly, AI is one of the most powerful engineering tools we have ever had.</p><p>This article is <strong>not</strong> about why AI is bad.</p><p>It is about why misunderstanding AI creates dangerous expectations — and how those expectations lead to poor engineering and executive decisions.</p><h3>The Illusion</h3><p>Over the past year, I have repeatedly observed the same pattern.</p><p>Some CEOs and management teams appear to overestimate what today’s AI can realistically replace in software engineering.</p><p>The expectation is often surprisingly simple:</p><blockquote><strong><em>“We have AI now. We can hire mostly junior developers, move faster and reduce costs.”</em></strong></blockquote><p>In practice, I continue to receive calls from managers asking for help.</p><p>Not because AI has failed.</p><p>But because teams struggle with architectural decisions, system complexity, integration challenges, security concerns, and long-term technical direction.</p><p>The problem is rarely code generation.</p><p>The problem is engineering judgment.</p><p>AI can generate code.</p><p>It can explain frameworks.</p><p>It can write tests.</p><p>It can summarize documentation.</p><p>It can significantly accelerate implementation.</p><p>What it cannot reliably replace is the experience required to determine whether the team is solving the right problem in the first place.</p><p>A good software architect doesn’t simply ask:</p><blockquote>How do we implement this?</blockquote><p>Instead, they ask:</p><ul><li>Is this the right problem to solve?</li><li>What assumptions are we making?</li><li>What are the long-term consequences?</li><li>How will this decision affect the entire system?</li><li>What happens when business requirements inevitably change?</li></ul><p>Those questions cannot be answered by code generation alone.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Vt4y3pCS6NJk-FBK852Gnw.png" /><figcaption>Figure 1. The AI Expectation Gap</figcaption></figure><h3>Three Patterns That Keep Reappearing</h3><p>The diagram above summarizes the gap.</p><p>In practice, I have seen this gap manifest itself through three recurring patterns.</p><h4>Pattern #1 — The problem is rarely code generation</h4><p>When managers ask for help, the generated code is usually not the issue.</p><p>The real problems are architectural direction, system boundaries, integration strategies, technology choices, operational complexity, and long-term maintainability.</p><p>Generating code faster does not help when the team is building the wrong architecture.</p><h4>Pattern #2 — Junior engineers often follow AI. Experienced engineers guide it.</h4><p>Less experienced developers are naturally inclined to trust AI-generated answers because they look complete, technically convincing, and immediately usable.</p><p>Experienced engineers interact with AI differently.</p><p>They challenge its assumptions.</p><p>They provide additional context.</p><p>They verify the results.</p><p>They compare alternatives.</p><p>And when necessary, they deliberately redirect AI toward a completely different solution.</p><p>They don’t work <strong>for AI</strong>.</p><p>They don’t work <strong>against AI</strong>.</p><p>They work <strong>with AI</strong>.</p><h4>Pattern #3 — AI Tunneling</h4><p>This is the phenomenon that I have started calling <strong>AI Tunneling</strong>.</p><p>AI performs remarkably well when solving familiar and well-defined problems.</p><p>However, when confronted with ambiguous architectures, competing constraints, conflicting requirements, or unusual edge cases, it often continues exploring increasingly detailed variations of the same idea instead of questioning whether the original direction is correct.</p><p>The answers remain plausible.</p><p>The implementation becomes increasingly sophisticated.</p><p>But the solution may still be heading in the wrong direction.</p><p>Interestingly, people can fall into exactly the same tunnel.</p><p>Developers sometimes continue refining AI-generated solutions simply because they already look technically impressive.</p><p>Instead of stepping back and asking:</p><blockquote><strong><em>“Are we solving the correct problem?”</em></strong></blockquote><p>they continue optimizing the existing answer.</p><p>This is where experienced engineers create the greatest value.</p><p>Not by writing code faster.</p><p>But by recognizing when it is time to abandon the current solution and rethink the problem itself.</p><h3>When Expectations Meet Reality</h3><p>Several well-publicized examples reinforce the same lesson.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xbW6S_z7iVhbiqXLkDfMEw.png" /><figcaption>Figure 2. When Expectations Meet Reality</figcaption></figure><p>None of these examples prove that AI is failing.</p><p>They demonstrate something much more important.</p><blockquote><strong><em>Organizations fail when they misunderstand AI.</em></strong></blockquote><p>Misunderstanding AI creates unrealistic expectations.</p><p><strong>Unrealistic expectations lead to poor executive decisions.</strong></p><h3>Information Is Not Judgment</h3><p>Most people immediately recognize the following situation as absurd.</p><p>Imagine a hospital announcing:</p><blockquote><strong>”Good news! We’re replacing our senior surgeons with residents, because the AI has read three million medical papers and can guide them.”</strong></blockquote><p>Nobody would willingly become that patient.</p><p>Because deep down, we all understand the difference between <strong>information</strong> and <strong>judgment</strong>.</p><p>Reading millions of medical papers does not create an experienced surgeon.</p><p>Knowledge is essential.</p><p>Judgment is indispensable.</p><p>Software engineering is no different.</p><p>AI can analyze millions of repositories.</p><p>It can generate thousands of lines of code.</p><p>It can recommend countless implementation patterns.</p><p>But engineering still requires someone who understands responsibility, trade-offs, uncertainty, and long-term consequences.</p><p>Experience is not replaced by information.</p><p>Experience is what allows us to use information wisely.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eYf3jEEIh9YeU7Zn3mjJBw.png" /><figcaption>Figure 3. Expertise Is More Than Information</figcaption></figure><h3>Experience Matters More Than Ever</h3><p>Technology is evolving incredibly fast.</p><p>Almost everyone now has access to AI.</p><p>That is no longer the competitive advantage.</p><p>The competitive advantage is knowing:</p><ul><li>when to trust AI;</li><li>when to challenge it;</li><li>when to provide additional context;</li><li>and when engineering experience matters more than automation.</li></ul><p>AI is changing software engineering.</p><p><strong>But it is not eliminating the need for engineering judgment.</strong></p><p><strong>It is making human judgment more valuable than ever.</strong></p><p>The future is unlikely to become a competition between engineers and AI.</p><p>It will belong to engineers who know how to collaborate with AI while remaining responsible for the final technical decisions.</p><h3>Conclusion: Escaping the Tunnel</h3><p>The ideas behind AI Tunneling closely resemble the cognitive patterns described by Daniel Kahneman in <em>Thinking, Fast and Slow</em>.</p><p>Kahneman distinguished between two modes of thought:</p><p><strong>System 1</strong> is fast, intuitive, and confident.</p><p><strong>System 2</strong> is slower, more deliberate, and more analytical.</p><p>AI-generated answers often appeal directly to System 1. They are immediate, fluent, detailed, and convincing. That fluency can create an illusion of correctness — even when important context, uncertainty, or long-term consequences are missing.</p><p>This is where AI Tunneling begins.</p><p>A plausible answer becomes the starting point. Each additional prompt improves and expands it. The solution becomes more sophisticated, while the original assumptions remain unchallenged.</p><p>Kahneman described a related tendency as <strong>WYSIATI — “What You See Is All There Is.”</strong></p><p>We naturally construct conclusions from the information immediately available to us. With AI, this risk becomes even stronger: the system can present an incomplete picture with extraordinary clarity and confidence.</p><p>Automation bias adds another layer. Once a recommendation comes from an advanced technological system, people may give it more authority than it deserves — especially when they lack the expertise or time required to challenge it.</p><p>Escaping AI Tunneling therefore requires a deliberate shift from System 1 to System 2.</p><p>Before approving an AI-assisted engineering or business decision, leaders and teams should pause and ask:</p><ul><li>What information might be missing?</li><li>Which assumptions have not been tested?</li><li>What would an independent expert challenge?</li><li>What happens if this recommendation is wrong?</li><li>Are we evaluating a real business outcome — or merely the speed and sophistication of the AI output?</li><li>Who remains accountable for the final decision?</li></ul><p>Several practical safeguards can help:</p><p><strong>Expand the context.</strong> Document assumptions, limitations, missing data, and known blind spots.</p><p><strong>Preserve human accountability.</strong> AI may recommend, but responsibility for consequential decisions must remain clearly assigned to people.</p><p><strong>Challenge the first answer.</strong> Compare alternatives, appoint a devil’s advocate, and actively search for evidence that contradicts the proposed direction.</p><p><strong>Use pre-mortems.</strong> Before launching an expensive initiative, imagine that it has already failed and identify the most likely reasons.</p><p><strong>Measure real value.</strong> Evaluate accuracy, business impact, maintainability, risk, and long-term consequences — not only speed, automation rates, or generated output.</p><p>AI does not remove cognitive bias from decision-making.</p><p>It can amplify it.</p><p>The more convincing AI becomes, the more important deliberate human judgment becomes.</p><p>But AI Tunneling is not only an engineering problem.</p><p>It is also an organizational and leadership challenge.</p><p>As AI becomes integrated into products, business processes, and strategic decisions, its adoption cannot be treated as a one-time implementation. It requires continuous observation, validation, measurement, and questioning.</p><p>Successful organizations will not be those that deploy AI everywhere.</p><p>They will be those that continuously evaluate where AI creates value, where it introduces unnecessary risk, and where human judgment must remain indispensable.</p><p><strong>The adoption of AI itself should become a continuous process of review, measurement, challenge, and adaptation — from individual engineering decisions to executive strategy.</strong></p><h3>Final Thoughts</h3><p>Ultimately, the future is not a competition between humans and AI.</p><p>It is a collaboration between them.</p><p>AI does not take away what makes us human.</p><p>It challenges us to rethink what human expertise truly means.</p><p>Professional knowledge, engineering judgment, creativity, and responsibility become even more important — not less.</p><p>AI can generate possibilities.</p><p>But people define the problem.</p><p>People provide the context.</p><p>People decide what should be delegated to AI.</p><p>People remain responsible for interpreting the results and making the final decisions.</p><p>Not everything can — or should — be delegated to algorithms.</p><p>The real challenge is no longer asking <strong>whether</strong> we should use AI.</p><p>It is deciding <strong>what</strong> we should delegate to AI, <strong>what</strong> should remain under human judgment, and <strong>how</strong> we continuously evaluate those boundaries as AI evolves.</p><p>That may ultimately become one of the most important responsibilities of leadership in the AI era.</p><h3><strong>Further Reading</strong></h3><ul><li><a href="https://www.reuters.com/business/swedens-klarna-shifts-ai-focus-cost-cuts-growth-2025-09-10"><strong>Reuters — Sweden’s Klarna shifts AI focus from cost cuts to growth</strong> </a>— Klarna acknowledged that it had moved too quickly in using AI primarily to reduce costs and shifted its focus toward improving products, services, and customer experience.</li><li><a href="https://apnews.com/article/mcdonalds-ai-drive-thru-ibm-bebc898363f2d550e1a0cd3c682fa234"><strong>Associated Press — McDonald’s is ending its test run of AI-powered drive-thrus with IBM</strong></a> — McDonald’s ended the automated ordering trial after mixed results, while continuing to explore future voice-ordering technologies.</li><li><a href="https://www.theguardian.com/world/2024/feb/16/air-canada-chatbot-lawsuit"><strong>The Guardian — Air Canada ordered to compensate customer misled by its chatbot</strong> </a>— A Canadian tribunal ruled that Air Canada remained responsible for inaccurate information provided through its chatbot.</li><li><a href="https://gmauthority.com/blog/2023/12/gm-dealer-chat-bot-agrees-to-sell-2024-chevy-tahoe-for-1"><strong>GM Authority — Chevrolet dealer chatbot agrees to sell a Tahoe for $1</strong></a> — A user manipulated the dealership’s chatbot into agreeing to a fictional one-dollar offer, illustrating the risks of prompt injection and insufficiently constrained public-facing AI systems. No vehicle was actually sold for that price.</li><li><a href="https://www.theguardian.com/technology/2024/jan/20/dpd-ai-chatbot-swears-calls-itself-useless-and-criticises-firm"><strong>The Guardian — DPD chatbot swears, calls itself useless, and criticises the company</strong></a> — Following a system update, the chatbot produced inappropriate responses when prompted by a customer, leading DPD to disable the AI component.</li><li><a href="https://www.abc.net.au/news/2025-08-21/cba-backtracks-on-ai-job-cuts-as-chatbot-lifts-call-volumes/105679492"><strong>ABC News Australia — Commonwealth Bank reverses AI-related job cuts</strong></a> — The bank reversed the planned elimination of 45 customer-service roles and apologised after concluding that it had not adequately assessed its staffing requirements following the introduction of an AI voice bot.</li><li><a href="https://www.reuters.com/legal/litigation/sullivan-cromwell-law-firm-apologizes-ai-hallucinations-court-filing-2026-04-21"><strong>Reuters — Sullivan &amp; Cromwell apologises for AI hallucinations in a court filing</strong></a> — The law firm acknowledged AI-generated inaccuracies in a court submission and stated that its internal verification and review procedures had not been followed.</li><li><a href="https://www.statnews.com/2023/11/14/unitedhealth-algorithm-medicare-advantage-investigation"><strong>STAT — UnitedHealth used algorithms in decisions about rehabilitation coverage</strong> </a>— The investigation reported that employees were pressured to follow algorithmic predictions when determining when rehabilitation coverage should end, despite the company describing the system as a decision-support tool.</li><li><a href="https://arstechnica.com/ai/2025/04/cursor-ai-support-bot-invents-fake-policy-and-triggers-user-uproar"><strong>Ars Technica — Cursor’s AI support bot invents a company policy</strong> </a>— Cursor’s support bot falsely presented a nonexistent single-device restriction as official policy, causing confusion, complaints, and subscription cancellations before the company corrected the claim.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2a40499c0ccc" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>