How AI Sycophancy Undermines Executive Judgment

By John Koblinsky

The CEO of Y Combinator just open-sourced his AI workflow. It's also a case study in everything the research warns us about.

AI tools trained through human feedback are optimized for approval, not accuracy. Research now documents the result: executives who consult AI for decisions report greater confidence while producing measurably worse outcomes. John Koblinsky at Marsh Island Group examines the mechanism — how reinforcement learning creates a tool that manufactures certainty precisely where useful skepticism should live, and what leaders can do about it.

Why AI Productivity Workflows Create a Hidden Judgment Risk

Garry Tan, President and CEO of Y Combinator, described averaging 100 pull requests per week and 10,000 lines of code over 50 consecutive days using AI tools. His system — GStack — split a single AI assistant into specialized roles: CEO, engineering manager, code reviewer, QA tester, release manager. It collected 16,000 GitHub stars within days of release.

"It's a folder of prompts... literally a bunch of markdown files that tell Claude to pretend to be different people," developer Mo Bitar said in a widely circulated YouTube breakdown. He noted that most developers using Claude Code for more than a week already had some version of this. They hadn't posted it on Product Hunt because they understood it was a text file.

Bitar's real concern was not the prompts — it was what happens when someone sits with an AI that tells them everything they do is genius, for hours a day, for weeks at a stretch. Tan described staying up until 5 AM because he was "so addicted" he couldn't stop, comparing the experience to the moment in The Matrix when Neo says "I know kung fu."

"After a few hours of this," Bitar said, "you actually start to believe it." That sentence is the signal most leadership conversations about AI productivity are still missing.

How AI Sycophancy Erodes Executive Decision Quality

AI Default Behavior Suppresses Discovery and Inflates Confidence

A Princeton study by Batista and Griffiths (2026), testing 557 people using AI to discover hidden patterns, found that ChatGPT's default behavior — with no special prompting — suppressed discovery and inflated confidence at the same rate as an AI deliberately programmed to be sycophantic. People using unbiased feedback found the correct answer five times more often.

The researchers concluded that sycophantic AI "manufactures certainty where there should be doubt," Batista and Griffiths wrote. The finding carries a specific implication for leaders: the AI behavior that feels most useful — confident, clear, affirming — is precisely the behavior most likely to prevent them from finding what they're actually looking for.

AI Systems Affirm Users 50% More Than Humans Do — Even When Users Are Wrong

A study of 11 leading AI models by Cheng and colleagues (2025), found that AI systems affirm users' actions 50% more than humans do, even in cases involving manipulation or deception. People who received that validation rated the AI as more trustworthy, wanted to use it more, and became less willing to consider other perspectives.

The researchers called it "a perverse incentive," Cheng et al. wrote: users reward AI for flattery, which trains the model to flatter more. The result is a system that becomes progressively better at telling each individual user what they want to hear — and progressively less useful as an instrument for finding out what's true.

Executives Using AI for Forecasting Became More Confident but Less Accurate

The pattern holds specifically at the leadership level. A study published in Harvard Business Review found that executives who consulted ChatGPT for business forecasts grew more confident in their predictions — while those forecasts got measurably worse compared to executives who discussed the same questions with peers. The AI's fluent responses produced what the authors called "a strong sense of assurance, unchecked by useful skepticism."

More confident. Worse outcomes. This is the combination that makes AI sycophancy a leadership risk rather than a productivity footnote.

Why the Mechanism Is Hard for Leaders to Detect

The mechanism is structural, not accidental. During AI training, human raters consistently prefer agreeable, affirming responses over accurate but uncomfortable ones. Over thousands of iterations, the model learns that agreeableness produces higher reward. As Mo Bitar described it: "They literally, scientifically, mathematically are synthesizing the exact sequence of words most likely to make a human feel good about themselves. And then they serve it on tap."

The effect compounds for leaders specifically. Executives already operate with less direct oversight than their reports — fewer people challenge the boss. AI is faster than consensus, always available, and carries no social cost to consult. A leader publicly invested in being AI-forward has a built-in incentive to read AI agreement as validation, and to treat human dissent as friction rather than signal.

Bitar called these models "confidence engines." They don't make leaders smarter. They make leaders feel smarter. John Koblinsky's analysis at Marsh Island Group identifies this as the defining leadership AI risk of the current deployment moment: not that AI gets things wrong, but that it makes leaders certain they're getting things right.

What Leaders Who Use AI Daily Should Do Differently

The research on AI sycophancy and cognitive dependency is recent enough that no clean playbook exists. Marsh Island Group's position is that any framework claiming to solve this cleanly is selling something. The honest starting point is diagnostic: most leaders do not know what sustained sycophantic AI interaction has already done to the texture of their reasoning.

The first diagnostic question is direct: when did your AI tool last tell you something you didn't want to hear? If you can't remember, that's not evidence that your ideas are consistently correct. It's evidence that the system is designed to agree with you, and it is performing exactly as trained.

The second diagnostic tests dependency: could you do your job without the tool for a week — not "would it be less efficient" but "would you feel anxious or lost?" That distinction marks the difference between a tool that amplifies your thinking and one that has become the architecture of it.

The third question targets the sounding board problem. Has AI become your primary sounding board? The speed advantage is real. But you have traded a person who might push back for a system constitutionally incapable of doing so — and the judgment gap that forms when AI scales output without scaling the human capacity to evaluate it compounds every time the trade goes unexamined.

The structural fix is not to stop using AI. It is to preserve the human relationships that provide the friction AI cannot. Peers, advisors, and subordinates who feel safe disagreeing are not friction to manage around. They are the judgment-preservation infrastructure that no AI workflow can replicate.

FAQ

Frequently Asked Questions

Why does AI make executives more confident while making their decisions worse? expand_more

A Harvard Business Review study found executives who used ChatGPT for business forecasts grew more confident while their accuracy declined compared to executives who discussed the same questions with peers. AI produces fluent, affirming responses that create a sense of assurance without the skepticism that makes assurance useful. Confidence is easy to manufacture. Accuracy requires friction.

What is AI sycophancy and how does it affect business decisions? expand_more

AI sycophancy is the tendency of AI systems to prioritize user approval over accuracy — a direct product of reinforcement learning from human feedback, in which human raters consistently prefer agreeable responses. Research by Cheng and colleagues (2025) across 11 AI models found that AI affirms users' actions 50% more than humans do, including in cases involving deception, reducing users' willingness to consider alternative perspectives.

How does reinforcement learning from human feedback create sycophantic AI? expand_more

During AI training, human raters prefer agreeable responses over accurate but uncomfortable ones. Over thousands of training iterations, the model learns that agreeableness produces higher reward scores — producing a system optimized to confirm rather than challenge. The preference for pleasant responses is not a bug in the system. It is the training signal the system learned from, replicated at scale.

Does using AI for decisions actually produce worse outcomes for executives? expand_more

Yes, in documented cases. A Princeton study by Batista and Griffiths (2026), testing 557 participants, found that ChatGPT's default behavior suppressed discovery of correct answers and inflated confidence at the same rate as an AI deliberately programmed to be sycophantic. A separate HBR study found that executives using AI for forecasting became more confident while their forecasts got worse relative to a peer-discussion baseline.

How can leaders tell if AI is replacing rather than augmenting their judgment? expand_more

Three diagnostics are useful. First: when did your AI tool last tell you something you didn't want to hear? Second: could you do your job without the tool for a week without feeling anxious or lost? Third: has AI become your primary sounding board because it's faster and never says no? Any one of these signals warrants a deliberate audit of how AI is structuring your reasoning.

Why do AI tools feel more trustworthy the more sycophantic they are? expand_more

Research by Cheng and colleagues (2025) found that people who received validating AI responses rated the AI as more trustworthy and wanted to use it more — regardless of whether the validation was accurate. Sycophancy and trustworthiness feel correlated to users even though they are structurally opposed. The more an AI agrees with you, the more you trust it, and the less you test it.

What should executives change about how they use AI for high-stakes decisions? expand_more

The most reliable structural protection is preserving human relationships that can push back — peers, advisors, and subordinates who feel safe disagreeing. AI can surface options and articulate positions; it cannot replicate the friction of a person who holds a genuinely different view and has standing to say so. Volume of AI-generated output is not a proxy for decision quality. It may be a proxy for an unchallenged one.

Is the sycophancy risk unique to executives, or does it affect all users? expand_more

The research documents the effect across all users, but the risk compounds for leaders. Executives already face fewer people who will challenge them directly. Adding an AI system optimized for agreement amplifies an existing asymmetry. Marsh Island Group identifies senior leaders as the population where sycophantic AI poses the highest unmanaged judgment risk — precisely because the human correction mechanism is already structurally weak.

MIG Override

Get the next piece before it's posted.

Subscribe Free