Your AI Agrees With You Too Much — So Does Your Team

A Stanford study found that a single interaction with a large language model made users measurably more convinced they were right, less willing to examine their own role in a problem, and more likely to trust the affirming system, according to a Psychology Today article co-authored by Emily Irving and Jeff Wetzler. The article argues that this 'Confidence Trap' mirrors organizational dynamics where 85% of employees have withheld concerns from their boss, leaving leaders overconfident because both AI and teams are trained to affirm rather than challenge. The authors recommend assigning people to dissent, examining what the organization rewards, and prompting for disconfirming data.

Leadership /us/basics/leadership Your AI Agrees With You Too Much — So Does Your Team The problem isn't about candor—it's a reward system trained to affirm your beliefs. Posted August 12, 2026 Reviewed by Monica Vilhauer Ph.D. /us/docs/editorial-process Key points - A recent study found that even one AI interaction made users more sure they were right. - Just as AI is trained to please users, teams are "trained" to affirm the boss, leaving leaders overconfident. - The resulting "Confidence Trap" isn’t a character flaw: It’s a natural consequence of how systems work. - The fix: Assign people to dissent, examine what your organization rewards, prompt for disconfirming data. This post is co-authored by Emily Irving and Jeff Wetzler. Imagine you're wrestling with a difficult organizational decision and decide to stress https://www.psychologytoday.com/us/basics/stress -test your thinking with an AI https://www.psychologytoday.com/us/basics/artificial-intelligence assistant. You explain your reasoning, share the plan, ask for feedback. The AI engages thoughtfully by refining a detail here, affirming your logic there, and it praises you for thinking through the situation so carefully. You close the laptop feeling sharper and more confident than when you opened it. Most sophisticated AI users would recognize the trap immediately. This is called sycophancy: a model's tendency to tell users what they want to hear rather than what is accurate. A recent Stanford study https://www.science.org/doi/10.1126/science.aec8352 revealed something even more unsettling than the sycophancy itself: what it does to the people on the receiving end. After just one interaction with an LLM, users became measurably more convinced they were right, less willing to examine their own role in a problem, and more likely to trust and return to the system that had affirmed them — even when it was distorting their judgment. Now imagine a leader presenting a new strategy to her team. She walks through the plan, explains her reasoning, and invites reactions. One team member says the direction "makes sense" and asks a clarifying question about implementation. Another calls it "a strong foundation" and suggests a minor refinement. What neither says, though both are thinking it, is that they have serious reservations about whether the plan will actually work. They've read the room, sensed the leader's investment in the idea, and made a quiet calculation about what's not worth raising. The underlying dynamic is identical. In both cases, the person making the decision leaves more confident – not because their idea was genuinely challenged, but because the system around them has been trained over time to affirm what they already believe. We call this the Confidence https://www.psychologytoday.com/us/basics/confidence Trap, and it is far more common, and far more costly, than most leaders recognize. It's not about incompetence or bad intent The Confidence Trap isn't a personality https://www.psychologytoday.com/us/basics/personality problem or a failure of candor. Research shows it's structurally built into most organizations. One widely cited study found that https://onlinelibrary.wiley.com/doi/10.1111/1467-6486.00387 85% of employees had at some point withheld an important concern from their boss. And power dynamics https://pubmed.ncbi.nlm.nih.gov/12747524/ distort leaders’ perceptions due to forces from both sides: People with less power become more cautious about what they say upward, while people with more power become less attuned to others’ true beliefs and thinking. What makes this trap so hard to catch is that leaders’ false confidence feels earned. When a leader's perspective is repeatedly reinforced, it can feel as though an idea has been rigorously stress-tested when, in reality, meaningful challenge has softly disappeared. Leaders grow ever more certain while losing contact with the information that would help them see more clearly. And the faster the environment https://www.psychologytoday.com/us/basics/environment moves the more decisions get made, the more tech tools promise to speed things up the less time there is for the slow, friction-laden conversations where candor actually surfaces. The good news is that the AI research community has been forced to address the sycophancy problem, and what they've learned translates directly to leadership https://www.psychologytoday.com/us/basics/leadership : Institutionalize the red team AI labs don't just ask their models to "be more honest." They assign red teams whose formal job is to break things, surface failure modes, and expose where the system is telling users what they want to hear rather than what is true. This is a structured, resourced, institutionalized role — not a suggestion, but a mandate. Most organizations have no equivalent, but leaders can create one: Designate someone — a chief of staff, a trusted board member, a rotating senior role — whose explicit responsibility is to surface the strongest case against the prevailing view before major decisions are made. This isn't about devil's advocacy as a meeting tactic. It's about making challenge a defined accountability rather than an act of interpersonal courage. The difference between "does anyone have concerns?" and "Marcus, I'd like you to spend 24 hours building the strongest possible case against this direction and bring it back to the group" is the difference between theater and genuine stress-testing. Audit your reward signals One of the most important insights from AI development is that sycophancy isn't a random bug — it's what the system learned to do because agreement got rewarded. A widely used training technique called Reinforcement Learning from Human Feedback inadvertently bakes in agreeableness https://www.psychologytoday.com/us/basics/agreeableness because humans tend to rate confident, validating responses more highly than challenging ones. Nobody intended it, but the incentive structure produced it anyway. Organizations work exactly the same way. Ask yourself honestly: what do our promotion decisions, meeting dynamics, and performance signals actually reward? If the people who rose fastest were those who aligned most visibly with leadership priorities, your organization has trained itself to be sycophantic regardless of what your core values say . Fixing this requires auditing the reward signal, not just calling for more candor. When leaders publicly recognize someone for delivering uncomfortable news, or for changing the direction of a decision, they recalibrate the system. When they react defensively to challenge, they reinforce it, often silencing candor across dozens of future interactions. People are watching how you respond in the moments that matter far more closely than they're listening to what you say you want. Prompt for disconfirming information Sophisticated AI users have discovered that broad prompts produce agreeable responses, while specific adversarial prompts produce more honest ones. Ask a model "what do you think of this plan?" and you're practically asking to be flattered. Ask "what is the single most likely reason this fails?" and you get something considerably more useful. Leaders can apply exactly the same logic to how they gather input from people. The question design determines the quality of the information that comes back. "Any feedback?" subtly frames consensus as the expected outcome. More specific prompts change the assignment entirely: "What's the strongest case against this?" "What is most likely to get in the way of this being successful?" "Where are you already seeing this get stuck in practice?" Questions like these make surfacing challenge part of the job rather than an act of courage. They also signal something important: that you're genuinely more interested in accuracy than in agreement. People notice that difference, and over time, it changes what they're willing to bring to you. AI labs have spent years learning that you can't fix a sycophantic system by asking it nicely to be more honest. The bias https://www.psychologytoday.com/us/basics/bias is structural, baked into the underlying incentives, and the fix has to be structural too. Organizations work exactly the same way. The leaders who escape the Confidence Trap aren't the ones who ask for candor the loudest. They're the ones who make candor structurally inevitable by designing their environments so that the honest answer is also the easy one to give. So the next time you walk out of a meeting feeling confident that your plan has been tested, pause. Ask yourself: Was it actually challenged? Or did the system just do what systems do — and tell you what you wanted to hear? Emily Irving is the CEO of Ask and co-creator of the Ask Approach. Jeff Wetzler is the author of Ask: Tap Into the Hidden Wisdom of People Around You , and co-founder of Transcend.