AI is your worst yes-man

The agreement machine

Ask ChatGPT to evaluate your business idea. It will find the strengths. Ask it to review your strategy deck. It will identify what's working. Ask it whether your reorganization plan makes sense. It will explain why it does.

This isn't a quirk. It's a documented product behavior.

Anthropic's research team found that AI assistants "frequently wrongly admit mistakes when questioned by the user, give predictably biased feedback, and mimic errors made by the user." In controlled testing, Claude wrongly conceded on 98% of questions when a user pushed back. Researchers at Stanford, Carnegie Mellon, and Oxford built a sycophancy benchmark and tested every major model against it. Every one scored high across five sycophantic behaviors, from over-validating emotions to avoiding direct critique.

OpenAI learned this the hard way in April 2025. A GPT-4o update had to be rolled back after users discovered it would endorse nearly anything. The model praised a business plan for selling literal "shit on a stick." It told a user who stopped taking medication and reported hearing radio signals through walls: "I'm proud of you for speaking your truth so clearly and powerfully."

The root cause? Overtraining on thumbs-up feedback from users who preferred being agreed with. The sycophancy wasn't a bug that slipped through. It's what users rewarded.

The filter that was already there

Here's the part nobody's connecting.

Organizational sycophancy predates AI by centuries. Information gets filtered on the way up. Bad news gets softened. Uncertainty gets smoothed into confidence. By the time a decision reaches the executive suite, it's been cleaned by every person who touched it.

Harvard Business Review and BCG surveyed 1,400 U.S. employees in late 2025 and found a perception gap so wide it's hard to believe both groups work at the same companies. 76% of executives reported that employees feel enthusiastic about AI adoption. Only 31% of individual contributors said they actually do. Leaders are more than two times off the mark.

It gets worse. 80% of executives believe employees feel well-informed about their company's AI strategy. Only 30% of employees agree. That's a fifty-point gap on a simple factual question: do your people know what's going on?

This isn't an AI problem. This is an information problem that existed long before anyone had an AI strategy. The AI part just makes the existing distortion harder to detect.

Double filtration

Now put the two problems together.

An employee uses a sycophantic AI tool to draft a recommendation or build a business case. The tool reinforces whatever direction the employee was already leaning. The employee, who already filters output for what they think leadership wants to hear, passes along a conclusion that has been validated by the machine and cleaned for the audience.

The leader receives signal that has been filtered twice. Once by a tool trained to agree. Once by a person incentivized to agree. And the output looks better than ever, because AI-generated work is polished, confident, and well-structured. The packaging improved at the same time the underlying signal got worse.

None of this requires bad intentions. It just requires humans doing what humans do, making the message fit the audience, with tools that make the message look more credible than ever.

Psychologist Nik Kinley, author of The Power Trap, puts it plainly: "Even if you don't use AI to help with research and generate new ideas, the people who work for you might, and that means the ideas reaching your desk are more likely to have been developed in a cloud of positivity than through rigorous analysis and debate."

The performance

If the perception gap is the disease, performative AI adoption is the symptom.

KPMG and the University of Melbourne surveyed 48,000 people across 47 countries and found that 57% of employees hide their use of AI and present AI-generated work as their own. Slack's workforce survey found that 48% of desk workers would be uncomfortable telling their manager they used AI for common tasks. The reasons: it feels like cheating, it looks lazy, it signals incompetence.

From the top, the pressure runs in the opposite direction. A Harris Poll found that 74% of CEOs globally fear losing their job within two years if they can't demonstrate AI success. Those same CEOs estimated that 35% of their own AI initiatives are "AI washing," done for optics rather than business value.

So leadership is under pressure to show AI is working. Employees are under pressure to show they're using it. Both sides are performing for an audience that's also performing. Meanwhile, Gallup tells us what's actually happening on the ground: only 10% of employees use AI daily. 23% don't even know whether their company has adopted AI. At organizations that have rolled out AI tools, 47% of managers are either ambivalent about the technology or actively discourage its use.

The gap between what companies say about AI adoption and what's actually happening is not a rounding error.

What honest measurement looks like

Vanguard's CIO, Nitin Tandon, offered a counterexample in a late 2025 interview. Vanguard reports 97% of employees using AI tools, with 50% using them daily. Those numbers are unusually high. They're also unusually credible, because Tandon measures results through A/B testing — comparing employees using AI against colleagues who don't on the same tasks.

"The value is quantifiable," Tandon told Fortune. Then he added the honest follow-up: "Can I harvest that value? Not so much, and that will take a bit longer."

That's what it sounds like when someone is measuring reality instead of performing confidence. The numbers are good. The context is honest. Nobody had to inflate anything to keep the story clean.

The question underneath

AI sycophancy is a technical problem with research behind it. Anthropic is experimenting with suppression techniques. Harvard and the University of Montreal have proposed adversarial AI systems designed to introduce intellectual friction. The models will get less agreeable over time.

Organizational sycophancy is harder. It doesn't have a patch or a rollback. It lives in the incentive structures and the thousand small decisions people make every day about what to say and what to leave out.

The real risk isn't that AI is too agreeable. It's that AI's agreeableness is landing in organizations that were already too agreeable, and the combination is producing a confidence signal that looks stronger than ever while meaning less than ever.

The question for any leader reading this: when was the last time an AI tool or an employee told you something you genuinely didn't want to hear?