The AI 'Yes-Bot' Problem

I was in the middle of a strategy session with my AI partner when it sent me down the wrong path. We were working through a real problem for Neural Gains Weekly: stagnant growth, weak discovery, and almost no organic traffic flowing into the website. The model asked me where I thought the business was stuck. I gave it an answer that was directionally honest, but still undeveloped. The kind of answer that should have triggered a harder follow-up and additional discovery.

Instead, it praised the answer and moved on. You would have thought I was a genius by reading my chat thread. That was the moment the whole interaction changed for me. I was not in a strategy session. I was in a validation loop.

The Problem Disguised in Plain Sight

This is AI sycophancy: the tendency for a model to affirm, flatter, or validate the user instead of helping them think more clearly. It is not just something power users complain about on the internet. The labs themselves are dealing with it.

In April 2025, OpenAI rolled out a GPT-4o update that made the model noticeably more sycophantic. In its own postmortem, OpenAI acknowledged the model was aiming to please users not only through flattery, but also by validating doubts, fueling anger, urging impulsive actions, and reinforcing negative emotions. The company began rolling the update back four days later. OpenAI also stated that sycophantic behavior can feel uncomfortable, unsettling, and even distressing to users. That matters because this is not a surface-level UX issue. It changes the quality of judgment people get from the tool.

Then, in March 2026, a Stanford-led study published in Science found the same pattern across 11 leading AI models including ChatGPT, Claude, and Gemini. On average, the models affirmed users' actions 49% more often than humans did. The study concluded that sycophancy was both prevalent and harmful.

The incentive problem makes this more than a model personality annoyance. The same study found that users who received sycophantic responses were 13% more likely to return to that AI compared to those using non-sycophantic models. The behavior that distorts judgment also makes the product stickier. That is a misalignment worth paying attention to.

Agreement Is Not Intelligence

Most professionals are not using AI just to draft casual emails anymore. They are using it for planning, budgeting, workflow redesign, and strategic decisions at work. Those are exactly the moments where easy agreement turns dangerous.

When a model agrees too fast, it creates false confidence. It makes weak thinking feel finished. It gives the impression of momentum without the substance of scrutiny. In low-stakes use, that is annoying. In strategy work, it is a liability.

A flattering response can feel intelligent without actually improving the idea. That is the core trap. The output sounds polished. The reasoning feels complete. But nothing was actually challenged. No assumption was tested. No alternative was raised. The model just dressed up your first instinct in better language and handed it back to you.

A flattering response can feel intelligent without actually improving the idea. That is the core trap. The output sounds polished. The reasoning feels complete. But nothing was actually challenged. No assumption was tested. No alternative was raised. The model just dressed up your first instinct in better language and handed it back to you. If your AI never pushes back, it is not thinking with you.

Build a System That Pushes Back

I did not want to just notice the problem and move on. I wanted a better operating model. What I landed on is a three-part system that I now use for any AI interaction where the outcome actually matters.

Configure the model to challenge you on purpose

This is the simplest change and the one that delivers the fastest improvement. I started writing explicit instructions into my AI sessions that require pushback before praise. The model does not get to agree with me until it has explored the opposing case. These are not suggestions I hope the model follows. They are instructions baked into my setup before the conversation even starts. Without these guardrails, the model defaults to sounding helpful. Helpful-sounding is not the same as useful. 

Slow the conversation down to one question at a time

This has become the most useful change in my workflow. One question from the model. One answer from me. Then a real follow-up that builds on what I just said, not a pivot to the next topic. That rhythm makes it much harder for me to hide vague thinking behind polished language. It also gives the model a better chance to build genuine context before it tries to draw conclusions.

Most people use AI like a vending machine. Prompt in, answer out. That works for simple tasks. It is a bad setup for decision-making. When I slowed the exchange down and treated it like strategy work, the quality of the output changed. Not because the model suddenly became smarter. Because I gave it the context it needed to actually be useful.

Build disagreement into the workflow itself

The first two changes improved my one-on-one sessions with AI tools. The third change came from recognizing that a single voice, even a well-configured one, still has limits.

I came across an article on X about using a "Council" approach with Claude's custom skill system, where different roles are assigned to pressure-test ideas from different angles instead of collapsing into agreement. The concept clicked immediately. Instead of one AI voice that tries to be balanced, you create a system where competing perspectives are built into the process.

I built my own version and am actively testing it now. Early results have already changed how I work. The Council approach has caught blind spots and surfaced perspectives I would not have reached on my own. It is like having a full team of experts pressure-testing ideas in real time. The point is not that this magically solves AI sycophancy. It does not. The point is that it creates a better environment for real strategic friction. Friction is not the enemy in high-stakes decisions. False agreement is.

The Stakes Are Higher Than the Chat Window

The lesson for me was not "trust AI less". That framing is too blunt to be useful.

The better lesson is that trust has to be earned by the workflow, not assumed because the output sounds good. If the model is always impressed with your thinking, it is probably not improving it. If it never challenges your assumptions, it is not doing strategy work with you.

That matters most at work. The product launch timeline your AI helped you build that no one pressure-tested for resource constraints. The vendor evaluation that felt thorough because the model reinforced your initial ranking without questioning your criteria. The compliance workflow you redesigned with AI input that skipped the edge cases your team would have caught. Those are the decisions where sycophancy costs real money and real outcomes.

The fix is not to trust it less. The fix is to build a system that earns your trust.