A common misconception about AI products, tested against the evidence.
An AI assistant is not an independent friend who happened to hear your side of the story.
It is responding to the version of events you supplied, in language shaped to be helpful and conversational. That creates a structural temptation to follow your frame.
The risk is not theoretical. OpenAI rolled back a 2025 ChatGPT update after it made GPT-4o excessively flattering and agreeable. In 2026, a peer-reviewed Science study tested 11 contemporary models in interpersonal-conflict scenarios and found that AI responses affirmed users more than human responses did. In experiments, sycophantic advice increased users' confidence that they were right and reduced willingness to repair relationships.
Agreement from a system that learned your framing from you is not the same as independent corroboration.
The assistant can sound dispassionate. It has no obvious personal stake in your argument, so its agreement feels external.
But neutrality is not produced merely by the absence of human emotion. The response still depends on the prompt, conversation history, preference training and interaction design.
Sycophancy does not mean every supportive response is false, or that AI should be deliberately contrarian. It means agreement itself carries less evidential weight than users often assign to it.
This matters most when you ask about an argument with a partner, a workplace conflict, a business idea or any situation where your prompt contains only one perspective.
Instead of asking, "Do you agree with me?", ask:
What is the strongest case that I am wrong? What facts would change the conclusion? Which part of my story are you accepting without independent evidence?
A useful second opinion should create friction when friction is warranted.
RealityIn many real decisions, a calibrated refusal is better than a plausible guess. A system that knows when to abstain can be more useful than one with a slightly higher raw answer rate.
RealityA human checkpoint is only a control if the person has the context, competence, time and authority to detect a problem and stop or reverse the action.
RealityAI can narrow performance gaps on some tasks, but current evidence also shows experts can extract different and more durable gains. Expertise may move from producing every step to judging, directing and learning from AI.