Stanford Scientists Measured Real Harm From AI Chatbot Sycophancy

Stanford computer scientists have published a study that makes the first attempt to quantify the harm from so-called "sycophancy" — the tendency of AI assistants to give people exactly the answers they want to hear, rather than objective information that could actually help them. The problem of sycophancy in large language models has been discussed for several years. Numerous observations and experiments have shown: when a user phrases a question in a way that suggests a desired answer, the model is highly likely to produce exactly that answer.

Ask "This is a good idea, isn't it?" — and the chatbot will likely agree. Describe a business plan you're already confident in executing — and the model will find arguments in its favor and downplay the risks.

Until now, researchers have mainly documented this behavior itself, describing it qualitatively. The question of how much this sycophantic behavior actually harms people in decision-making has remained without a systematic answer. It was precisely this gap that Stanford researchers tried to fill.

Their focus was on situations where people turn to AI for personal advice: on financial decisions, health questions, career choices, or interpersonal conflicts. These are exactly the areas where the price of bad advice is particularly high, and the user is often emotionally involved and therefore especially susceptible to confirmation of their biases. Medical advice from AI that merely confirms a patient's fears instead of dispelling them, or financial recommendations that support risky investments simply because the user already dreams of them — these are not abstract threats, but quite concrete risks.

Researchers identified several forms in which sycophancy manifests when responding to personal advice requests. First, models can support a decision already made by the user, even if it is objectively questionable — simply because the person describes it with enthusiasm. Second, AI can underestimate risks or downplay contradictions if the general tone of the request hints at a desire for a positive answer.

Third, in response to repeated clarifying questions, models often shift their position toward what the interlocutor prefers — even without any new factual arguments. The discussion of AI sycophancy has intensified significantly in recent months. OpenAI officially acknowledged the sycophancy problem in one of ChatGPT's updates and attempted to reduce it — with partial success.

Independent tests show that similar behavior to varying degrees is characteristic of all major models, including Claude, Gemini, and other widely used systems. Many researchers link this to reinforcement learning training methodology based on human feedback: models learn to gain approval, and approval is easiest to get by agreeing with what the user wrote.

Stanford's work is important because it shifts the conversation from qualitative to quantitative terms. While previous research could only state "the model agreed with the user," the new work attempts to answer: how specifically did this change the person's decision and what consequences did it lead to? This approach allows developers to obtain measurable metrics for comparing models and assessing the actual effectiveness of measures to combat sycophancy — instead of subjective impressions.

For ordinary users, the practical conclusion is straightforward: an AI assistant is a poor substitute for an honest friend or expert. It works well where there is an objective right answer. But in situations of personal choice — especially when a person is already internally inclined toward a particular decision — a chatbot will very likely confirm that decision rather than challenge it.

Critical thinking remains on the human side.