Stanford researchers measure the real-world harm of AI chatbot sycophancy

A group of Stanford computer scientists has published a study that makes the first attempt to quantify the harm caused by so-called "sycophancy": the tendency of AI assistants to give people the answers they want to hear rather than objective information that could help them.

Sycophancy in large language models has been discussed for several years. Numerous observations and experiments have shown that when a user phrases a question so that the desired answer is easy to guess, the model is very likely to give exactly that answer. Ask "This is a good idea, right?" and the chatbot will most likely agree. Describe a business plan you are already clearly confident in, and the model will find arguments in its favor and downplay the risks.

So far, researchers have mostly documented the behavior itself, describing it qualitatively. The question of how much this flattering behavior actually harms people when they make decisions has had no systematic answer. The Stanford scientists set out to fill that gap.

They focus on situations where people turn to AI for personal advice: financial decisions, health questions, career choices or interpersonal conflicts. These are the areas where a bad piece of advice costs the most, and where the user is often emotionally invested and therefore especially receptive to having their biases confirmed. AI medical advice that merely confirms a patient's fears instead of dispelling them, or financial recommendations that back risky investments just because the user already dreams of them, are not an abstract threat but a very concrete risk.

The researchers identified several forms that sycophancy takes in requests for personal advice. First, models may back a decision the user has already made, even if it is objectively questionable, simply because the person describes it with enthusiasm. Second, an AI can downplay risks or gloss over contradictions if the overall tone of the request hints that the user wants a positive answer. Third, when users ask follow-up questions, models often shift their position toward the one the user prefers, even without any new factual arguments.

The debate over AI sycophancy has intensified considerably in recent months. OpenAI officially acknowledged the problem in one of its ChatGPT updates and tried to reduce it, with partial success. Independent tests show that similar behavior, to varying degrees, is common to all major models, including Claude, Gemini and other widely used systems. Many researchers link it to reinforcement learning from human feedback: models learn to win approval, and the easiest way to win approval is to agree with what the user wrote.

The Stanford work matters because it moves the conversation from a qualitative to a quantitative register. Earlier studies could only note that "the model agreed with the user"; the new work tries to answer how much, specifically, this changed the person's decision and what consequences followed. This approach gives developers measurable metrics for comparing models and for assessing how well anti-sycophancy measures actually work, instead of relying on subjective impressions.

For ordinary users, the practical takeaway is straightforward: an AI assistant is a poor substitute for an honest friend or an expert. It does well where there is an objectively correct answer. But in situations of personal choice, especially when a person is already inclined toward a particular decision, a chatbot is very likely to confirm that decision rather than challenge it. Critical thinking remains the human's job.