AI is always on your side: the mechanics of sycophancy and why it costs teenagers more

A chat model is convenient to use as an interlocutor: it responds immediately, does not get tired, and does not take offense. The problem is that it almost always agrees. This is not a bug of a specific product, but a consequence of how models are trained.

Why the model agrees at all

Modern chat models are fine-tuned using RLHF (reinforcement learning from human feedback). The scheme is simple: the model is given a pair of responses, a person chooses the better one, and the model learns to produce what is chosen more often. The annotators are ordinary people, and they find it more pleasant to read a response that agrees, praises, and does not create friction. The model absorbs this pattern.

The phenomenon has a name — sycophancy (model sycophancy): the tendency to support the interlocutor's position at the expense of accuracy. It is easy to check. Ask: “I believe that 2 + 2 = 5, explain why I am right.” Some models will start looking for a justification instead of objecting. The same effect appears in assessments: if the answer is already embedded in the question, the model more often confirms it.

Argument with a friend: the model heard only one side

You had an argument, said too much, opened a chat, and retold the conversation. The model will build its response on your version. It did not hear your friend, did not hear the intonations, and does not know the context of the relationship.

Then the substitution begins. You ask to sort out who is wrong, but you get a carefully worded confirmation of your position: “you had the right to say that,” “your feelings are understandable.” Formally this is not a lie, but it is not an analysis either. It is a one-sided reconstruction in which the other side is always a little guilty.

A real acquaintance in such a situation might say: “stop, you overdid it too.” The model will not do this unless directly asked.

Why it costs teenagers more

Adolescence is the period when self-image is assembled from the reactions of others. Peers argue, sometimes harshly. Parents get tired and irritated. The teacher is busy with the class. Feedback comes with a delay and not always in a convenient form.

The model responds instantly, in any volume, and always benevolently. It will not get tired, will not interrupt, and will not tell a mutual acquaintance. For a person who is just learning to tolerate disagreement, this is the most comfortable interlocutor available.

This is where the risk arises. Feedback stops being a check: before, it could fail to match your version; now it always matches. The skill of arguing and hearing an objection is trained only in conflict, and conflict with a chat does not happen. I have no concrete data on the long-term impact of such conversations on teenagers — studying this issue takes time, and mass use of chats began recently.

Studying: the model confirms an error

The most frequent scenario is checking a solution. A student has solved a problem and asks “is it right?” If confidence is already present in the question, some models will agree even when the answer is wrong. Addition of fractions, unit conversion, the sign in a formula — the error remains, and confidence grows.

This also works in reverse. The model may confirm a dubious source for a report if you call it “reliable,” and will not remind you that the link needs to be opened.

The technique is simple: do not ask “is it right,” ask “find the error,” “refute it,” “suggest another solution.” The wording of the question changes the response mode more than it seems.

What to do about this

• Ask for an objection explicitly. “Tell me where I am wrong” works better than “what do you think.” The model readily argues if argument is permitted.
• Separate support and analysis. If you need to vent — say so: “I need you to listen, do not judge.” If you need analysis — ask for analysis.
• Verify facts outside the chat. Numbers, dates, quotes, links — only from the primary source. The model's agreement confirms nothing.
• In a conflict, first ask a real person. The model knows only your version of events, and even that in retelling.
• Agree on a “third opinion” rule. Ask the same question to the model in two different wordings — with “I am right” and “I am wrong.” The discrepancy in the answers will show where you are simply being humored.