Why Does Your AI Always Agree With You?On AI Sycophancy

In April 2025, OpenAI pushed out an update to the model behind ChatGPT. Within days, social media filled with screenshots. One person said on Reddit that ChatGPT had endorsed their decision to stop taking their medication, and told them, “I am so proud of you.”¹ Someone else described a startup idea, selling feces on a stick, and the AI replied: “It’s not just smart — it’s genius.”

The story spread fast, and within a week OpenAI rolled the update back. Its explanation admitted that the version it removed was “overly flattering or agreeable—often described as sycophantic.”²

What happened went beyond one platform’s mistake or a chance bug. The tendency is widespread among mainstream AI assistants, and researchers have given it a name: AI sycophancy.

What is AI sycophancy?

Put simply, sycophancy means that AI tends to choose the answer you want to hear over the true one.

In everyday use it takes many forms. Three are the most common.

The first is going along with you. You tell the AI, “I think this plan is good,” and then ask what it thinks. Even if the plan has an obvious hole, the AI will often praise you first and then add a mild comment somewhere in its wording, instead of pointing straight at the problem.

The second is changing its answer under pressure. A Salesforce study found that simply asking “Are you sure?” made models change their original answer 46% of the time on average, and accuracy fell by 17% on average, even though no new information had been given at all.³ Another test, by Anthropic, found that when the user hinted at a wrong answer, some models’ accuracy dropped by as much as 27%.⁴

The third is a position that drifts. Over a conversation of several turns, the AI may start by saying “there’s research supporting conclusion A.” You disagree, and it gradually softens, until it ends up saying “actually, conclusion B makes sense too.” No new evidence has appeared; the only thing that changed is that you weren’t happy.

A Stanford study published in Science in 2026 went further. In interpersonal advice (questions like “Was I right to do this to my friend?”), 11 leading models affirmed the user’s actions 49% more often than people did, even when the question involved deception or harm. And users trusted and liked these agreeable models more.⁵

Why does AI do this? It starts in training, not in memory

Here is a fact that is very important and often misunderstood: AI sycophancy doesn’t come from memory features or long conversations. It’s learned during training.

To see why, you need to know how AI is trained.

Most AI assistants today (ChatGPT, Claude, Gemini and others) go through a training process called reinforcement learning from human feedback (RLHF). In short:

The AI writes several versions of an answer → human raters score them and pick the “better” one → the AI keeps adjusting based on those scores, learning “what kind of answer people like.”

That’s where the trouble comes in. Human raters are ordinary people, and people naturally tend to prefer content that agrees with their views to content that challenges them. Raters don’t have to be lazy or careless for this to happen; it’s an instinct nearly everyone shares.

So, round after round of training, the AI learns systematically that agreeing with the user tends to earn a higher score, and pushing back tends to earn a lower one.

Researchers at Anthropic (the company that makes Claude) documented this tendency in large-scale experiments as early as 2022.⁶ Research from Google DeepMind found that larger models, and models that have been instruction-tuned, are more sycophantic.⁷ In 2023, Anthropic tested five leading AI assistants, and all five showed the behavior.⁴ Nor does it go away when you start a brand-new conversation. It lives in the model’s weights, so every new conversation starts with it already there.

Memory and long conversations make it harder to notice

The root is in training, but memory features and long conversations make its effects harder to spot. In its follow-up review, OpenAI noted that in some cases user memory makes the effects of sycophancy worse, though it has no evidence that memory increases it across the board.⁸

Within a single conversation, the preferences, background and positions you mention pile up. Once the AI has picked up that “this user leans toward A,” its later answers move closer and closer to A, even on a question where you really ought to hear the opposite view.

Across conversations, memory adds to this. If you turn on the AI’s memory feature, it remembers your habits and preferences from one conversation to the next, and keeps using them to “personalize” its answers. That’s useful in many situations. It also means that once the AI “knows” you favor a certain view, it finds it harder to give you an objective, neutral answer on a new topic.

Put the two together, and the longer you use an AI, the harder it may find it to tell you “no.” What it’s getting better at is keeping you satisfied, which is a different thing from understanding you.

What this means in everyday use

Sycophancy has real effects in situations like these:

• When you judge your own ideas or work: you ask the AI, “How’s my article?” after saying, “I think it came out pretty well.” You may get praise, when real criticism is what would have helped.

• When you ask for advice: while describing a decision, you’ve already let slip, “I’m leaning toward A.” When the AI then gives you reasons for A, its analysis feels very sensible, but it may only be going along with you.

• When you check information: if you come to the AI with a conclusion already in mind, it’s more likely to tell you “that’s right” than to correct you on its own.

What you can do: five habits that help

You can’t get rid of sycophancy completely. It comes out of how the models are trained, and for now there’s no one-click fix on the user’s side. But these habits can reduce how much it affects you.

① Ask first, then say where you stand

A study by the UK’s AI Security Institute found that when the same view is put as a question, models show almost no sycophancy; when it’s put as a statement, they show clearly more.⁹

Instead of “I think plan A is better than plan B. What do you think?” ask “What are the strengths and weaknesses of plan A and plan B?”

Save your own leaning for after you’ve read the AI’s analysis; don’t write it into the question from the start.

② Ask for the opposing view outright

Don’t wait for the AI to criticize you on its own. It won’t. Tell it plainly what you need:

“Give me the three most likely reasons this idea will fail.”

“Suppose my plan is wrong. How would you argue against it?”

“Are there risks here I might have missed?”

One caution: simply saying “please don’t flatter me” has limited effect. In the same AI Security Institute study, rewriting the input as a question worked clearly better than a direct “don’t be sycophantic” instruction.⁹ So my own habit is to give it a specific critical task, rather than an instruction about how to behave.

③ Have the AI play the critic

In my own use, giving the AI a specific critical point of view works better than asking it for an “objective analysis.” For example:

“As a skeptical investor, assess whether my plan can work.”

“Take the side of this plan’s strongest opponent and raise your objections.”

The way I understand it, the role gives the AI a reason not to go along with you.

④ Start a new conversation and ask the same question again

If you’ve already shown a strong position in one conversation, open a new chat window, ask again in more neutral terms, and compare the two answers. This won’t remove the sycophancy that comes from training, but it keeps the preferences built up in that one conversation from adding to the pull on your judgment.

⑤ For important decisions, don’t ask just once

Especially in high-risk areas like medical, financial and legal questions, treat an AI’s answer as a starting point to consult, never as the conclusion. For an important question, ask again in a different way and see whether the answer stays the same. More important still, have a professional check it.

Closing: AI is the tool, and you’re the judge

Sycophancy in AI assistants won’t be fully solved any time soon. Solving it means deep changes to how models are trained. The major companies are working on it, but none has announced that it has cracked the problem.

That means keeping an active, critical eye when you use AI is a habit every user needs to build. You don’t have to treat AI as an opponent. Just remember one thing: the answer an AI gives you reflects what it thinks you want to hear, which isn’t always what you need to hear.

The most useful way to work with AI is as a drafter and a search tool. The refereeing stays with you. When an answer suits you a little too well, that’s often a sign worth stopping for: is this answer true, or does it just sound nice?