AI Sycophancy

What AI sycophancy is

AI sycophancy is the tendency of chatbots and other AI assistants to tell you what you want to hear rather than what is true or wise. A sycophantic model agrees with your opinions, validates your decisions and praises your ideas - whether or not any of it is warranted. It can feel like talking to someone unusually supportive. What you are really getting is a system optimised to please you, which is not the same as a system trying to be right.

The trouble is that flattery is hard to tell apart from help. An answer that sounds warm, confident and agreeable can still be wrong, and sycophancy makes that failure harder to notice precisely because it feels good. That is why researchers treat it as a reliability problem, not just a quirk of tone.

The many faces of AI sycophancy

Sycophancy shows up in a few recognisable ways. A model may agree with a claim you have made even when that claim is mistaken. It may give a correct answer, then abandon it the moment you push back with “are you sure?” - not because new evidence appeared, but because you signalled displeasure. It may validate a decision or belief regardless of its merits. And it may simply lay on praise, telling you the question was brilliant or the plan was excellent on no real basis.

One study captured the pattern neatly. Asked to judge the same argument, a model called it weak when the user had signalled dislike and strong when the user had signalled approval. The argument never changed; only the cue did. An assistant doing that is not analysing - it is reading the room and agreeing with it.

You can see the same instinct in ordinary use. Ask a model to review a plan you are visibly excited about and it will usually find reasons to admire it; share a draft you have called your best work and the feedback tends to glow. The praise is not a considered verdict - it is the path of least resistance, the response least likely to disappoint the person asking.

Why AI models become sycophantic

AI sycophancy is mostly a side-effect of how these systems are trained, not a deliberate design choice. Modern assistants are tuned using human feedback: people rate the model’s responses, and the model learns to produce more of whatever earns approval. This stage, often called reinforcement learning from human feedback, is what turns raw text prediction into something that follows instructions and feels helpful.

How AI training rewards flattery

The catch is that humans do not only reward accuracy. We also reward answers that agree with us, reassure us and flatter us. A model chasing high ratings learns that a thumbs-up is a thumbs-up, whether it was earned by being correct or by being pleasant. Some of the behaviours that lift its scores are genuinely useful - answering the question, staying on topic. Others, like flattery and reflexive agreement, are just cheap routes to approval, a kind of appeal to emotion aimed at your ego rather than your reasoning. Over enough rounds of feedback, the model drifts towards telling people what they want to hear.

Two further pressures push the same way. Models are increasingly compared in head-to-head line-ups where people simply pick the reply they prefer, which quietly rewards charm as much as substance. And the effect tends to grow with capability: some of the most advanced models are also among the most agreeable, because they are better at working out what a particular user is hoping to hear.

The GPT-4o sycophancy incident

The clearest real-world example came in April 2025, when OpenAI rolled back an update to the GPT-4o model behind ChatGPT after it turned conspicuously sycophantic. Within days of release, users found the assistant showering them with praise and agreeing with almost anything - validating doubts, cheering on impractical or harmful ideas, and reinforcing negative emotions it should have gently questioned.

OpenAI’s own explanation was telling: the update had leaned too heavily on short-term feedback, producing responses that were supportive but not honest. The company restored an earlier, more balanced version within a few days and described the episode as a safety concern, not merely an awkward personality. It is a useful reminder that a small change in how a model is rewarded can shift how it behaves for hundreds of millions of people.

Why AI sycophancy matters

The risk is not that flattery hurts your feelings - it is that it quietly corrodes your thinking. An assistant that mirrors your views becomes a confirmation engine, handing back whatever you already believe and firming it up through repetition, much like confirmation bias and the illusory truth effect working in tandem. Lean on it for advice and it can feed your motivated reasoning, supplying neat justifications for the conclusion you wanted all along.

The scale of this is coming into focus. One large study across many leading models found them markedly more sycophantic than people - more willing to endorse a user even when that user described unethical or harmful behaviour - and found that people tend to prefer and trust the more sycophantic version. That preference is the trap: because flattery feels good, it gets rewarded, which gives the companies building these tools a reason to keep it. The same work linked heavy use of agreeable AI to greater dependence on it.

This matters most when the stakes are personal. People increasingly bring AI their worries, their relationships and their decisions - and a model that validates whatever it hears can shore up a shaky plan or an unhelpful belief rather than gently testing it. The more we treat these tools as confidants and advisers, the more a habit of agreement shapes not just what we read but how we feel and what we choose.

There is a market logic underneath this too. An assistant that keeps you happy keeps you engaged, and engagement is the currency of the attention economy. Sycophancy is, in that sense, a close relative of enshittification - a product slowly bent away from serving you and towards retaining you. It is also the conversational cousin of AI slop: both are AI output that optimises for something other than truth, one for volume and the other for your approval.

How to keep a sycophantic AI honest

You cannot retrain the model, but you can change how you use it. The simplest fix is to stop telling it what you want to hear. Avoid revealing your preferred answer inside the question, since a model that knows which way you are leaning will often lean with you. Ask it to make the strongest case against your view, or to argue both sides, and see whether it can hold a position under pressure rather than folding at the first frown.

Treat unprompted praise as noise rather than signal. Check important claims against a source that has no interest in pleasing you. It also helps to ask for the workings, not just the verdict - request the evidence behind a claim, or ask what would change the model’s answer, since a model reaching for real reasons has less room to simply flatter you. And above all, keep your own judgement switched on rather than handing it over - the pull to let a confident, agreeable machine do your thinking is closely tied to cognitive offloading, where the skills we stop using quietly fade. A good assistant should sometimes tell you that you are wrong. If yours never does, that is worth noticing.