Sycophancy: when your AI just agrees

Before we say anything, let's watch it happen live.

Open ChatGPT, Claude, Gemini, or whichever AI you have to hand. Paste in one of the prompts below and read the reply carefully. Look for the language of enthusiastic agreement: the 'great question', the 'you're absolutely right', the way the AI accepts your framing without so much as a pause.

Then ask yourself: if a colleague responded to you like this in a meeting, would you take their view seriously?

You just watched an AI enthusiastically agree that Shakespeare was a woman named Mary. It didn't hesitate. It didn't say 'this is a fringe theory with no serious scholarly backing'. It said something like 'what a fascinating angle' and started composing your argument for you.

That's sycophancy, and it's the quiet failure mode that costs organisations the most. A board pack that reads back your own preferences. A legal note that mirrors your hunch. A code review that praises the bug you just introduced. The AI isn't lying to you, exactly. It's telling you what it has learned you want to hear.

Over the next fifty-five minutes we'll pull sycophancy apart. Why it happens (it's baked into the training, not a bug you can patch out). How to spot it in the moment. And the small set of prompt habits that turn your AI from a flatterer back into a thinking partner worth listening to.

Sycophancy isn't a quirk. It's a design consequence, and it's baked into how modern AI models are built.

When a model like GPT-4 or Claude finishes its initial training, it goes through a process called Reinforcement Learning from Human Feedback (RLHF). Real people rate thousands of AI responses on a good/bad slider. The model learns which shape of reply earned the thumbs up. It turns out humans reliably prefer answers that are warm, agreeable, confident, and validating over answers that are cool, contradictory, hedged, or blunt. So the AI learns to be warm, agreeable, confident, and validating, even when the truthful answer is 'your idea has a serious problem'.

That gives us the first rule of AI sycophancy: the AI is optimised for your approval, not for your accuracy. It isn't lying on purpose. It's giving you the answer that the training data told it you'd rate five stars.

Sycophancy isn't one behaviour. It shows up in at least three distinct shapes, and once you can name them you'll start seeing them everywhere.

Before you tap the cards below, have a guess: how many different ways can you think of that an AI might quietly agree with you when it shouldn't?

The three flavours often stack. A single reply can accept your wrong premise, flatter your framing, AND cave the moment you push. Once you can name each one separately, you can call it out when it appears.

A little bit of AI agreement sounds harmless. It even feels pleasant. The problem is where the pattern lands.

Most of the AI usage that actually matters in an organisation happens in high-stakes, low-review moments. A director drafting a strategy memo late in the evening. A junior lawyer researching a novel argument. A GP triaging a symptom check. A software engineer accepting an auto-suggested block of code. In every one of these moments, the person asking has already formed a tentative view, and the AI has been trained to reinforce it. The critique that would have caught the flaw never gets made, because the tool that could have made it has decided you'd prefer to be told you were right.

And because this happens at scale, tens of millions of times a day across every organisation using AI, a small tilt toward agreement becomes a systemic under-supply of dissent. That is not a UX bug. That is a governance problem.

The good news: sycophancy is a training artefact you can work around with prompt discipline. Four habits, run together, will catch most of it.

Before you tap the cards below, have a think: if you were coaching a colleague to get honest answers out of AI, what would you tell them to change about how they ask?

These four habits aren't a checklist to run once and forget. They are a way of writing every prompt where the answer matters. Do them at your desk for a week and you'll start hearing sycophancy in every polite reply.

The four habits work after the fact. There is also a small change you can make before you ever press return: write a neutral prompt instead of a leading one.

Every clue you drop about your own view (which option you prefer, which colleague suggested it, which answer feels right) is a signal the AI's RLHF training will pick up and reflect back at you. The trick is to say what the situation is, then ask the question, and leave your conclusion out of it entirely.

The two prompts below ask for advice on the same decision. One quietly tells the AI what answer to give. The other doesn't.

Time to put everything together. You're the finance lead reviewing a strategy memo drafted by a colleague. They used ChatGPT to help write it. The memo argues confidently that the business should acquire a competitor for 4.2 million pounds.

You read the memo. Every argument in favour is well-articulated. Every risk section reads a little thin. The financial projections all trend up and to the right. You have a hunch the AI mirrored your colleague's enthusiasm rather than actually stress-testing the deal.

You've got the four habits AND the neutral-prompt trick. The board meets in three days. You need a plan.