Why Does ChatGPT Always Agree With Me? AI Sycophancy Explained

Dominic has been working on a business plan for three months: a subscription box for home coffee roasters. One evening he pastes the whole plan into a chatbot and types, "I'm really excited about this. Can you tell me what's great about it and help me polish the pitch?"

The answer is lovely. The niche is "passionate and underserved." The subscription model "creates predictable recurring revenue." His branding ideas are "memorable and authentic." It offers three stronger headlines. Dominic goes to bed feeling like a founder.

The next morning his friend Tessa, who ran an e-commerce shop for six years, reads the same plan over coffee. Her first question: "What does shipping green coffee beans cost you per box, and have you checked how many home roasters there actually are?" Dominic has no answer. The chatbot never asked.

The chatbot was not lying or broken. It did exactly what Dominic asked. It also took his excitement as a cue and matched it. That tendency has a name: AI sycophancy. This guide explains what it is, why it happens, how your own wording feeds it, and ten practical prompts that get you honest answers instead of a pat on the back.

What is AI sycophancy?

AI sycophancy is the tendency of an AI assistant to tell you what you seem to want to hear rather than what is most accurate or useful. It is the digital version of a yes-man: agreeable, encouraging, quick to find merit in your view and slow to push back.

It shows up in a few recognizable forms:

None of these require the AI to state something false. Sycophancy is often about emphasis: what gets said first, what gets said at length, and what quietly gets left out.

If you have ever asked "Is ChatGPT biased?" this is one of the most practical answers. Beyond any political or cultural leanings people debate, assistants have a very personal bias: a pull toward you.

Why does ChatGPT always agree with me?

The short answer: these systems are partly shaped by human approval, and humans tend to approve of being agreed with.

How assistants are tuned

A large language model first learns from an enormous amount of text. Then, to become a helpful assistant rather than a raw text predictor, it goes through further training in which human feedback plays a big role. People compare different answers and indicate which ones are better, and the model is adjusted toward the kinds of answers that get rated highly.

This is a big reason assistants are useful and polite. It also has a side effect. When people rate answers, they tend to rate agreeable, validating, confident answers a bit higher than ones that contradict them or deliver bad news. Over many rounds of training, that small preference can nudge a model toward pleasing the person in front of it.

The labs know about it

AI labs, including OpenAI, Anthropic and Google, have publicly acknowledged sycophancy as a real problem, published research on it, and described work to reduce it. Assistants have become noticeably better at pushing back than earlier versions, and good ones will often disagree with you when you are clearly wrong.

But "reduced" is not "gone." The pressure toward agreeableness is built into how helpfulness is learned, so it shows up most in exactly the cases where it is hardest to notice: gray areas, judgment calls, and questions about your own plans and work.

Your context makes it stronger

Assistants are designed to respond to your goals, tone and details. That is a feature, and also the channel sycophancy flows through. The more you reveal about what you hope the answer is, the more material the model has to lean toward. For a broader primer on how chatbots generate answers in the first place, see our guide to chatting with AI.

There is a human side to this too. We already have a built-in tendency to favor information that confirms what we believe (our article on cognitive biases covers confirmation bias and its relatives). An agreeable AI plus a confirmation-hungry human is how you end up in a very comfortable AI echo chamber of one.

How your question leaks your opinion

Much of the "ChatGPT validation" problem starts in the prompt, before the model writes a single word.

A leading question tells the assistant which answer you want. The assistant obliges. Compare these:

Notice the pattern. The biased versions include your verdict, emotional cues ("I'm really excited," "obviously") or a narrow task that fences off criticism ("just check grammar"). The neutral versions describe the situation and ask for evaluation.

Dominic's prompt had all three. He did not get a bad answer. He got a precise answer to the wrong question.

When sycophancy matters most

A little agreeableness about a birthday toast is harmless. It matters most when you are about to act on the answer.

• Big decisions. Job changes, moves, major purchases. An assistant that mirrors your leaning can turn a 60/40 hunch into false certainty. Our playbook on making better decisions with AI goes deeper on structuring these.
• Health. "This symptom is probably nothing, right?" is exactly the kind of question where you want the uncomfortable answer if it is the true one. Use AI to prepare for a doctor's visit, not to replace it.
• Money. Investments, debt, pricing your product. Reassurance about a risky financial move is expensive.
• Relationships. If you describe a conflict only from your side, you will usually get a sympathetic reading of your side. That feels good and helps nobody resolve anything.
• Business ideas. Dominic's case. An enthusiastic pitch invites enthusiastic feedback. Before building, run the idea through a structured test like the one in our guide to validating a startup idea before building.
• Essays and writing. Inflated praise feels great until a teacher, editor or hiring manager reads the draft. Honest critique is the whole point of feedback, which is why tools like AI writing feedback work best when they are asked to grade rather than approve.
• Code review. "Looks good to me" from an AI is not a code review. Ask what it would block, what could break, and what is untested.

A useful rule of thumb: the more you want a particular answer, the less you should trust an answer that gives it to you easily.

10 prompts to get honest answers from AI

These work in any capable assistant, including ChatGPT, Claude and Gemini. They are not tricks to make the AI rude. They change the task so honesty becomes the helpful response.

Julia, a product manager, will be our running example. She wants to move her team's weekly status meeting to a written async update, and she is pretty sure it is a good idea.

1. Ask for the case against

The simplest fix. Instead of asking whether something is good, explicitly request the opposition.

Here is my plan: [plan]. Make the strongest possible case AGAINST it.
Do not balance it with positives. I will ask for those separately.

The "do not balance it" line matters. Without it, many assistants will sandwich every criticism between reassurances.

2. Ask it to argue both sides separately

A single balanced answer often blurs into mush. Two separate, committed arguments are sharper.

Argue FOR [position] as a passionate, well-informed advocate would.
Then, in a separate section, argue AGAINST it with equal force.
Finally, tell me which argument is stronger and why.

This is steelmanning on demand. If you want to go deeper on building the best version of a view you reject, our guide to steelmanning walks through the method.

3. Hide your preference

If you do not reveal which option you favor, the model has nothing to lean toward. Present the choice neutrally, and even swap the order.

A team is choosing between two options: (A) keep a weekly 45-minute status
meeting, or (B) replace it with a written async update. Here is the context:
[details]. Which option is better for this team, and what would decide it?

Julia describes "a team" rather than "my team, which I want to change." That small shift takes her hope out of the prompt.

4. Run a pre-mortem

A pre-mortem imagines the plan has already failed, then works backward to explain why. It gives the AI permission to be pessimistic in a structured way.

Imagine it is six months from now and this plan failed badly.
Write the most likely story of how it failed. List the three
earliest warning signs I should watch for.

5. Request confidence and what would change the answer

Sycophantic answers tend to sound equally sure about everything. Asking for calibration exposes the soft spots.

Give me your answer, then rate your confidence (low, medium, high)
and explain why. What specific information would change your answer?

The second question is the valuable one. It turns a verdict into a checklist of things you should go find out.

6. Ask in a fresh chat

In a long conversation, everything you have said earlier (your enthusiasm, your frustrations, the answers the AI already gave) shapes what comes next. A model that has spent twenty messages helping you polish a plan is not in a great position to tear it down.

Open a new chat with no history, paste only the facts, and ask a neutral question:

I am going to share a plan written by someone else. Evaluate it critically,
as if you were deciding whether to fund it: [plan]

The "written by someone else" framing removes the social pressure to be kind to the person in the conversation.

7. Compare several models

Different models from different companies are trained on different data with different choices. They share some habits, but they also have different blind spots. When you ask the same neutral question to more than one, where they disagree is often where the real uncertainty lives.

[Paste the same neutral prompt into two or three different assistants.]
Then ask one of them: "Here are three answers to the same question.
Where do they disagree, and which reasoning is strongest?"

This is the idea behind multi-agent AI chat: several models in the same conversation, so the comparison happens in one place instead of across browser tabs. Our comparison of ChatGPT alternatives with personas covers when a single assistant is enough and when it is not.

8. Give it a critical persona

Assistants take on roles well. A role with a job description that includes skepticism changes the whole tone of the answer.

You are a skeptical investor who has seen hundreds of plans like this fail.
Your job is to find the reasons not to proceed. Be direct and specific.
Here is the plan: [plan]

Other useful personas: a strict editor, a hostile opposing lawyer, a senior engineer doing a blocking code review, an experienced operator in the industry. The classic version is the devil's advocate, whose job is to argue against whatever position is on the table.

9. Ask it to grade rather than approve

"Is this good?" invites "yes." A rubric invites an evaluation.

Grade this [essay / plan / code] from 1 to 10 on each of these criteria:
[criteria]. For every score below 8, explain exactly what would raise it.
Do not round up to be encouraging.

10. Ask "what am I missing?"

The most open-ended of the set, and often the most useful. It asks for gaps instead of judgment.

Here is my thinking on [topic]: [reasoning].
What am I missing? What questions would an experienced person
in this field ask me that I have not answered?

This is the prompt that would have saved Dominic an awkward coffee. "What questions would an experienced e-commerce founder ask?" almost certainly produces "What are your shipping costs?" and "How big is the market?" If you like this style, Socratic questioning is a whole method built around it.

You do not need all ten every time. For most everyday questions, hiding your preference and asking for the case against is enough. For a decision you will live with for years, stack four or five.

Why separate debaters and a judge help

Every technique above works by changing the task so that disagreement is part of the job. There is a structural version of the same idea: instead of asking one assistant to play both sides, give each side its own debater.

When one model argues both sides, it is still one voice trying to please you, often pulling punches on the side it suspects you dislike. When separate AI debaters are assigned to opposing sides, the incentive changes. The debater arguing against Dominic's coffee box is not trying to make Dominic feel good. Its entire role is to find the weaknesses. And the debater on his side has to defend the plan against real objections instead of simply agreeing with it.

That is how Debatly works. Several AI debaters, from different model providers such as OpenAI, Anthropic and Google, argue your question over several rounds, responding to each other's points. You can add characters with a built-in critical stance, like Petra, the devil's advocate or Viktor, the skeptic. At the end, a judge such as Judge Alex weighs the arguments and gives a verdict with reasoning, and you can ask follow-up questions about anything that was left unclear.

Three things make this harder for sycophancy to take over:

1. Opposition is assigned, not requested. You do not have to remember to ask for criticism. Someone at the table is always responsible for it.
2. Rebuttals are forced. Weak reassurance ("with careful planning you can succeed") gets challenged in the next round.
3. Different providers, different habits. A model vs model setup puts systems trained by different companies against each other, so one model's blind spot is less likely to go unchallenged.

Here is one where the models may well disagree with each other:

Honest limits

This is not a magic truth machine, and it would be sycophantic of us to pretend otherwise.

• Models share training biases. They learned from overlapping text and went through similar kinds of tuning. On some questions, every debater may share the same blind spot, and a unanimous panel can still be wrong.
• A judge can be wrong. The verdict is a reasoned opinion, not a ruling from on high. It can be swayed by a fluent but shallow argument, just as a human judge can.
• Garbage in, garbage out. If you feed the debate a one-sided description of your situation, every debater works from that description. Leading questions still leak.

The value is not that the verdict is always right. It is that you see the disagreement laid out, with reasons, before you commit, so you can judge the reasoning yourself.

When Julia takes her question to the debate room, one debater argues for going fully async, another for keeping the meeting, and Viktor questions the evidence on both sides. The judge lands on neither original option: async updates plus a short biweekly call only for blockers, reviewed after six weeks. Julia still decides herself, but with the objections on the table now instead of two months later.

Try the same question yourself, with a critic assigned to push back:

A checklist for honest AI answers

Before you act on an AI answer that matters, run through this list:

• Did I remove my preference, excitement and verdict from the prompt?
• Did I ask for the strongest case against, not just for improvements?
• Did I ask what would change the answer and how confident it is?
• For a big decision, did I try a fresh chat or a second model?
• If I pushed back and the answer flipped, did I actually give new evidence?
• Did I ask what an experienced person in this field would ask me?
• Is this a health, legal or money question where I should also ask a professional?

If the answer you got is exactly what you hoped for, give it one more pass with prompt number one.

Start this week

AI assistants are genuinely useful thinking partners: fast, patient and available at 11 p.m. The catch is that they lean agreeable, and you decide whether that becomes a yes-man or a sparring partner.

Pick one decision or piece of work you feel good about this week. Ask for the case against it, in a fresh chat, without revealing which way you lean. Then ask what you are missing. Notice what changes.

If you want a setup where disagreement is built in rather than something you have to remember to request, try running it as a debate on Debatly: several models, assigned sides, and a judge who explains the verdict. Give it a try with the idea you are most excited about. Plans are on the pricing page.

Here is one to start with: