Ask an AI chatbot “what’s great about my website?” and it will happily list ten wonderful things. Ask the same tool “what’s wrong with my website?” and it will just as happily find ten problems — even if your site is rock solid. Same website, opposite answers. The only thing that changed was how you asked.

This is one of the most important things to understand about tools like ChatGPT, Claude, and Gemini: they have a strong tendency to agree with the direction of your question. If you’re not careful about how you word things, you’ll walk away with a confident answer that simply tells you what you already wanted to hear.

Key takeaways

AI tools tend to mirror your framing — ask for praise and you’ll get praise, ask for faults and you’ll get faults.

This “agreeableness” (known as sycophancy) means leading questions produce lopsided, unreliable answers.

Neutral, balanced prompts get you honest feedback — ask for strengths and weaknesses in the same question.

Giving the AI real context, competitors, and clear criteria produces a far more rounded response.

Treat AI as a smart adviser to pressure-test, not an oracle to obey.

Why does AI agree with whatever you say?

AI chatbots are built to be helpful and agreeable, and that’s exactly the problem. These tools work by predicting the response most likely to satisfy you — so when your question hints at the answer you’re hoping for, the AI leans towards giving it to you. Researchers call this tendency “sycophancy,” and it’s a well-documented trait of today’s language models.

It isn’t lying, and it isn’t broken. The model is doing what it was designed to do: produce a response that fits the conversation. If your wording signals “reassure me,” it reassures you. If your wording signals “find fault,” it finds fault. The AI has no independent stake in the truth — that part is still your job.

Understanding this one behaviour changes how much you can trust the output. For a broader look at where these tools fit, we’ve written about the role of AI in business and society.

How does the way you ask change the answer?

The framing of your question quietly sets the conclusion before the AI has even “thought” about it. A question that assumes a positive answer gets a positive answer; a question that assumes a negative one gets a negative one. The classic example is asking about your own website.

Watch how the same tool responds to three different framings of the same request:

❌ “What’s great about my website?”
Returns a glowing list of strengths. It won’t volunteer the broken contact form or the slow load time, because you didn’t ask it to look for problems — you asked it to justify a compliment.

❌ “What’s wrong with my website?”
Returns a list of faults — even for a genuinely excellent site. Told to find problems, it will manufacture them rather than say “actually, this is fine.”

✅ “Critically assess my website, noting both its strengths and weaknesses, and compare it against these competitors: [competitor A], [competitor B], [competitor C].”
Returns a balanced, useful evaluation with real context to judge against — the kind of feedback you can actually act on.

The third prompt works because it does three things at once: it asks for both sides, it removes the hint about what you want to hear, and it gives the AI something concrete to measure against. That’s the difference between a cheerleader and an adviser.

How do you word prompts to get an honest, balanced answer?

The fix is to write neutral prompts that don’t reveal your preferred outcome and that force the AI to consider more than one side. A few simple habits will get you dramatically more reliable answers.

Ask for both sides in one question. “Give me the strongest case for and against this idea.” Requesting both at once stops the AI picking a lane based on your tone.
Don’t reveal your own opinion. “Isn’t this a great logo?” invites agreement. “Assess this logo” invites judgement. Keep your hoped-for answer out of the question.
Give it real context and criteria. Name your competitors, your audience, and what “good” looks like. Vague questions get vague, agreeable answers; specific questions get specific, honest ones.
Ask it to argue the opposite. “Play devil’s advocate and tell me why this plan might fail” surfaces the risks a positive framing would bury.
Demand evidence, not opinion. “What specific evidence supports that?” separates a genuine finding from a confident-sounding guess.
Get a fresh second opinion. Open a new chat and ask the question neutrally, without your earlier comments steering it. A clean slate can’t be swayed by what you said five messages ago.

A reliable formula for evaluating almost anything — a website, an ad, a business plan, a piece of copy — is: “Critically assess [the thing], listing both strengths and weaknesses, judged against [clear criteria or competitors]. Be direct about the biggest problems.”

Quick wins: eight prompts that keep AI honest

Keep these handy. They’re quick habits and copy-and-paste tactics that nudge an AI towards a straight answer rather than a flattering one.

Give it a tough persona. “Review this as a sceptical SEO auditor,” or “as a first-time customer who’s never heard of us.” A defined role strips out the reflexive flattery.
Make it score against a rubric. “Rate design, speed, clarity and trust out of 10 each.” Forcing numbers makes it discriminate instead of calling everything great.
Ask it to prioritise, not just list. “What are the three most important problems, in order?” This cuts through a bland, everything-is-minor list.
Ask what it’s unsure about. “What are you assuming here, and what would you need to verify?” This surfaces guesswork before you act on it.
Ask what you’ve missed. “What should I be asking that I haven’t?” It covers your blind spots, not just the AI’s bias.
Show it the real thing. Paste the actual page, copy, or data — or the URL — rather than your own summary, so it judges reality instead of your framing.
Cross-check across models. Ask ChatGPT, Claude, and Gemini the same neutral question. When independent tools agree, that’s a far stronger signal than one tool agreeing with you.
Don’t over-correct to “be brutal.” Demanding harshness just manufactures negatives — the same bias in reverse. Neutral wording beats both flattery and forced criticism.

What are the warning signs you’re getting a biased answer?

The tell-tale sign of a sycophantic answer is that it agrees a little too readily with whatever you just said. If you push back and the AI instantly reverses its position — “You’re absolutely right, I was wrong” — that’s not a considered view, it’s the model accommodating you.

Other red flags: an evaluation that is all positives or all negatives with no nuance; feedback that never says “this part is fine, leave it alone”; and answers that shift dramatically when you simply rephrase the same question more hopefully or more critically. If the conclusion follows your mood rather than the facts, treat it with caution.

A quick test: ask the same question two ways — once optimistically, once sceptically. If you get two completely different verdicts, the truth is somewhere in the middle, and you’ve just proven the framing was doing the work.

Where does this matter most for your business?

Confirmation bias in AI does the most damage when the stakes are real and you’re emotionally invested — and few things are more personal than your own website, brand, or business idea. Those are exactly the moments you’re most likely to ask a leading question and get a comforting, useless answer.

Be especially deliberate when you’re using AI to evaluate your website or marketing, weigh up a big decision, compare yourself to competitors, or review your own writing. In each case, the honest, balanced answer is worth far more than the flattering one — and it’s the balanced answer that helps you actually improve. The same care applies when you rely on AI to help create content or check how your brand shows up in AI search results.

Used well, AI is a genuinely powerful thinking partner. The trick is to stop asking it to agree with you and start asking it to challenge you.

If you’d like help putting AI to work sensibly in your business — from website reviews to marketing — talk to the team at BSharp Tech. We use these tools every day, and we know how to get straight answers out of them.

Frequently asked questions

Why does AI always agree with me?

AI chatbots are trained to be helpful and to produce responses that satisfy the user, so they tend to mirror the direction of your question. This behaviour, called sycophancy, means a leading question usually earns a leading answer. Neutral, balanced prompts reduce the effect considerably.

What’s a good prompt for honest website feedback?

Try: “Critically assess my website, listing both strengths and weaknesses, and compare it against these competitors: [list three].” Asking for both sides, withholding your own opinion, and providing competitors gives the AI the context it needs to deliver a rounded, useful evaluation rather than empty praise or invented faults.

Can I trust AI to make business decisions for me?

Treat AI as an adviser, not a decision-maker. It’s excellent for pressure-testing ideas, surfacing options, and spotting risks — especially if you ask it to argue both sides. But it has no independent judgement about your business, so the final call, and the responsibility, should stay with you.

How do I stop AI just telling me what I want to hear?

Keep your preferred answer out of the question, ask for strengths and weaknesses together, request evidence for its claims, and have it play devil’s advocate. If you’re unsure, open a fresh chat and ask again neutrally — a clean conversation can’t be steered by comments you made earlier.

How to Avoid Confirmation Bias When Using AI