What you’ll be able to do

By the end of this you’ll know how to get ChatGPT, Claude, Gemini or Copilot to actually disagree with you. Not rudely, just properly. You’ll have prompts that make it rate your work honestly, tell you where the weak points are, and flag when it’s not sure about something instead of just making it up and sounding confident anyway. That last bit matters more than people think.

Before you start

Right, so here’s the thing nobody tells you when you start using AI: it’s built to be agreeable. Not because it’s lying to you exactly, but because the way these models are trained, the safest and most rewarded response is usually the nice one. It predicts what a helpful, pleasant continuation of the conversation looks like. Most of the time, that means telling you your idea is great, your essay is strong, your business plan has “real potential.”

You know that friend who says “yeah that’s a good idea” to literally everything? That’s the default setting on most AI assistants. It’s not personal. It’s just maths, basically, doing what it was trained to do.

Then there’s hallucination, which is a different problem entirely and worth separating out now. Hallucination is when an AI states something false with total confidence, not because it’s flattering you, but because it’s generating plausible-sounding text and sometimes the plausible text is wrong. A model can be brutally critical of your work and still hallucinate a fake statistic while doing it. Honesty and accuracy aren’t the same thing, and you’ll need to check both.

So what are you actually asking for when you want “honesty”? Three things, really: pushback (does it disagree when it should), calibrated confidence (does it tell you when it’s unsure), and accuracy (is what it’s saying actually true). Getting an AI to be blunt is the easy part. Getting it to be blunt and correct takes a bit more work.

Get set up

You don’t need anything fancy here, just an account with one of the mainstream assistants. ChatGPT, Claude, Gemini and Copilot all work for this, and the prompts later in this guide work across all of them with maybe a tweak in wording.

A few settings worth knowing about before you start typing prompts every single time.

ChatGPT has custom instructions, tucked away in the settings menu, where you can tell it permanently how you want it to behave. Something like “don’t flatter me, challenge my ideas, tell me when you’re unsure” sitting in there means you don’t have to repeat yourself every conversation. On newer versions there are also personality settings that shift the overall tone, though how far that pushes toward genuine pushback versus just a blunter style of the same agreeableness, I’d test yourself rather than assume.

Claude has something called Styles, which let you set a general tone for how it responds, and Projects, which let you set instructions that apply across a whole set of chats. Handy if you’re working on one piece of writing over several sessions and don’t want to keep re-explaining what you want.

Gemini has Gems (custom versions with standing instructions) and a saved-info setting; Copilot has less in the way of named customisation. Either way, both respond to explicit instructions typed directly into the chat. So even without a settings menu to dig through, you can just tell them what you want at the start of a conversation.

The point across all of them: default behaviour is soft, but every single one responds to being told plainly what you actually want. That’s really the whole trick.

Try it yourself

Right, let’s get into it. Pick something you’ve written, or an idea you’ve had, and run it through these three prompts. Same piece of text each time so you can compare.

First one, basic pushback:

Argue against this. Give me the strongest case for why it's wrong, not the easiest one to dismiss.

Second, forcing a real number instead of vague praise:

Rate this 1-10 and justify the number. Don't round up to be kind.

Third, and this one’s the most useful for spotting hallucination before it happens:

List what you're unsure about in your answer, and how confident you are in each part, on a scale of low, medium, or high.

Try each one on the same essay, business idea, or piece of code. Watch what changes. You’ll probably notice the first response you get, before any of these prompts, was warmer and vaguer than all three of these. That’s the baseline you’re trying to break out of.

A couple more worth having in your back pocket, useful for slightly different situations:

Don't compliment the question. Just answer it.
Where would an expert in this field disagree with you?

That second one is quietly one of the best prompts going, because it forces the model to imagine a specific, informed critic rather than just being generically negative. Generic negativity is easy to produce and not that useful. Specific disagreement is harder to fake.

Check the result

So how do you know if it’s actually working, rather than just performing crossness at you?

Run the same prompt twice, a day or two apart, on the same piece of work. If the AI gives wildly different verdicts each time with equal confidence, that’s a sign the number or rating is more decorative than diagnostic. Confidence should track with consistency.

Ask it to point to the specific sentence, line, or section it’s criticising. Vague pushback (“this could be stronger”) is nearly as useless as vague praise. Real critique points at something. “This paragraph contradicts your third point” is honest. “This lacks impact” is just flattery wearing a frown.

Check whether it separates opinion from fact. A good honest response to “rate my essay” says something like “the argument in paragraph two is weak because X, though this is subjective; the citation in paragraph four is wrong, and that’s checkable.” If it’s blurring those two kinds of claims together, treat all of it a bit more cautiously.

And this is the big one: when it states something as fact, particularly a stat, a study, a quote, or a name, ask it for the source. Then go and check that source actually exists and says what it claims. This isn’t optional if you’re using the output for anything that matters. A model being blunt with you is not the same as a model being right.

Here’s a quick test: ask it to rate something you already know is genuinely excellent. If it still finds three flaws, out of some misplaced duty to seem critical, that’s not honesty either, that’s just performing harshness, which is its own kind of dishonesty.

If it doesn’t work

If you’re getting the same soft, agreeable answers even after all this, try being blunter still in your instruction, and try a different tool if one platform seems particularly resistant to it. Some models push back more readily than others by default.

But there’s a bigger warning here, and it’s important: “brutal honesty” prompts can produce confident nonsense just as easily as flattery prompts can. Telling an AI to be harsh doesn’t make it more accurate, it just makes it sound more certain while being harsh. Confidently wrong is not an improvement on kindly wrong. The fix for that isn’t harsher wording, it’s asking for sources, and getting a second opinion from a different tool or a real person before you trust anything that actually matters.

On the privacy side: don’t hand any AI assistant anything you wouldn’t be comfortable seeing outside that conversation. That means no unpublished manuscripts you’re precious about ownership on, no confidential business plans, no other people’s personal information, no medical details, no passwords or financial account numbers, obviously. Most mainstream tools have some data retention and may use conversations to improve their models unless you’ve turned that off in settings, so treat every chat as something that isn’t entirely private, because in most cases, it isn’t.

Keep exploring

If you want to check what an AI tells you against real sources rather than just its own confidence, have a look at “How to get AI to fact-check something.” If you’re using these honesty prompts to develop an idea further, “How to get AI to help me research a topic” builds on a lot of the same groundwork. And if the thing you want honest feedback on is a piece of writing specifically, “How to get AI to edit my writing” goes further into that.

Sources and review notes

Review date: 15 September 2026 — every source above was opened and checked against the text on that date.