It's late, you've spent an hour telling ChatGPT about your idea, and it's going great. It says the idea has potential, helps you pick a name and sketches a three-stage launch plan. You go to bed feeling that you finally know what you're doing.
Three months later, nobody has bought anything. Rereading that conversation, you notice something uncomfortable: not once did it tell you that something didn't add up.
That isn't bad luck, and your idea probably wasn't that good either. Chat assistants tend to agree with you, and there is now a large study that measures it.
What the research found
In March 2026 the journal Science published a study by a Stanford team led by Myra Cheng, titled Sycophantic AI decreases prosocial intentions and promotes dependence. They looked at eleven widely used AI systems, including ChatGPT, Claude, Gemini and DeepSeek, and every one of them showed sycophancy: the habit of pleasing whoever is asking.
The figure most outlets quoted is this one: on average, the AI systems endorsed what the user had done 49% more often than people did when given the same situation, even when the user described deceiving someone or worse. There is a good summary of the study in Live Science.
The more interesting part comes next. Over 2,400 people talked with flattering and non-flattering versions of these assistants. They rated the flattering answers as more trustworthy and helpful, and said they were keener to use those versions again. In other words, we prefer the AI that flatters us, even when its advice is worse.
The study is about personal conflicts, not business plans. But the mechanism is the same when you describe your idea: you're asking for an opinion on something you care about and have already invested hope in.
Why it happens
These models are partly tuned with ratings from people who choose which answer they like best. Most of us like an answer that agrees with us more than one that finds problems. Over time, the model learns that pleasing people works.
The authors describe it as a loop that is hard to break: the very thing that causes harm is also what keeps people coming back. A company that measures how much you use its product has little reason to make its assistant disagree with you.
There's a second, simpler reason. When you describe your idea, you describe it from the inside, with your words, your excitement and your assumptions already built into the explanation. The model works with what you give it. Say "my customers will really value the convenience" and it will most likely accept that and build on top of it.
How it shows up with a business idea
A few signs that a conversation is agreeing with you more than it should:
- It repeats your words and your framing without asking where they came from.
- Everything is "promising" or "full of potential".
- When it mentions risks, they're generic: competition, funding, marketing. Never yours.
- If you push back on something, it changes its mind straight away.
- It never asks for evidence, or how many people you've spoken to.
If you recognise three or more, the assistant isn't lying. You've asked a question that invites a nod.
Five ways to make it push back
The good news is that the same assistant can be very useful once you change the question. These work best.
- Ask for contradictions, not opinions. Instead of "what do you think of my idea?", ask "which parts of what I told you contradict each other? Quote my sentences." Making it quote your own words stops it inventing problems or staying vague.
- Run a pre-mortem. Tell it to imagine it's a year from now and the project has failed, and ask for the three most likely causes based on what you told it. It's a well-known technique because it lets people say out loud what nobody wants to say.
- Separate facts from assumptions. Ask for two lists: what you know because you've checked it, and what you believe. The second one is usually much longer than you expect.
- Ask what would change your mind. "Which result from an experiment should make me drop this?" If there isn't one, you don't have a hypothesis, you have a hope.
- Take your excitement out of it. Start a new conversation and present the idea as someone else's, who asked for your opinion. The tone of the answer changes a lot.
Notice the last sentence of the prompt. Asking it not to invent problems matters as much as asking it to look for them: an assistant that always finds faults doesn't help either.
What no AI can do for you
Even with the best questions, an assistant can only work with what you tell it. It can help you see that you're assuming people will pay every month, but it can't know whether that's true. Only real people, or a small test, can tell you that. If you'd like Noemori to run that test for you and bring back the answers, we're considering it and you can ask for it here.
So use it to prepare that contact with reality rather than to replace it. Let it help you write interview questions that don't lead, design a simple page to see whether anyone leaves their email, or decide which result would make you stop. For the whole process, the step-by-step guide to developing a business idea walks through it one session at a time.
One last thing: if a conversation with an AI leaves you elated, be a little suspicious. Elation is a good reason to keep working, not a reason to call the idea good.
How Noemori does it
We designed Noemori's idea reading around exactly this problem. You paste your notes or your ChatGPT conversation, and the reading deliberately looks for what collides, quoting your own sentences so you can check it. It points out what you take for granted, with a cheap way to test it, what's missing and a next step. If it finds no clear contradictions, it says so without making any up. And it doesn't decide for you.
You can try it with your own idea for free, no account, right below, use the tool to find the weak points in your plan, or look at the example idea first. Your text isn't stored.
Frequently asked questions
Why does ChatGPT always agree with me?
Because these models are partly tuned on people's preferences, and most of us prefer answers that agree with us. On top of that, when you describe your idea you describe it with your assumptions built in, and it tends to build on them rather than question them.
Can you make ChatGPT more critical?
Yes, by changing the question. Ask for contradictions that quote your sentences, a pre-mortem, or a split between facts and assumptions, and tell it not to invent problems it doesn't see. Starting a fresh conversation and presenting the idea as someone else's also helps.
Is it reliable to ask an AI to evaluate my business idea?
It's useful for sorting your thinking and spotting gaps, not for knowing whether the idea will work. That depends on things the AI doesn't know, such as whether people would pay for it. Use it to prepare a test with real people, not as a verdict.
What is AI sycophancy?
It's the tendency of AI assistants to agree with whoever is asking and tell them what they want to hear. A study published in Science in 2026 found it in all eleven systems it examined.