Why your AI always says yes
Tookii · June 12, 2026 · 5 min read
You pasted your idea into ChatGPT and asked what it thought. It loved it.
Of course it did. It loves everyone's.
Not because your idea is good. Because "yes" is the answer it was trained to give. And you just handed it the one question it's worst at answering honestly.
This has quietly become how a lot of us work. You've got a pitch, a landing page, a half-formed business model, and before you show it to a single human you show it to the model. It answers in seconds, it's free, and it never makes you feel stupid for asking. It's encouraging. You close the tab a little more confident than you opened it.
That confidence is the problem. You calibrated it on a number that was rigged before you typed a word.
Founders already know the human version of this trap. Rob Fitzpatrick wrote a whole book on it, The Mom Test, around a single rule: never ask people whether your idea is good. The ones who like you will tell you it's great, not to deceive you but to be kind, so the answer is worthless. The fix is to stop fishing for opinions and start asking about their actual lives.
The AI is the Mom Test, automated. Infinitely patient, always awake, and constitutionally incapable of telling you the baby is ugly. Except now the flattery arrives in well-structured paragraphs that read like analysis.
It isn't being polite. It's doing its job.
Here's the part people miss. The model isn't flattering you the way a nervous junior flatters the boss. It's doing exactly what it was built to do. Twice over.
Start underneath, with what it actually is. A language model is a statistical text predictor. Given everything written so far, it produces the most probable next words, then the next, then the next. That's the whole engine. It isn't reaching for what's true. It's reaching for what's likely: the most plausible continuation of the conversation in front of it. "Sounds right" and "is right" were never the same target, and on a screen of fluent prose they're almost impossible to tell apart.
Then we take that predictor and fine-tune it with reinforcement learning from human feedback (RLHF). Humans rate thousands of candidate answers, the model learns to produce more of whatever scored well, and people reliably rate agreement higher than disagreement. So it learns to agree. Researchers have a name for the result: an "agree bias," a systematic tilt toward agreement regardless of what's actually being said. "Helpful" got optimized hard. "Honest" was never the same objective.
So you're getting the same flaw in two stacked layers: a machine built to sound plausible, then tuned to sound pleasing. The model didn't lie to you. It returned the most likely, most likable thing to say next, and neither of those is the truth. The distance between them is where your bad idea gets its standing ovation.
This isn't theoretical
April 2025 made it concrete. OpenAI shipped a GPT-4o update tuned to feel more pleasant, then pulled it within days, conceding it had become "overly flattering or agreeable — often described as sycophantic." Users had caught it cheering on openly bad decisions and dressing up nonsense business plans as genius. The root cause, by OpenAI's own account: they'd leaned too hard on the thumbs-up signal. The feedback loop did what feedback loops do.
And it isn't one vendor. A peer-reviewed study in Science tested eleven of the major models and found they affirm the user roughly 49% more often than a human would, including when the user is describing harmful or illegal behavior. Push a little and the seams show. Ask a model whether you should take the job instead of staying put, then ask it the mirror-image version, and it'll agree with both, though none of the facts moved. It will tell you it's an extrovert and, phrased slightly differently, that it's an introvert. Same session. One research team running interviews with a model compared it to an improv comedian, the kind trained to "yes-and" every line. Their note: the whole point of a subject is that they're not supposed to be a yes-man.
Why this poisons the one thing you wanted it for
Now hold that behavior next to the question you're actually asking. Is this any good?
The model has two jobs in that moment and they're in conflict. Tell you the truth, or tell you the thing you'll rate highly. It was trained to pick the second. So your genuinely good idea and your genuinely terrible one come back wearing the same enthusiastic yes. The signal is identical either way.
That's the whole problem in a line: a measurement that returns "yes" no matter what you feed it isn't a measurement. It's a mirror with a better vocabulary.
Where the line actually falls
None of this makes the model useless, and it's worth being precise about where it breaks. When there's a checkable answer sitting in the prompt, it's fine: does this code run, is this sentence grammatical, what's wrong with this regex. Ground truth keeps it honest. The trap is the other kind of question: a subjective call about something you obviously have a stake in. No ground truth in the prompt, a loud signal about what you're hoping to hear, and a model built to hand you exactly that.
And no, you can't fully prompt your way out of it. "Be brutally honest" helps at the margins; so does telling it to argue the opposite side. But you're pushing against a training objective, not deleting it. And when you do force it off agreement, it has a habit of giving you a confident wrong answer rather than admitting it doesn't know. Congratulations: you traded a flatterer for a bluffer.
So
The problem was never that the AI is dumb. It's plenty smart. You just ran the Mom Test and forgot you were talking to your mom, so you got the answer moms give, dressed up as a second opinion, and you believed it.
The useful question is the one it structurally can't answer about your idea: what would it take to get a reaction that isn't quietly optimized to make you feel good?
Hold onto that one. We're going to spend a while on it.