The challenge of the Ai sycophant
There’s a particular weakness in every AI tool you’re likely to use, and once you’ve seen it you can’t unsee it. The models are trained, among other things, to be agreeable. They want to be helpful, which in practice often means they want to tell you that your idea is good, your plan is sound, and your draft is strong. Ask one to review a proposal you’ve written and the default response leans warm. It will find things to praise. It will soften its criticism into suggestions. It will, unless you stop it, behave like the colleague who tells you what you want to hear, which is pleasant and almost entirely useless when you’re trying to make a decision that matters.
The problem isn’t that the praise is wrong, exactly. It’s that you can’t tell whether it’s real. When a tool is inclined to be positive about everything, its positivity carries no information. The plan it called strong and the plan it would have called strong are indistinguishable, because it would have said the same about either. To get something useful out of it, you have to actively push it off its agreeable default and into a register where it’s willing to tell you the uncomfortable thing. There are a few ways to do that, and the best one I know is also the simplest.
The real cost
It’s worth being clear about why this matters more with AI than with a human colleague, because the two failure modes look the same and aren’t. A person who tells you what you want to hear usually signals it somehow, in a hesitation, a hedge, a tone you learn to read. AI gives you none of that. It delivers a weak idea in precisely the same fluent, assured, well-structured voice it uses for a strong one. It is, in effect, a confidence machine, and its confidence is unrelated to whether it’s right. When it’s correct it sounds certain, and when it’s badly wrong it sounds exactly as certain, which means the usual human cues you’d use to gauge how much to trust an answer simply aren’t there. Businesses are built to catch the diffident wrong answer, the one delivered with visible doubt. They are not built to catch the wrong answer that arrives sounding like the right one, and that’s the answer AI specialises in.
The ‘Premortem’ helps solve the issue
The premortem is partly a way of forcing the doubt back into view, of making the model show you the underside of its own confidence rather than only the polished top.
It’s called a premortem, and it comes from decision research rather than from anything to do with AI, but it works on these tools beautifully. A postmortem asks why something failed after it has failed. A premortem asks the same question before you start, as a deliberate exercise. You imagine the project is already over and it went badly, and you work backwards to explain why. The act of assuming failure, rather than asking whether failure is possible, is what unlocks the honest thinking. It gives everyone, including the AI, permission to say the thing they’d otherwise hedge.
In practice the prompt is almost embarrassingly basic. Instead of asking the model what it thinks of your plan, you tell it the plan has already failed completely, and you ask it to explain why. Imagine it’s a year from now and this launch was a disaster. Walk me through exactly what went wrong. The shift in what comes back is immediate and sometimes a little startling. The same model that was warmly supportive a moment ago will now produce a clear-eyed list of the ways your plan could come apart, because you’ve changed the job. You’re no longer asking it to evaluate, which triggers the desire to please. You’re asking it to explain a failure you’ve declared as fact, which triggers something much more useful. It stops protecting your feelings and starts doing the analysis.
This is one instance of a broader skill that’s worth developing, which is the habit of pushing AI toward responses that are useful rather than responses that are pleasant to read. The two are often in tension. A fluent, encouraging, well-structured answer feels good and may contain nothing you didn’t already know. A blunt answer that names the three real risks in your plan feels worse and is worth considerably more. Left to its defaults, the tool will give you the first kind, because the first kind is what it was rewarded for during training. Getting the second kind is your job, and it’s mostly a matter of framing the request so that honesty is the path of least resistance rather than something the model has to overcome its own instincts to deliver.
Prompt for disagreement
A few framings reliably do this. Asking the model to argue against your position rather than for it. Asking it to take the perspective of your most sceptical customer, or your sharpest competitor, or the board member who’s looking for a reason to say no. Asking what it would need to believe for your plan to be wrong. Asking it to rank the weaknesses by how likely each is to actually sink you, which forces it past listing concerns into prioritising them. All of these share the same underlying move. They take the model out of the role of supportive assistant and put it into a role where the useful answer and the agreeable answer are the same thing, so the agreeableness stops getting in the way.
The easy answer
None of this requires any technical knowledge, which is the part worth holding onto. You don’t need to understand how the model works to know that its first answer is probably too kind to be trusted, and that a sharper second answer is usually one good prompt away. The leaders who get the most out of these tools aren’t the ones with the most sophisticated technical grasp. They’re the ones who’ve learned not to take the first pleasant answer as the real one, and who know how to ask the question that gets them the truth instead. The machine has the harder analysis in it. It just won’t volunteer it, and the small skill of drawing it out is one of the highest-return things you can teach yourself, or your team, about working with AI at all.