AI says wrong things in a right-sounding voice. That's the dangerous part.

You trust it. It says it with total certainty. You send it off, then find out it's wrong — and now it's your mistake, not its.

The pain isn't that it was wrong. Everyone is wrong sometimes. The pain is that it was wrong without a flicker of hesitation — no signal at all saying "I'm guessing here." A colleague who isn't sure says "let me double-check that," voice dipping. AI returns a fabricated number in exactly the voice it uses for a number it's certain of. And because the two voices are identical, you have no way to tell them apart — so you trust the wrong one.

01It's equally confident whether right or wrong

You have to understand this to stop blaming it wrongly: AI doesn't know when it's unsure. Humans have a physical feeling — the unease in the gut when saying something you don't have a grip on — and that unease makes us lower our voice, hedge, go check. AI has none of it. To it, every answer is just the most-likely string of words following the question. A correct answer and a plausible-sounding fabrication can be equally likely — and it produces both with the same fluency.

In other words: the confidence in AI's voice has nothing to do with whether it's right. It's a writing habit, not a measure of certainty. Grasping this defuses half the trap — because from now on you stop reading that firm tone as evidence.

The good news is that even though it won't frown for you, a guessed answer usually still leaves traces. Here are four worth stopping to inspect:

Red flags — it may be guessing

Numbers too neat, too roundspecific stats, dates, percentages with no source — real life is rarely that tidy

Names that sound reala book title, a regulation, a feature — sounds perfectly plausible, but you've never heard of it and can't check it on the spot

Not one "it depends" or "may"an answer scrubbed clean of any uncertainty, even where there clearly should be some

On a topic you don't know eithermost dangerous when you lack the expertise to catch the error — this is exactly when to be most suspicious, not least

What the four share: they all "look right." That's why they're dangerous — an error that looks wrong, you'd have caught already.

02Make it reveal its own uncertainty

Since it won't volunteer its weak spots, you have to ask it to. And here's the pleasant surprise: a couple of extra lines in your request turn invisible confidence into something you can see. These two, tacked onto an important question, defuse most of the confidently-wrong cases:

Make it rate its own certainty

"For each main claim, tell me how sure you are — certain / fairly sure / guessing. Where you're guessing, say so plainly."

Permit it to say "I'm not sure"

"If you don't know for certain, say 'I'm not sure' outright instead of guessing to fill space. That's a good answer, not a failure."

Line one makes certainty printable, so you know where to look. Line two removes the must-answer pressure — the thing that pushes it to fabricate in the first place.

These two do something subtle: they change the rules so "I'm not sure" becomes a valid outcome. By default, AI behaves as if every question must have a decisive answer — so when it doesn't know, it still manufactures one that sounds decisive. When you say up front that admitting uncertainty is allowed, you give it a way out of performing. And that scrap of "I'm guessing here" is worth far more than a smooth, hollow answer.

For the genuinely high-stakes spots — the number going into a report, the regulation you're about to cite — don't stop at asking for certainty. Verify one thing. It sounds heavy, but it's surprisingly cheap: checking one number takes thirty seconds, while unwinding a wrong number already sitting in a report sent to your boss takes half a day and a little credibility. The cost isn't in verifying — it's in skipping it.

03A confident voice is the cheapest thing it has

Here's what to carry away: of everything AI produces, confidence is the cheapest. It costs nothing to sound certain, because sounding certain is just a style it learned, not a conclusion it reached after weighing things. The expensive thing — the thing you should demand — is evidence: a source, a self-rated certainty, a checkable number. An answer with a confident voice but no evidence is identical to a wild guess, right up until the moment someone leans on it.

So your job isn't to make AI less confident — you can't, and you don't need to. Your job is to stop reading the confident voice as a signal of correctness, and to always make it reveal its weak spots before you sign your name to it. There's a deliberate method for this: why "done" is just a claim, not evidence, and how to make the agent ask before it guesses. The wrong line that sounds most right is always the one you forgot to ask "are you sure" about.