Know what AI is actually good at

The skill isn't trusting AI or distrusting it. It's knowing, fast, which kind of question you're holding.

Two failures sit at opposite ends of the same mistake. The over-truster takes a confident answer at face value and ships it. The reflexive sceptic refuses to use a genuinely capable tool and quietly pays for it in time and quality. Both have skipped the only question that matters: is this the sort of thing AI is reliable at?

That question has reasonably knowable answers. Models are strong at fluent transformation — rephrasing, summarising, translating the shape of an idea, generating options, explaining a well-trodden concept. They're weak, and confidently so, at anything where being precisely right matters and the truth lives outside their training: live facts, specific figures, recent events, niche specialisms, and the consequences of high-stakes decisions about your health, money, or legal position.

A two-part test

Before you lean on an answer, run it through two quick filters. First, stakes: if this is wrong, what does it cost me? Low-stakes work — a casual draft, a first pass, a list of options — barely needs checking, and over-verifying it is its own miscalibration. Second, verifiability against the world: does the claim depend on a specific external fact I could check, or is it a transformation of things I already gave it? The first kind needs a primary source or a human; the second usually doesn't.

High stakes or fast-moving facts go to a human or a primary source. Everything else, let the tool carry — and stop second-guessing the parts that don't matter.

There's a quieter case worth naming: when the model contradicts you in your own area of expertise. The reflex to defer — "it's read more than I have" — is exactly backwards. Your hard-won experience is a signal worth more than its fluency. Don't simply accept the contradiction, and don't dismiss it either; treat it as a flag to investigate the gap.