Field notes·AI·2026

Can You Trust AI? The Five-Second Question That Actually Answers It

Ask a chatbot for anniversary gift ideas and it answers in a confident, helpful voice. Ask it whether two medications mix, and it answers in the exact same voice. That's the whole problem — and the fix isn't in the machine.

For three years the public argument about AI has been stuck on one question: can you trust it? Enthusiasts say yes, it's usually right. Skeptics say no, it makes things up. Both are correct, and both are useless — because the question itself is broken. It's like asking "can you trust cars?" Trust a car to roll downhill? Sure. To drive itself through a school zone in a snowstorm? Different conversation entirely.

Trust was never a property of the machine. It's a property of the task. Swap the broken question for one that answers itself in five seconds: what would a wrong answer cost me here? Ask that, and every AI decision you'll ever face sorts itself onto one of four rungs.

The four rungs

• Rung one: ideas. Brainstorms, drafts, dinner plans, ways to phrase a hard sentence. A wrong answer costs nothing, because a bad idea on a list simply doesn't get picked — you were always the filter. Trust freely, full speed, zero anxiety. This territory is bigger than people think.
• Rung two: everyday facts. How long to boil eggs, what that dashboard light usually means. Wrong costs minor annoyance. Trust, but spot-check before you act — and the more obscure the fact, the harder you check.
• Rung three: money, health, legal, reputation. The appeal deadline, the tax rule, the lease clause, the statistic going in the report with your name on it. Never act on the machine's word alone — verify against the source of record first: the agency's own page, the institution's own site, the actual document. That check usually takes about sixty seconds.
• Rung four: the irreversible. Moving money, taking or skipping medication, signing, filing, accusing. Never delegate. The machine can help you prepare — organize questions for the doctor, draft the letter your lawyer reviews — but the decision belongs to you and, usually, to a qualified human. "Usually right" is not a standard anyone accepts for the irreversible. You wouldn't board a plane that usually lands.

Why the machine can't do this sorting for you

Here's the mechanical truth that makes the ladder necessary: a chatbot works by predicting plausible next words, and it produces them with identical confidence whether they're brilliant or invented. It doesn't hesitate. It doesn't hedge more when it's guessing. Lawyers have been sanctioned in real courts for filing briefs with AI-invented case citations — real-sounding names, real-looking formats, entirely fictional. The machine sounded exactly as sure as when it's right, because it always does. Confidence, from this machine, is a costume it never takes off.

So the judgment has to live on your side of the screen. The trick that makes it a reflex: place the question on the ladder before you read the answer. Before, not after — because once you've read a fluent, organized, persuasive answer, your judgment is already being massaged. Anniversary ideas: rung one, read freely. Mom's medication question: rung four, and now you know before the eloquent paragraph arrives that it's preparation, not permission.

The question "can I trust AI?" has no answer. The question "what would a wrong answer cost me here?" answers itself in five seconds.

One more habit worth stealing: conversations change rungs without announcing it. A friendly rung-one brainstorm can suddenly contain a claim about your insurance coverage — a rung-three fact wearing rung-one clothes. When an answer starts stating facts you'd act on, re-place it. And if the whole subject of installed defaults sounds familiar, it should — your own settings run the same way.