The ten-minute habit that keeps AI from embarrassing you.

Every AI embarrassment you've read about has the same anatomy. The output was fluent. It was specific. It sounded like something a competent person would write. And it was wrong in a way that took a reader with actual knowledge about a minute to spot. The person who shipped it didn't lack intelligence or tools. They skimmed, because the output looked finished, and looking finished is the one thing AI does flawlessly. There's a structural reason. OpenAI's own researchers concluded in 2025 that models hallucinate because training and evaluation reward a confident guess over an admission of uncertainty. Confident wrongness is baked into the grading.

Lawyers learned this the expensive way. Damien Charlotin's AI Hallucination Cases database, which tracks court decisions where a party relied on fabricated citations or other invented material, reached 1,600 cases worldwide in June 2026. The profession's working standard has hardened accordingly, from 'trust but verify' to what Thomson Reuters now calls 'do not trust until verified,' and we covered why verification is now a survival skill in legal work. But the pattern isn't a legal problem. It's a delegation problem. The moment you hand work to anything, a junior employee, a contractor, a model, your job shifts from producing to reviewing. People who skip the review aren't moving faster. They're accumulating risk with their name on it.

The good news: knowing how to verify AI output is not a talent. It's a short checklist, and the right check depends on the kind of output you're holding. Run it and you get AI's speed with your reputation intact. Skip it and you're trusting a system built to sound right, not to be right.

The strongest case against the habit comes with real data. In a 2025 randomized trial by METR, experienced open-source developers using AI tools took 19% longer to finish tasks than developers working without them. They had expected AI to make them 24% faster, and even after the study they believed it had made them 20% faster. Review and rework ate the gains, and nobody felt it happening. If checking can quietly erase the speedup, why build a habit around more of it?

Read that result again, though, and it argues for the checklist. Those developers' instincts were off by nearly forty points. When you can be measurably slower while feeling faster, 'it feels fine' tells you nothing about your own workflow, so you verify AI output with a named check instead of a vibe. The overhead also isn't fixed. When METR re-ran the experiment with newer tools starting in late 2025, the slowdown disappeared from their estimates, though the researchers caution that selection effects make the new numbers weak evidence. Checking speeds up with practice, because you learn where models fail and look there first.

The other half of the answer is asymmetry. Verification costs minutes, and the cost is visible. Skipping it costs nothing until it costs a sanction, a retraction, or a client who catches the invented number, and those 1,600 court cases are just the failures public enough to count. Proportional checking spends the minutes where a mistake would be expensive and spends nothing where it wouldn't.

Not every output earns the full treatment, the part most advice on how to verify AI output gets wrong. A brainstorm you'll discard by lunch needs no verification at all. A draft email to a colleague gets the read-aloud pass. A number going into a client deliverable or a board deck gets the full trace, every figure back to a source you opened. Accountants are living this calibration right now as AI absorbs the first-pass work and they own the review. The question is never 'did I check everything?' It's 'did I check in proportion to what this mistake would cost?'

Second principle: verification means running a named checklist; a vibe doesn't count. 'It looks right' is exactly the instinct AI output is optimized to trigger, so your gut is the one tool you can't use here. There's research behind that warning. When Microsoft and Carnegie Mellon surveyed 319 knowledge workers in 2025, the more confidence people had in the AI, the less critical thinking they applied to its output; confidence in their own expertise pushed the other way. The checks above work because each one forces contact with ground truth: a source you opened, a result you already knew, a section you can recite. If you can't name which check you ran, you didn't verify, you skimmed with extra confidence.

This is the meta-skill of the AI era, and it earns back far more than the minutes it costs. Verification is what makes delegation safe enough to scale. The people getting the most out of AI aren't the ones with the cleverest prompts. They're the ones who can hand off far more work because they have a fast, reliable way to catch failures before anyone else sees them. It's the same reason experienced reviewers are suddenly so valuable: judging output you didn't produce is the skill that survives every model upgrade. It sits at the core of the skills we teach, and it's learnable on your real work instead of in theory.