Most professionals who use AI tools daily have never formally described how they decide whether to trust what the tool hands back. Dr. Jonah Tebaa has, and the habit he describes is notably unglamorous: a printed page, a pen, and a specific question asked in a specific order, three times, before he lets a draft near a client.
The scene he uses to explain it says as much as the method itself. Standing at his desk shortly before a client meeting, memo in hand, he is not proofreading. He is looking for the paragraph most likely to draw a challenge the moment the meeting starts. In his account, that single habit, reading a document standing up while hunting for its weakest claim, is close to the entire discipline compressed into one motion.
A Habit Built Around Three Separate Questions
What distinguishes Dr. Tebaa's approach from the more common instinct to simply "check if it sounds right" is that he treats fluency and correctness as unrelated properties. A draft can be well written and still answer the wrong question, and it can answer the right question while still containing a claim that would not survive scrutiny in the room where it matters.
His first read ignores style altogether and checks only whether the draft engaged with the actual question that was asked, not a nearby, easier version of it. He has observed that AI systems tend to quietly narrow a hard question into a more general one, and that a confident, well-organized answer to the wrong question is often more dangerous than an obviously weak one, because the polish disguises the gap.
The reflex his first pass is built to interrupt is documented in regulator guidance rather than folklore: the UK Information Commissioner's Office, explaining automation bias, notes that because AI models rest on mathematics and data, people tend to think of them as objective and trust their output regardless of how statistically accurate it is. Dr. Tebaa's three passes are one practitioner's answer to that tendency, run at a desk rather than written into a compliance policy.
The Adversarial Middle Step Most People Skip
The part of Dr. Tebaa's method that sets it apart from the general advice circulating about AI oversight is the second pass, which he treats as a deliberately hostile read rather than a review. Before anyone, including himself, is permitted to praise a draft, someone has to attack it: a colleague briefed to look only for weaknesses, or Dr. Tebaa himself, consciously switching from author to opponent.
He argues that most people stop one step too early, treating an AI draft as finished once it reads plausibly, and that this is precisely where errors travel furthest, because a plausible-sounding claim is the hardest kind to catch later. The questions he runs at this stage are practical rather than theoretical: which number in the draft looks like a confident guess rather than a verified figure, where a skeptical client is likely to push back first, and what the draft may have quietly left out that would change the conclusion if it were put back in.
Only after a draft has been actively attacked, in his framing, does it earn the right to be judged on its merits.
Judging Output by What It Changes, Not How It Reads
The final stage of Dr. Tebaa's method departs from writing quality entirely. He evaluates an AI draft by asking whether it altered a decision he was about to make. A document that is accurate, well argued, and confirms what he already believed, in his view, has not done useful work, regardless of how polished it is. A shorter, rougher draft that changes a single number he plans to challenge or a recommendation he decides to drop has done its job.
This closing test functions as a check on the two passes that precede it, since it is easy to treat an intent check or an adversarial read as procedural steps performed for their own sake. By insisting that a draft prove its value in terms of a real decision, Dr. Tebaa's framework stays oriented toward judgment rather than documentation, a distinction he returns to often in his broader commentary on how practitioners, not organizations, ought to work alongside AI tools day to day.
Related evidence: The Chain-of-Verification method has a model draft an initial response, plan verification questions to fact-check that draft, answer them independently, and only then produce a final verified response, which reduces hallucination. (the Chain-of-Verification research on checking a draft)
