On this pageJump to a section
The scariest AI error is the one that reads fine.
A broken sentence gets caught. A confident paragraph with one wrong number, one invented quote or one claim nobody can support sails through, because it sounds like everything around it. I use AI tools constantly, and the lesson I keep relearning is that the tool does not tell you when it is wrong. You need a process that does. Here is mine, built to take about five minutes and to scale up when the stakes do.
Why do you need a QA step at all?
Because your sense of how well it went is unreliable.
METR ran a randomized trial with 16 experienced open-source developers working on 246 real issues. Before the study they expected AI to speed them up by 24%. Afterward, they believed it had sped them up by 20%. The measured result was that issues took 19% longer with AI allowed.1
19%
Read that result carefully
METR lists strong caveats: a small sample, early-2025 tools, possible selection effects and no claim that AI fails to help in other settings. It also notes its results are now outdated and that it published an update in February 2026.1 I am not citing it to say AI slows people down. I am citing it for one narrow reason: feeling fast is not the same as being right or being faster. That is what a QA step is for.
Failures in the wild look the same
Anthropic's Project Vend (I broke it down in what an AI running a shop teaches about delegation) let an AI agent run an office snack shop for about a month. It gave customers a payment account it had invented, sold items below cost and kept offering discounts after saying it would stop.2 And in the airline chatbot case (see AI Customer Support Without the Air Canada Mistake), a company ended up owing a refund because its bot stated a policy that did not exist.3 None of those errors sounded wrong. That is the point.
What are the five checks?
Run them in this order. Each takes a minute or two.
- Numbers. Trace every figure to a source or record. If you cannot find where it came from, delete it or replace it with something you can verify.
- Names. Check people, companies, products and places. Spellings, titles and relationships are a common quiet failure.
- Claims. Underline every statement that promises, guarantees or asserts a result. Ask whether you can prove it today. If not, cut or soften it.
- Sources. Click every link and search every quote. Make sure the page exists, says what the draft says it says, and carries a date that fits.
- Voice. Read it aloud. Would you say this to this person? Strip filler, hype and anything that sounds like a brochure.
The claims check matters most
For anything that reaches a customer, claims are where legal and reputational risk lives. Models will happily produce a plausible testimonial, a round-number result or a comparison to a competitor. Treat each of those as guilty until traced to a record you hold.
The quiet failures to look for
- The round number. A tidy percentage with no source is often invented to fill a gap.
- The borrowed authority. "Studies show" or "experts agree" with nothing attached.
- The stale fact. True last year, false now, delivered in the present tense.
- The almost-right voice. Correct grammar, but warm where you are blunt, or salesy where you are plain.
- The helpful extra. A guarantee, a deadline or a discount you never offered, added to be friendly.
Teaching yourself to see these makes the five checks faster, because you know where to look first.
Anthropic's own prompt engineers made the same point in a roundtable posted on September 5, 2024. One of them was explaining why reading outputs closely is the core skill of working with a model.4
If you aren't reading the model outputs, you might not even notice that it's making that mistake.
That is the whole case for the five checks. The people who build these models still read the output line by line.
Use the tool to build the checklist, not to pass it
One useful trick: ask the tool to list every factual claim in its draft and name where each came from. That gives you a checklist. But the list itself can be wrong, so you still verify each line against a real source. The tool is good at producing the list. You are the one who closes it.
- AI draftFast, fluent and unchecked
- Five checksNumbers, names, claims, sources, voice
- Risk tierScan, full pass or a second person
- ShipA human decision, every time
- Miss logFeed what the checks caught back into the instructions
How deep should the review go?
Match depth to risk, so you do not skip checks on big things or waste time on small ones.
| Risk level | Review depth |
|---|---|
| Low: internal notes, brainstorms, private drafts | A quick scan for sense and tone |
| Medium: client emails, posts, proposals | The full five-check pass |
| High: claims, prices, health, legal, money, bulk sends | Five checks plus a second person and source documents |
| Irreversible: deleting, publishing, sending to a list | Explicit human approval every time |
Define the tiers in writing
If you do not decide the tiers in advance, a busy afternoon will decide them for you. Write down which tasks fall in which tier, and put the list where you can see it. Your AI policy is the natural home for it.
How do you make QA fast enough to stick?
Make it small, repeatable and recorded.
A five-minute habit
Keep one checklist document. Open it each time. Tick the boxes. A process that lives in your head disappears the first time you are rushed, which is exactly when the errors arrive.
Keep a miss log
Every time the checks catch something, write a one-line note: what the tool got wrong and why. After a few weeks, patterns show up. Maybe it invents figures when it lacks data, or drifts into hype when the brief is thin. Each pattern becomes an instruction for next time, such as "never state a number that is not in the source I paste."
What does this look like in my own work?
I rely on AI agents for a lot of the building in my own business and in the system I build for clients, and the part I never hand off is the review. The agents draft and assemble. I check what reaches a client, and the client approves every claim.
AI can write the first draft in a minute. Whether it ships is still a human decision.
Start tomorrow with one thing: check before you ship. Pick the next AI-assisted piece you plan to send, run the five checks and write down what you find. Chances are you will catch something. That is the process working, not the tool failing.
Keep building.Ekai
Questions people ask
How do I check AI output for errors?
Trace every number and claim to a source, click every link, search every quote, and read the piece aloud for voice. Scale the depth of review to the risk: a quick scan for internal notes, a full checklist for anything a client or customer will read.
Why do people misjudge how much AI helps?
METR's 2025 randomized trial found that experienced open-source developers expected AI to speed them up by 24% and, after the study, believed it had sped them up by 20%, while the measured result was that tasks took 19% longer. The study was small and used early-2025 tools, so it is a caution about self-assessment, not a verdict on AI.
Can AI check its own work?
It can help by listing the claims in a draft and where each came from, which gives you a checklist. But a person must verify the list against real sources, because the tool can be wrong about its own sources.
What should always get a second human review?
Anything that makes a claim about results, health, money or law, anything that spends money or sends to a list, and anything you cannot easily undo.
Sources
- 1
- 2
- 3
- 4
