Checking the result: spotting when the AI lied or missed the point

You've learned to brief a task. What's left is the dullest and most expensive part: working out what came back. Because an AI hands over its result with the same confident face whether everything is right or half of it is invented.

Remember the hallucinations lesson: a model has no built-in "I don't know this" flag. The confident tone is identical for truth and invention. Add last lesson's conclusion — the model is obliged to fill the gap — and you get an uncomfortable truth: how plausible an answer looks tells you nothing at all about it. Smooth text, tidy structure, precise numbers — that's form. Form is the one thing a model gets perfect every time.

Two different failures people confuse

A beginner says "the AI got it wrong". A professional first works out what actually happened, because the fixes are different:

• Failure 1: it lied. The answer solves your task, but the facts in it are false. A number, a quote, a link, a date — invented. Cured by checking.
• Failure 2: it missed the point. Every fact is honest, but it solved a different task. You asked for a teardown, got a recap. You asked for 120 words, got 400. You said don't promise a discount — it promised one. Cured by feedback or by fixing the brief.

The difference is practical. Against lying, only a source check helps — no amount of chatting with the model will do it. Against missing the point, a source is useless — you have to fix the wording. The diagnosis picks the medicine.

What you always check: five red zones

Checking everything is impossible — you'd spend more time than you saved. So here's the short list of things that can't be invented safely. Not in the answer? Relax. In the answer? Check it, every time:

1. Numbers — sums, percentages, deadlines, statistics. The shape of a number is always perfect; the substance may be made up.
2. Quotes — "as X once said". Models are virtuosos at composing convincing quotes nobody ever uttered.
3. Links and sources — risk zone number one from the hallucinations lesson. A plausible URL is the easiest thing of all to invent.
4. Names and dates — especially of lesser-known people and local events: little data in training, plenty of filling-in.
5. Claims about "the law" or "the rules" — where the cost of error peaks, and the model sounds equally confident about a rule in force and one repealed five years ago.

A heuristic for life: check whatever you'd be embarrassed by. If you're going to wave that number around in a meeting or put that quote in a post, then you are the one answerable for it, not the model. Remember the lawyer from the hallucinations lesson who filed six invented precedents? He skipped exactly these five zones.

A pause to think

A tempting idea: if AI makes mistakes, let another AI check it. Think for a second — where's the hole?

The hole is that the second model checks the same way: by predicting what's likely. Ask a model "is all of this correct?" and it will confirm its own invention with equal confidence, because the invention looks likely — that's the whole reason it was generated. Another model will happily "confirm" five non-existent precedents.

There is a useful move here, though — a different one. Not "check it", but "list everything in this answer a human needs to verify, and why". Models are decent at listing risky spots, because that's a task about form, not facts. You do the verifying yourself, in the source. Use the list; never use the verdict.

Here is the answer you just gave. Do not rewrite it.
Task: list everything in it that a human needs to verify: numbers, quotes, links, names, dates, claims about rules.
Format: a table — what to check / where it can be checked / what an error would cost.
Constraints: do not confirm or deny any fact yourself, you are not a source. The list only.

Checking for "missed the point" — match it against the brief

The second failure takes 30 seconds to catch, and you already own the tool: your own brief from the first lesson. Read your five parts and match the answer against them:

• Role — is the tone right?
• Task — did it do what I asked, or recap the topic?
• Context — were my details used or ignored?
• Format — is there a table where I asked for a table? Is the length right?
• Constraints — is the fence intact? Did anything forbidden slip through?

The point where you stumble is the exact address of the fix. That's what separates a person with a brief from a person without one: the first has something to compare against. The second has only a feeling that something's off, and nothing to fix it with.

Feedback: fix on the second pass, don't start over

What does a beginner do when the answer misses? Wipes everything and rewrites the prompt from scratch — losing the good parts and the conversation's context along with them. What does a pro do? Gives feedback in three parts:

1. What to keep — "the first paragraph is great, the tone is right".
2. What's wrong — specifically, not "I don't like it": "the third point is vague, and you promised a discount".
3. What to do instead — "replace the third point with an example using the numbers from my context, and drop the discount promise entirely".

Compare for yourself: "redo it, it's bad" leaves the model guessing what was bad, and it often breaks what was working. The three-part formula locks in what worked and fixes only what didn't. It's the same principle as the brief: leave no gaps for someone else to fill.

Tone and structure are exactly right, keep them.
Wrong: in the second paragraph you promised deadlines I never gave you, and the last paragraph is generic filler with no detail.
Do this: remove any deadlines. Rewrite the last paragraph using only details from my context. Leave everything else untouched.

The rule of three iterations

So when do you start over? There's a simple threshold: three iterations. If after three rounds of fixes the answer still isn't right, the problem isn't the answer — it's the brief, and you're fixing the wrong thing. Go back to the five parts and find the piece that was missing all along. Usually it's context: something you kept in your head and never wrote down.

The tell that you're in this trap: you're fixing the same spot for the fourth time. The model isn't stupid — it's faithfully executing a brief with a hole in it.

A common misconception: "checking takes longer than doing it myself"

Sometimes it does — and that's an honest signal, not a reason to abandon AI. If verifying the answer costs more than doing the work by hand, you picked the wrong task. AI pays off where checking is easy and doing is slow: shortening your text (check = read it), tearing apart your document with word-for-word quotes (check = find the quote), giving you 12 ideas (check = pick one). It doesn't pay off where it's the reverse: assembling market statistics with sources takes a second to generate and half a day to verify.

This is, incidentally, the best filter for AI tasks there is. Before you hand something over, ask yourself: how will I check the result? No quick answer — don't hand it over.

Do this now

Take any answer an AI gave you during this course and walk the five red zones: any numbers? quotes? links? names? claims about rules? Find at least one spot you'd be embarrassed by if it turned out to be invented. Verify it in the source. That habit is what separates a professional from the person in the news story.