Checking the Output

AI is a confident writer. It writes well-formed, reasonable-sounding sentences even when it's completely wrong. Part of using AI well is knowing how to double-check.

The word professionals use for "AI made something up" is hallucination. It looks like a real fact. It sounds right. It's not. Hallucinations happen most often with names, numbers, dates, quotes, and links, exactly the things that are embarrassing if you pass them along.

1. Grounded answers vs. remembered answers

This is the most important idea on the page, and it changes how you read every reply.

An AI can answer you in two very different ways, and they are not equally trustworthy:

• From memory (recall). The answer is dredged up from the hazy, averaged patterns the AI soaked up in training. This is where generic answers and confident made-up facts come from. It has no specific source, it's a best guess shaped like a fact.
• From a document you gave it (grounding). The answer is pulled from a file you uploaded, a page it just searched, or text you pasted in. There's a real source sitting right there, and you can go read it.

The single most useful habit you can build is to always know which one you're getting. When it matters, make the AI tell you. The cheapest possible check is one sentence:

Quote the exact sentence you got that from.

If it can quote a real sentence from a real source, the claim is grounded, and you can verify it in seconds. If it starts hedging, paraphrasing, or can't point to anything, the answer was recall, and you should treat it as unverified until you check it yourself.

2. Citations: what a sourced answer feels like

The clearest example of grounding is NotebookLM, which we met on the Projects page. It answers only from the sources you gave it, and it attaches a little numbered marker to nearly every sentence. Click a marker and it jumps you straight to the exact line in the exact document the claim came from.

Once you've used it, ordinary chat feels different by comparison. In a normal chat, a fact just appears and you have to wonder where it came from. With citations, every claim wears its receipt. You're not trusting the AI's memory, you're reading your own document, with the AI acting as a fast index into it. That is what "sourced" feels like, and it's why, when accuracy really matters, numbers, quotes, who said what, a tool that cites beats one that just talks.

3. Push back on confident-but-generic answers

A lot of AI output sounds intelligent but could apply to anyone. If the advice doesn't mention your specific situation, your actual details, your real constraints, it's filler.

Push back:

That's too generic. Tie each point to something specific in the files I gave you. If you don't have enough info, tell me what's missing.

The AI will usually rewrite with far more useful specificity. Sometimes it'll say "I don't have info on X," which is great news. It's being honest.

4. Make it be short

AI is biased toward long answers. Length often disguises weak thinking: six paragraphs of polite hedging that amount to nothing. Force brevity, especially for analysis:

Give me the three most important points. No more. No preamble.

If the three-point version is weak, the six-paragraph version was weak too. It was just easier to hide.

5. Swap the order when comparing

This is a small but striking trick. AI has a habit of favoring whichever option you list first. If you're asking it to compare two things, "should I go with A or B?", ask the same question a second time with the order reversed.

Which laptop is the better buy for a student on a budget: the A14 or the B12?

(Later, in a fresh chat:)

Which laptop is the better buy for a student on a budget: the B12 or the A14?

If the answer flips based on order alone, the AI doesn't actually have a strong reason. Treat the "answer" as a starting point for your own thinking, not as a verdict.

Where models are reliably weak

Hallucinations aren't random. There are a few specific places the AI is predictably shaky. Learn the list and you'll know exactly when to slow down and check.

• Exact counts and arithmetic. The AI reads text in chunks, not digit by digit, so counting and math sit right over a blind spot: it can botch grade-school arithmetic in the middle of an otherwise brilliant reply. The fix is the one from the Basics page: tell it to "use code," so it computes the exact figure instead of guessing.
• Anything after its training cutoff. The AI's built-in knowledge is frozen at the date it was trained. It doesn't know this week's news, today's prices, or a place's current hours, unless it actually searched the web. For anything recent or live, make it search rather than recall.
• The middle of very long documents. Models attend best to the start and end of a long input and skim the middle. Ask about page 40 of an 80-page PDF and the answer may be vaguer than it looks. When it matters, ask it to quote the exact passage, or point it at the specific section.
• Confident citations to things that don't exist. Papers, court cases, books, URLs: the AI can produce a perfectly formatted reference to a source that was never written. A real-looking author, a real-looking year, a dead link. Always click the link and search for the title before you rely on it.

The Swiss cheese problem

AI ability is like Swiss cheese: mostly solid, with random holes. It can write you an excellent plan and, in the very same reply, botch a piece of grade-school arithmetic. These models have stumbled on questions as simple as "which is bigger, 9.11 or 9.9?" or "how many R's are in strawberry?"

The dangerous part is the lesson people draw: they see the AI ace a hard task and assume it must be right about the easy one too. It doesn't work that way. Being brilliant at one thing tells you nothing about whether it's correct on the next thing.

Common mistakes to watch for

• Fake URLs. AI sometimes invents website addresses that look real but go nowhere. Always click before you rely on one.
• Wrong numbers. If the AI quotes a price, a statistic, or a measurement, assume it could be wrong until you've checked.
• Made-up quotes or studies. "A 2023 study found...": does that study actually exist? Often it doesn't. Ask for the title and author and search for it.
• Overconfident medical, legal, or financial advice. AI will give it freely. Don't act on it as-is. Treat it as a starting point, then talk to a real professional.
• Old information. Models have a knowledge cutoff date and don't know recent events unless connected to the web. For news or anything current, use a tool with web search turned on, not the model's memory alone.

The "so what?" test

When the AI gives you a summary or a list of findings, ask it one last question:

So what? Give me the three specific things I should do based on this.

A good answer holds up here. A weak answer means the summary was fluff, either the grounding was too thin, or the AI was coasting. Either way, you've caught it before anything went out the door.

Cite what you opened

Everything above works in a chat window. Once I started letting AI agents do research and write notes for me, I needed a stricter version of the same idea, and it's now a rule every agent on my computer works under: no claim without a pointer you can check. That pointer can be a file and line number, a link the agent actually opened, or a command plus what it printed. If the agent didn't open the source in this session, it doesn't get to cite it. A link sitting in an answer doesn't prove anyone read the page.

The second half of the rule is to keep three things apart. First, what the source says: the exact quote and where it lives. Second, what the AI concludes from it, marked as its own reading. Third, what you decided to do about it. The failure this prevents is subtle. A conclusion gets written down where a quote should be, and a week later nobody can tell which parts came from the document and which were the AI's guess.

When there's no evidence, the claim stays in and gets the word unverified next to it. It doesn't get quietly dropped, and it doesn't get dressed up to sound sourced. When two sources disagree, both stay in the answer, with a note that they conflict. You can ask for all of this in plain words:

For every factual claim, give the exact quote and where it came from: the file and line, or the link you actually opened. Mark your own conclusions as yours. If you can't point to a source, write "unverified" next to the claim. If two sources disagree, show both and say so.

Records are append-only. If a saved note turns out to be wrong, I don't edit it in place. I write a new note that points at the old one and says what replaced it, so the history of what we believed and when stays readable.

The writer doesn't check their own work

Whoever wrote something is badly placed to verify it, and I treat AI agents the same way. For anything that matters, a second reader checks it before it counts as done: a different agent, ideally a model from another company, or me. And "done" means a check actually ran and passed, like a test, a command, or a page loaded and looked at. It doesn't mean the agent says it thinks it works.

One last pass before it goes public

Anything that leaves my computer with my name on it (a commit message, a pull request, an issue, a comment) gets one more pass through a "humanizer" prompt before it goes out. The prompt is a checklist of the habits that make text read as machine-written: puffed-up words, filler phrases, lists of three where two would do, a dash in every other sentence, an upbeat line tacked on at the end. It rewrites wording only. It's told outright not to add any fact, name or number that wasn't already there, and a rewrite that sneaks one in counts as a failure even if it sounds better.