Treat the First Draft as a Lying Witness
AI doesn't know things. It predicts the next plausible word. That means it will hand you a quote, a statistic, and a citation with the exact same confident tone whether they're real or invented. There's no flicker of doubt in the output, because there's no doubt mechanism at all.
So flip your default. Assume every specific claim is wrong until you've checked it. The vague stuff ("regular exercise improves mood") is usually safe. The specific stuff — names, dates, numbers, page references, direct quotes, study titles — is where AI fabricates. Those are the parts that make your writing credible, and they're exactly the parts most likely to be fake.
The cost of one fabricated citation in a graded essay or a published post isn't a small ding. It torches your credibility on everything else you wrote. One fake source and a professor starts re-checking your real ones.
The Four Failure Modes
Know what you're hunting for. AI mistakes aren't random; they cluster.
Fabricated quotes. The model invents a sentence and attributes it to a real person. It sounds like something Einstein would say, which is the problem. If you can't find the exact quote in a primary source, it doesn't exist.
Fake citations. This is the famous one. AI generates citations that look perfect — real-sounding journal, plausible authors, a DOI-shaped string — for papers that were never written. Lawyers have been sanctioned for this. Don't join them.
Confident errors. Real entities, wrong facts. Wrong publication year, wrong author, a feature attributed to the wrong tool, a law misstated. Harder to catch because half of it checks out.
Stale knowledge. The model's training has a cutoff. Ask about anything recent — a 2026 pricing change, a new release — and it may confidently describe a world that no longer exists.
Build a Verification Routine
Speed comes from a routine, not from trusting your gut. Run every draft through the same pass.
First, extract the claims. Make the AI surface its own factual load:
List every factual claim in the text below that a reader
could dispute: names, dates, numbers, quotes, and any
cited source. Output as a numbered list. Do not add or
defend anything — just extract.
Now you have a checklist instead of a wall of prose. Go down it.
For each item, the rule is primary source or it didn't happen. A quote needs the original speech, book, or interview. A statistic needs the actual report, not a blog that cites a blog. A study needs to resolve to a real paper you can open. Paste the title into Google Scholar or the journal site. If it returns nothing, the citation is fake — full stop.
Never verify a claim by asking the same AI "are you sure?" It will cheerfully apologize and invent a new fake source to replace the old one. The check has to happen outside the model, in a search engine or a real document. If you want to use AI to help, use a tool with live web access and make it link to sources you then open yourself.
For separating signal from noise online — sketchy stats, manipulated screenshots, "studies" that don't exist — the course at /courses/ai-literacy-spot-misinformation-beginners is a useful companion to this routine.
Make the AI Show Its Confidence
You can pressure-test a draft before you even start checking sources. Force the model to flag its own weak spots:
Review your previous answer. For each factual claim,
rate your confidence as HIGH, MEDIUM, or LOW, and mark
any claim you cannot tie to a specific, verifiable
source. Be honest about what you're unsure of.
This isn't magic — a confident model can be confidently wrong — but it's a triage tool. The LOW and MEDIUM flags tell you where to look first. You'll often watch the model downgrade a citation it stated as fact thirty seconds earlier. That's your tell.
Another sharp move: ask for sources separately from the prose.
Give me three sources that support the claim above.
For each, provide the exact title, author, year, and a
URL or DOI I can open right now.
Then open them. Half the time the links 404, point somewhere irrelevant, or don't exist. That failure rate is the whole point — it shows you how much you'd have shipped on faith.
Know When AI Is the Wrong Tool
Some facts you should never source from a model, period. Anything legal, medical, financial, or safety-critical. Anything tied to a specific recent date. Anything where being wrong has real consequences for a real person. For these, the AI is a starting point for questions, not a source of answers. Go to the statute, the doctor, the official documentation, the primary dataset.
And watch your own laziness. The danger moment is when a claim is convenient — it makes your argument land perfectly, so you want it to be true. That's precisely when you skip the check. Catch yourself there.
The reader who fact-checks ships slower than the one who doesn't, for exactly as long as it takes to get caught once. After that, the careful writer is the only one anyone trusts. Build the routine now, run it on everything, and your name stays attached to things that are actually true.