The dangerous thing about AI errors isn't that they happen. It's that they look exactly like the correct parts.
There's no hedging, no tell, no change in tone. A fabricated statistic arrives with the same confident phrasing as a real one, attached to a plausible-sounding source, in a sentence that reads perfectly. You can't catch it by reading carefully, because careful reading is what it's designed to survive.
So you don't catch it by reading. You catch it by checking specific categories of claim, systematically, in a dedicated pass.
The short version
Where AI reliably invents:
- Statistics — precise-sounding numbers with no source
- Citations — real-looking references to studies that don't exist
- Quotes — plausible statements attributed to real people
- Dates — when things launched, changed, or were published
- Specifics of how things work — features, settings, menu paths
- URLs — links that 404
The process: one dedicated pass, checking only these categories, before you read for anything else.
Why this happens
Worth understanding, because it tells you where to look.
A language model predicts likely continuations. Asked for a statistic about email open rates, the most probable continuation is a plausible-looking percentage attached to a plausible-looking source — because that's the shape such sentences take in its training data.
It isn't lying, and it isn't malfunctioning. It's producing the statistically likely form of the sentence you asked for. The number is fictional in the same way a plausible-sounding name in a novel is fictional.
Which means: the more specific and authoritative a claim sounds, the more worth checking it is. Precision is not evidence of accuracy. "Studies show that 73.4% of…" is exactly the shape of both a real finding and an invented one.
And web search doesn't fully solve it. A model with search can retrieve real sources — but it can also misattribute a real figure to the wrong source, quote a number out of context, or blend a retrieved fact with a remembered one. Retrieval reduces the problem; it doesn't remove the need to check.
The six categories
1 — Statistics
The tell: a specific percentage with no linked source, or attributed vaguely to "research" or "studies."
How to check: search the exact figure plus a distinctive keyword. If you can't find it at a primary source in two minutes, cut it.
The trap: finding the figure repeated on three blogs isn't verification. Fabricated statistics propagate — one article invents it, others quote it, and now it appears authoritative through repetition alone. Trace it to the organisation that produced it.
2 — Citations
The tell: references to studies, papers, or reports that sound exactly right for the argument.
How to check: search the title. Then search the author. A fabricated citation usually has a real-sounding author, a plausible journal, and no existence.
Worth knowing: this is the category with the highest embarrassment cost. A wrong percentage is a correction; a cited study that doesn't exist is a credibility problem, and it's the specific failure that has ended careers in law and academia.
3 — Quotes
The tell: a quotation from a named person that neatly supports your point.
How to check: search the exact phrase in quotation marks. If it exists, you'll find it. If you find only paraphrases, it's probably reconstructed.
The rule: never publish a quote you haven't seen at its source. Attributing invented words to a real person is worse than every other error here.
4 — Dates
The tell: any claim about when something launched, changed, or was deprecated.
How to check: the vendor's own announcement or changelog.
Why it matters more than it sounds: dates anchor claims about what's current. "Google removed this feature in 2023" is either useful context or an actively misleading statement, depending entirely on whether it's right — and readers use dates to judge whether your piece is worth trusting.
5 — How things actually work
The tell: specific menu paths, setting names, feature descriptions, pricing tiers.
How to check: open the product and look.
The most common version of this error: a menu path that was correct two versions ago. The model learned it from documentation written in 2023, the vendor moved the setting in 2025, and your instructions send readers somewhere that no longer exists. Deeply frustrating for a reader following along, and it makes everything else you wrote look unreliable.
6 — URLs
The tell: any link you didn't paste in yourself.
How to check: click every one.
Models produce plausible URL structures for real sites. example.com/blog/the-thing-youre-writing-about is exactly the shape a real URL takes, and frequently a 404.
The process
Do this as a dedicated pass, before reading for flow or voice.
1. Highlight every checkable claim. Numbers, names, dates, citations, quotes, menu paths, links. Don't evaluate yet — just mark them. Most articles have between fifteen and forty.
2. Work through the list, not the article. Reading the article to check facts means your attention keeps sliding to the prose. Working through a list of claims keeps you in checking mode.
3. For each: verify, replace, or cut.
Verify against a primary source you can see. Not a blog quoting it — the organisation that produced it.
Replace with a figure you can verify, if the point stands without the specific claim.
Cut if you can't verify it. This is the discipline that matters. A piece with fewer specifics is better than a piece with one invented statistic — one fabricated figure discovered by a reader costs more than the whole article earned.
4. Add the sources you used. Both because it's honest and because named sources with dates are what makes a passage citable rather than replaceable.
Two habits that reduce the work
Constrain at the brief stage. Give the model your sources and instruct it explicitly: use only what's provided, and flag any claim that would need a statistic I haven't supplied. Models genuinely will flag gaps when told that's preferred to filling them — and a flagged gap takes thirty seconds to resolve while a fabricated one takes a reader to discover. The brief.
Ask for sourcing inline as it drafts. Requiring a source next to every factual claim makes unsourced claims visible immediately rather than requiring you to hunt them down afterwards.
Neither eliminates the checking pass. Both reduce it substantially.
A note on your own claims
The checking pass applies to everything, including what you wrote yourself.
Your own remembered statistics deserve the same treatment. A figure you've quoted for three years may have been superseded, or may have been wrong when you learned it. Half-remembered numbers are their own category of error, and they're easier to miss because you trust the source.
And your own data needs stating carefully. "In 50 audits I ran last year, 34 had no valid SPF record" is checkable and honest. "Most companies have authentication problems" is neither — and the first version is also far more citable.
Frequently asked questions
What does AI most often get wrong?
Statistics, citations, quotes, dates, and the specifics of how products work — anything requiring a precise fact rather than general understanding. What makes these dangerous is that fabricated material arrives with exactly the same confidence and phrasing as accurate material, so you can't catch it by reading carefully. The most common practical error is a menu path or setting name that was correct two versions ago and has since moved.
Can AI cite real sources?
Sometimes, and not reliably. Without web search, citations are generated from patterns in training data and may reference studies that don't exist while having entirely plausible authors and journals. With search enabled, a model can retrieve genuine sources — but it can still misattribute a real figure to the wrong source or quote a number out of context. Always open the citation and confirm it says what the draft claims it says.
How do I verify a statistic quickly?
Search the exact figure alongside one distinctive keyword from the claim, and trace it to the organisation that produced it rather than a blog repeating it. Finding the number on three websites isn't verification — fabricated statistics propagate exactly that way, gaining apparent authority through repetition. If you can't reach a primary source in about two minutes, cut the claim; the piece is stronger without it.
What to do next
Take the last AI-assisted piece you published and highlight every number, date, and named source in it.
Then check three of them properly — trace them to a primary source, not a blog repeating them.
Most people find at least one thing that doesn't hold up. Better to find it yourself than to have a reader find it, and it calibrates how much checking your process actually needs.
Free: The content brief template.
Related guides
- AI content production without the slop — where this pass fits
- The content brief — reducing the work upstream
- Human-in-the-loop — when to check versus sample
- Entity SEO — why verified specifics make content citable
Join the Newsletter
Get practical marketing tactics delivered straight to your inbox.
Written by
Muhammad Basim
Related Articles
What Marketing Automation Actually Costs at Scale
Nobody's automation bill jumps because they added more automations. It jumps because they added steps to workflows that were already running — a filter here, an enrichment lookup there, a second notification — and the billing model charges for every one of them, on every run. That's the mechanic behind almost every "why did this […]
One Workflow, Three Tools: A Build Comparison
Feature tables don't tell you what a tool is like to use. The only thing that does is building the same thing twice and noticing where you got annoyed. So here's one realistic workflow — lead routing with enrichment and conditional assignment — specified once and mapped across Zapier, Make, and n8n. What changes between […]
GoHighLevel: 7 Automation Workflows to Build First
GoHighLevel gives you a hundred things you could automate, which is exactly why most accounts end up with forty half-built workflows and no measurable result. Seven are worth building first. They're the ones tied directly to revenue — catching leads before they cool, recovering appointments that would otherwise vanish, and asking for reviews at the […]