A draft I was editing recently cited a research finding and attributed it to a specific source. The finding was real. The researchers were real. The source page discussed the study. The problem was that the source page was a marketing blog that mentioned the research in passing. It wasn’t where the finding came from. ChatGPT saw the blog reference the study and decided the blog was the source.
That’s not an obvious fabrication. You’d never catch it by reading the draft. You’d only catch it by fact-checking the AI-generated draft against its cited sources.
Each category of AI fact error presents differently and, whether it’s ChatGPT, Claude or some other LLM, they state these incorrect facts confidently, which makes them even more difficult to spot. A single-pass read won’t catch them. This post names all five and gives you a process for finding them.
At a Glance
- Wrong sources. AI credits real findings to wrong sources or fabricates plausible-sounding citations.
- Mashed-up stories. When AI merges details from separate real events into a single narrative that never happened.
- Made-up numbers. Generating specific numbers that don’t exist in any source to manufacture credibility.
- Outdated data. AI treats every fact as equally stable, so counts, versions, and dates become outdated without warning.
- Missing fine print. AI drops qualifiers and limitations, presenting the most impressive version of every feature. (These surface-level tells overlap with the patterns covered in how to humanize AI writing for nonfiction.)
- Systematic process required. Every number, named source, and composite narrative needs independent verification against the original — a category-by-category process catches what a single read-through won’t.
One bad citation and your reader questions the whole book
If you’re publishing nonfiction, providing correct information is arguably more important than anything else. Readers can get past a writing style they don’t like. But when a reader catches a factual error in your work, they don’t just question that claim. They question every claim you made.
For those of us using AI as part of the writing process, fact-checking is a core piece of meeting AI book quality standards — and generic “check your facts” advice doesn’t help because it doesn’t tell us what we’re checking for.
When a reader catches one bad citation, they don’t question that citation. They question every claim you made.
The misattributed citation problem
In the manuscripts I’ve edited, wrong-source errors are the most common AI failure type. Your draft cites a study, you search the source for the exact finding, and it isn’t there. It happens in two patterns:
- Misattributed sources. ChatGPT credits a finding to the wrong source. It might point to a blog that mentioned a study instead of the original research, or flatten a citation chain so that paper A gets credit for data that actually came from paper C. The citation looks right. It points to the wrong thing.
- Fabricated sources. ChatGPT invents references that sound real. In one manuscript, the draft quoted a passage supposedly from a published industry report. The report was real. The passage wasn’t. ChatGPT generated a plausible-sounding quote and presented it as fact.
For every attributed claim in your draft, search the cited source for the exact quote or finding. If it’s not there, the attribution is fabricated or misplaced. There is no shortcut for this step.
Two real stories, one fake narrative
Sometimes each fact in a passage checks out individually, but something about the story still feels wrong. The dates don’t quite line up. The numbers seem too perfect together.
That’s a mashed-up story. The AI merges details from two or more separate real events into a single narrative that never happened. Each individual detail traces to reality. The composite is fiction.
Here’s what it looks like in practice. I found a manuscript where the opening paragraph implied that more than 18,000 people’s data had leaked through 16 security vulnerabilities as part of Matt Palmer’s May 2025 research. In reality, there were two completely separate incidents. Palmer’s work in May 2025 found 170 apps with a specific misconfiguration and a CVSS score of 8.26.
A different researcher in February 2026 found a single app with 16 vulnerabilities and 18,697 user records. ChatGPT combined the dates, actors, and metrics from both incidents because they involved the same platform.
This is the hardest error type to catch because it passes a surface-level fact check. It sounds convincing. You can verify each number independently and they’ll all check out. But the narrative connecting them is fabricated.
The detection method: when a passage combines a named person, a specific date, and multiple quantitative metrics, verify that all details come from the same event. If any element traces to a different incident, the composite is fiction.
The more vivid and precise an AI-generated anecdote sounds, the more likely it’s fabricated. Real case studies have messy details.
Numbers can look real even when they aren’t
You’ll find numbers in your AI draft that look like they came from real research, but some were never real to begin with, and others have gone stale since you last checked.
Made-up numbers
Phantom statistics. ChatGPT generates suspiciously specific numbers that look researched but don’t exist in any source. I found a draft citing “562 upvotes and 252 comments” for a Reddit thread that couldn’t be independently verified. In another section, a security score was stated as 9.3 when the actual score from the source was different.
Specific numbers are how ChatGPT manufactures credibility.
The detection rule: any number in your draft that wasn’t in your original research is a red flag. Search the cited source for the exact figure. If it’s not there, cut it or find the real number. Don’t try to read the whole article to find it either. Do a Control+F on a phrase or distinct word from the statistic to try to find where the factual claim lives in the source document.
Outdated data
Stale numbers pass the first check. These numbers were accurate when you researched them. They’ve changed since. A book about social media marketing quoted a platform’s user count from the research phase. By the time the manuscript reached factcheck, the real number was 60% higher.
Another draft described a company’s product lineup that had been reorganized months before publication. One draft still called the platform “Twitter” two years after the rebrand to X.
Some of this is unavoidable in nonfiction, especially in certain niches like technology. The best you can do is make sure that factual information is correct at the time you publish.
The detection rule for stale data: any count, version number, dated event, or model lineup requires re-verification with a timestamp.
The five-step fact-check process for AI manuscripts
Knowing the error types matters, but you need a repeatable process you can run on every draft. These five steps are ordered by what catches the most errors fastest. If you run all five and the draft survives, the remaining errors are minimal.
- Find every claim that names a source. Highlight every sentence that attributes a finding, quote, or statistic to a named source. These are your highest-risk items. In my experience, wrong-source errors are the most frequent AI failure type, so starting here gives you the highest return on your checking time. You can use AI to speed up the process of finding what information needs to be cited, but a human should do the final fact-checking.
- Check the numbers against the original. For every number in the draft, search the cited source for the exact figure. Statistics, dates, counts, scores. If the number doesn’t appear in the source, it’s either fabricated or transplanted from a different context. Cut it or find the real number.
- Make sure each story is actually one story. When a passage combines a named person, a specific date, and multiple quantitative details, verify that all details come from the same event. AI merges separate incidents that involve the same platform, topic, or person.
- Confirm the data isn’t outdated. Any count, version number, or dated event requires a freshness check. Search for the current figure. Note the verification date in your manuscript records so you know when each number was last confirmed.
- Look for missing disclaimers and caveats. For every feature claim, product description, or capability statement, check the qualifiers. Is this paid or free? Which plan tier? Is opt-in required? Experimental, beta, or generally available? OS-specific? AI gravitates toward the most impressive version of every feature. I found cases where features were described as universally available when it required opting in on a specific paid plan, and where a desktop application was described as not requiring certain software when installation was actually mandatory.
What to do next
You have two paths from here. You can run the five-step process yourself on your existing draft. It takes time, but it works, and the order matters.
Or you can work with a production team that builds fact-checking into every manuscript. A multi-stage production process catches these errors at each phase instead of hoping a single editing pass finds them all. If you’d rather hand off the verification work, that’s what a manuscript production service handles.
Either way, the error types don’t change. The question is whether you want to run the process or have someone run it for you.
Sources
- Originality.ai. “AI Fact Checking Accuracy Study.” Source for the ~87% automated fact-checking accuracy figure.
FAQ
Can AI fact-checking tools replace human verification for nonfiction books?
AI fact-checking tools are useful as a first pass, but they can’t replace human verification for book publication. Originality.ai reports that GPT-5 achieves roughly 87% accuracy on its automated checks. That means a 13% miss rate. For a 50,000-word nonfiction book with hundreds of claims, 13% is dozens of potential errors reaching print. Use automated tools to flag candidates for review, then verify every flagged claim against the original source yourself.
What are the most common AI errors in nonfiction manuscripts?
The five main types are wrong sources (crediting findings to the wrong researcher), mashed-up stories (merging separate events into one narrative), made-up numbers (generating fake statistics), outdated data (using stale figures), and missing fine print (omitting feature limitations). Wrong-source errors are the most frequent. Each type hides differently in your manuscript.
How do you check if AI made up a source or citation?
Search the cited source for the exact quote or finding. If it’s not there, the attribution is fabricated. AI often credits secondary blogs instead of original research, or invents plausible-sounding references that don’t exist in any source. The confidence of the language is not a signal of accuracy. The most authoritative-sounding citations are often the most fabricated.
Why does AI writing fail differently for nonfiction than fiction?
Fiction AI failures show up as flat characters and plot inconsistency. Nonfiction AI failures destroy credibility through fabricated citations, misattributed research, and composite events presented as real incidents. The damage hits the reader’s trust in the author’s expertise. A reader who catches one fabricated statistic stops trusting the rest.
How often does AI fabricate case studies in nonfiction books?
Frequently enough that any case study with a named person, a named project, and three or more specific quantitative metrics should be independently verified. The more vivid and precise an AI-generated anecdote sounds, the more likely it’s fabricated. Real case studies have messy details. AI-generated ones have numbers that illustrate the teaching point too perfectly.
What should I fact-check first in an AI manuscript?
Start with attributed claims. Look for every sentence that credits a finding, quote, or statistic to a named source. These are the highest-risk items because wrong-source errors are the most common AI failure type in nonfiction. After source claims, move to numbers, then composite stories, then stale data, then missing caveats. That order catches the most errors in the least time.