An AI tool once handed us a line claiming that "73% of consumers trust brands that use authentic language." It had a percentage, a confident tone, and would have looked great on a landing page. It also doesn't exist in any study we could track down. That's the problem with AI-generated marketing copy: it doesn't just write well, it writes convincingly, even when it's making things up.
If you're using ChatGPT, Claude, or any other AI tool to draft blog posts, ads, or email copy, you need a verification pass before publishing. Not a quick skim. An actual process for catching the specific things AI tools tend to invent.
Why AI Tools Invent Facts
Large language models predict what text should come next based on patterns, not a database of verified facts. When you ask for "a statistic about email marketing," the model generates something that sounds like a statistic because that's the pattern it learned, not because it pulled a number from a real report.
This is often called "hallucination," but that word makes it sound rare or glitchy. In our experience, it's closer to the model's default behavior when it doesn't have a real answer and doesn't say so. It fills the gap with something plausible instead of leaving it blank.
The fix isn't avoiding AI tools. It's knowing exactly which parts of the output need a human check before anything goes live.
The Four Things AI Loves to Fabricate
After running AI drafts through fact-checks, four categories cause the most trouble.
Statistics and percentages
Any number with a percent sign, a dollar amount, or a "studies show" attached to it is a candidate for invention. AI models are especially good at generating numbers that feel realistic (not round, not too extreme) which makes them harder to spot as fake.
Quotes attributed to real people
Ask an AI tool for "a quote from a marketing expert about customer retention" and it may return a full sentence attributed to a real, named person who never said it. The quote might even sound like something that person would say, which makes it more dangerous to publish, not less.
Named studies, sources, or organizations
Phrases like "according to a Harvard study" or "research from the Content Marketing Institute found" show up often in AI drafts. Sometimes the organization is real but the specific finding is not. Sometimes both are invented.
Specific product or feature claims
If you asked an AI tool to write about your own product and it added a feature, integration, or guarantee you don't actually offer, that's not a factual error in the traditional sense. It's a liability. Legal and customer service will both notice.
A Five-Step Verification Pass Before You Publish
Here is the process we use on any AI-drafted copy before it goes out the door.
Isolate every factual claim. Read the draft once just to underline or highlight anything that states a fact: a number, a date, a quote, a named source, a specific product detail. Ignore style and tone on this pass. Just hunt for claims.
Search for the exact number or quote. Copy the specific statistic or quote and search for it directly, in quotation marks. If nothing matches, or if the only matches are other AI-generated pages repeating the same unsourced claim, treat it as fabricated.
Check that the source actually says what's claimed. If the AI cites a real organization or study, go find the original report or article. AI tools sometimes attach a real source to a made-up number, or slightly exaggerate a real finding. "Nearly half" in the original study can turn into "52%" in the AI draft.
Verify quotes against their original context. If a quote is attributed to a real person, search for their actual public statements on the topic. If you can't find the exact quote anywhere, or the person has never publicly commented on that subject, remove it or replace it with something you can source.
Confirm every product or company detail against your own records. For claims about your own business (features, pricing, guarantees, timelines) check them against your actual product documentation or ask the team responsible. Don't assume the AI understood your product correctly just because it sounded confident.
What to Do When You Can't Verify a Claim
Most fact-checks end in one of three outcomes: the claim checks out, it's flat wrong, or you simply can't find a source either way. That third case happens more than people expect, and it's where most bad copy slips through.
Here's how to handle each:
- Verified: Keep it, and consider adding the source as a citation if it strengthens your credibility.
- False or fabricated: Cut it. Don't soften it, replace it entirely.
- Unverifiable: Rewrite it as an honest, hedged statement instead of a hard fact.
Before and after rewrites
Here's what that rewrite looks like in practice.
Before: "Studies show that 68% of small businesses fail within their first year due to poor marketing." After: "Many small businesses struggle in their first year, and weak marketing is often part of the reason."
Before: "As marketing expert Jane Whitfield once said, 'Your brand is a promise, not a logo.'" After: "There's a common idea in branding that a brand is really a promise, not just a logo."
Before: "Our software integrates with over 200 platforms." After: "Our software integrates with the major platforms our customers use most, including [list the ones you've actually confirmed]."
Notice that the rewritten versions aren't weaker marketing copy. They're just accurate ones. Readers don't need a fake statistic to believe a real point.
Building This Into Your Workflow
A one-time fact-check habit doesn't stick. Build it into how your team actually produces content.
- Keep a simple checklist next to your content calendar: claims underlined, sources checked, quotes verified, product details confirmed.
- Assign one person as the final check before publishing, even if that person didn't write the draft.
- Treat any AI-generated number or quote as "unverified" by default until someone finds the original source, not the other way around.
- Save verified sources in a shared doc so your team isn't re-searching the same statistics every month.
This takes maybe fifteen extra minutes per piece of content. Compare that to the time it takes to issue a correction, respond to a customer who caught a fake statistic, or explain to a client why a quote in their blog post was never actually said by that person.
Actionable Summary
- Treat every AI-generated statistic, quote, and named source as unverified until you've checked it yourself.
- Do a dedicated read-through just to find factual claims before you edit for style or tone.
- Search exact numbers and quotes in quotation marks to see if a real source exists.
- Confirm product or company claims against your own documentation, not the AI's assumption.
- When you can't verify something, rewrite it as an honest, hedged statement instead of publishing it as fact.
- Build a simple checklist into your workflow so fact-checking happens on every piece, not just the ones that feel risky.