AI hallucination in writing: how to check for invented quotes, facts and stories

You have a draft in front of you. It sounds like you. It reads well. And somewhere in the third paragraph is a sentence you have no memory of saying: “A 2021 study found that 70 percent of clients improve within six weeks.” Did you say that? Did you read it somewhere? Or did the tool produce a plausible-sounding fact on its own?

This guide gives you a method for answering that question for any AI-written draft, with any tool. It is the method we built Frazo around, but you do not need Frazo to use it.

Why AI invents things in the first place

Language tools generate text by predicting what is likely to come next. When you ask for a post about anxiety, a likely next sentence is one with a statistic, a study, or a client anecdote, because that is what posts about anxiety usually contain. The tool does not know whether the sentence is true. It knows it is likely.

The result is invention that sounds credible. It is not random nonsense, which would be easy to spot. It is a fluent, specific, confident detail that nobody told it to include.

The risk is not theoretical. In 2023 a US federal court sanctioned lawyers who filed a brief that cited cases a chatbot had made up; the court’s June 22, 2023 order in Mata v. Avianca is public. An expert posting under their own name faces a smaller version of the same problem: a reader, a colleague or a regulator finds a claim you cannot support.

Step 1: Keep a source of truth

You cannot check a draft against your memory. Memory is the very thing that blurs when you read fluent text. You need a fixed record of what you actually said or wrote.

The simplest version: answer by voice, keep the recording, and get a transcript. If you write from notes instead, keep the notes. The point is that there is a document to compare against that the tool did not write.

If you work with a tool that never shows you the original material, you are checking a draft against nothing.

Step 2: Split the draft into claims

Read the draft sentence by sentence and put every sentence in one of three groups.

1. Said. You can find it in your source, in nearly the same words.
2. Reworded. You can find the idea in your source but the wording is new. Check that it says the same thing, not a stronger version of it.
3. New. You cannot find it. Stop here and ask where it came from.

The third group is where invented quotes live. Most drafts have a few such sentences, and many of them are harmless connectors like “Here is why that matters.” Mark them anyway, because the harmless ones and the dangerous ones look the same until you read them closely.

Step 3: Treat these six categories as high risk

When a sentence lands in the “new” group, ask whether it falls into one of these categories. These are the details most worth checking.

Numbers and statistics. “70 percent”, “three out of four”, “in just 21 days.” If you did not say the number, delete it or replace it with a source you can link to.

Named studies and experts. “Research from Harvard shows…” Search for it. If you cannot find it in two minutes, it probably does not exist in that form.

Quotation marks. Any sentence in quotation marks that you are presenting as someone’s words. A quote needs a real speaker and a real source, or it should become a paraphrase, or go.

Client stories. Anything beginning with “One client…” If the story is not something you told, it is invented. If it is something you told, it may still identify a real person, which is a separate problem.

Credentials and experience. “As a therapist with 15 years of practice…” A tool can easily write a credential for you that is not yours.

Causes and promises. “This will reduce your stress”, “always works.” Cause-and-effect claims and promises of results can be wrong even when every individual fact is correct.

Step 4: Decide what to do with each flagged sentence

For every sentence in the “new” group you have three honest choices.

• Remove it. Usually the best option. Drafts are better with fewer unsupported claims.
• Source it. If it is a fact, find a real reference, read it, and link or name it.
• Own it. If it is something you believe and want to say, say it in your own words and let it be yours. It is fine for a post to contain things you add after the interview. The failure is passing them off as something that was in the source.

What you should not do is leave a sentence in because it reads well.

Step 5: Spot-check the reworded sentences

Rewording can slide into overstatement. If you said “it can help some people,” a draft might say “it helps people.” Read your original phrase and the draft phrase side by side and check for three kinds of drift:

• Hedges dropped. “Sometimes” becomes “always.”
• Scope widened. “In my practice” becomes “in general.”
• Order of cause changed. “People who sleep badly feel more anxious” becomes “Anxiety makes you sleep badly.”

Step 6: Do a read-aloud test

Read the final post aloud. If a sentence feels like it came from someone else, trace it. This is not a substitute for the comparison above, only a cheap last check, and it catches tone slips that a transcript comparison would not.

A 10-minute checklist

• I have the recording or notes the draft was built from.
• I marked every sentence as said, reworded, or new.
• Every number has a source or is deleted.
• Every named study or expert is real and I found it myself.
• Every quotation has a real speaker, or I rewrote it as a paraphrase.
• Every client story is mine, and no one could be identified from it.
• Every credential is true.
• No “always”, “never”, “will” or “promised” appears that I did not say.
• I read the final text aloud.

Doing it by hand versus building it into the tool

The method above works, and it is slow. Comparing a draft to a transcript sentence by sentence takes ten minutes or more for a single post, which is why people skip it.

Frazo is building this comparison into the product, but it is not available for English drafts yet, so for English you do it by hand as above. Where the check exists, it works like this. You answer questions by voice. Every sentence in the draft carries the place in the recording it came from, shown as a timecode, so the source is one click away. A sentence that is not in the recording is flagged “Not in the recording,” and the draft cannot be accepted until you decide whether to remove the phrase or keep it. Sentences you write yourself are marked as written by you, so the record stays honest in both directions. A check line at the bottom reports how many sentences were checked against the recording.

This does not make a draft perfect, and no tool should promise that. It makes the checking faster and puts the evidence next to the claim. The check is not available for English drafts yet, and English drafts are in early access, so for English text the manual method above is the one to use. To see how Frazo works, try Frazo.

Using this on other people’s drafts too

If you work with a human ghostwriter or an editor, the same method applies. Ask for the source material along with the draft: the interview recording, the notes, the links. A good writer will be glad to give it to you. A draft with no traceable source is a draft you cannot sign.

For ideas on what to record in the first place, see our list of content interview questions for experts. To see how price and process compare between a tool and a person, read AI ghostwriter vs human ghostwriter: what each costs.

The short version

Keep the original. Sort every sentence into said, reworded or new. Be most suspicious of numbers, studies, quotes, client stories, credentials and promises. Remove, source, or own each new sentence. Whatever tool you use, the rule is the same: a post under your name should contain nothing you cannot trace.