Case study
Why your AI ad writer makes up facts and promises you never made
A client's AI-written ad scripts sounded polished and the automated review passed them, yet some promised things the brief never said. Here is how the invented lines appeared, and the checks that now catch them.
Key takeaways
- A brief should say what you do not have, because an AI ad writer treats silence as room to invent.
- A second AI gives you an opinion, not a check; facts need a rule that compares each claim with your brief.
- Any term, deadline or figure in a script that your brief does not state should be treated as invented.
- If a tool reports a length problem, count the words yourself before changing the script.
- Keep customer quotes marked unverified until someone confirms them with the customer.
An AI ad writer makes up facts when the brief leaves a gap, and the result reads as smoothly as everything around it. When a client asked us to move their AI-written ad scripts into the video platform their team works in, the lines were polished and the automated review passed them, yet some of them said things the brief never said. The writing was in the platform within about a day of the owner asking, and the checks were tightened over the following day, each one after a live run showed a gap.
What an invented line in an ad costs you
An invented line in an ad is a promise you have to keep or explain, and it reads exactly like the true lines around it. A long sales script, written on a brief whose proof section was empty, planned a part called real results and filled it with a named customer and an invented figure. Short ads for an offer whose charge the brief left open told viewers it was a one-off charge, with no subscription and nothing hidden. One ad added a deadline the offer never had. Each line reads like the true lines around it, which is why a busy team approving scripts in batches can miss it. An invented testimonial or term is the kind of claim platforms turn down, and one the business then has to keep or explain.
Why an AI ad writer makes up facts
The writer was not lying on purpose: it was not told what the brief left out, so it filled the silence the way ads usually do. Give an ad writer a brief with no proof and it sees a sales script without a proof section. Scripts of that shape usually carry a customer quote and a figure, so it wrote both. Leave the charge open and it reaches for the reassuring words ads usually carry. Nothing in the tool knew the difference between a fact the business gave it and a fact it had supplied itself.
Why the AI reviewer did not catch them
The reviewer was a second AI model, asked whether every claim was supported, and it was as loose with facts and numbers as the first. It passed most of these lines. It also miscounted. On a live run it called all three scripts over the length asked for and marked it serious, while a plain count put them inside the range; a rewrite on that basis would have padded or cut good scripts. In another run it passed every check and still asked for a revision, with nothing to revise. In a third, two of the three scripts shared the same long sentence word for word, and the reviewer called them distinct.
What we changed
We took the facts out of the AI's hands and left it only the writing; everything that can be checked is now checked by plain rules against the brief.
- Every claim in a script is quoted word for word and must point to the part of the brief it rests on. A claim that points to nothing the run was given fails, whatever the reviewer said about it.
- Rules read every hook, body and call to action for offer terms and figures. A term the brief does not state fails the script.
- A brief with no proof now says so at every step, so no part is planned around results nobody gave.
- A customer quote in the brief stays marked unverified until the team itself confirms it.
- The rules count the words and overrule the reviewer, flag a passage of eight words copied from another script in the same set, and require every answer to fit a fixed form so no run is lost to an answer that cannot be read.
What the business has now
The writing now sits in the client's video platform, where the team writes, and the earlier scripts were moved into it. On the first live run after the changes, the rules caught one script carrying the added deadline and one repeating another word for word, and sent both back for a targeted revision. When the long script was run again on the brief that now says it has no proof, it planned no proof section, named no customer and used no figure.
What you can check yourself
- Take three AI-written ads and mark every fact, figure and promise in them. Find each one in your brief; anything you cannot find was made up.
- Leave one section of a brief empty on purpose, such as proof, and see whether the AI fills it anyway.
- Count the words of a script yourself and compare the result with what your AI tool says it counted.
- Read the offer lines twice: one-off, no subscription, nothing hidden and refund promises are the ones it adds most easily.
- Put two scripts side by side and look for a sentence they share word for word; a repeat usually means one of them was not written from the brief in front of it.
What to do next
The useful question about an AI-written ad is not whether it sounds right but whether every fact in it came from you. Ask whoever set up your writing tool for three things: every claim pointing back to the line of the brief it came from, offer terms checked by plain rules rather than by a second model, and a word count you can verify. Until those exist, read the proof lines and the offer lines last, and treat any term you never wrote down as invented. Teaching an ad writer to answer to your brief, rather than to its own idea of an ad, is work our engineers do for clients.
Questions people ask
Can I trust a second AI to check the first one's facts?
Not on its own. In this work the checking model passed most of the unsupported lines and once asked for a revision with nothing to revise. A model gives you an opinion; a fixed rule that compares a claim with your brief gives you an answer.
What happens if I leave part of my brief empty?
The tool usually fills it. A brief with an empty proof section produced a script with a named customer and a figure, because that is the shape ad scripts usually take. Write "no proof yet" in the brief rather than leaving the space blank.
How do I stop an AI ad writer inventing a price promise or a deadline?
State the offer in the brief in the plainest terms you can: what it costs, what that includes, whether the charge repeats, whether places are limited and until when. Have every mention of a term or figure checked against that text by a rule rather than by another model. Anything the brief does not state should fail the script.
Why does my AI tool say my script is the wrong length when it isn't?
A language model estimates rather than counts. If your tool keeps complaining about length, have the count done by a plain rule and compare it with the count you make yourself.
Should a customer quote in my brief be treated as confirmed?
No. A quote your team has not yet checked is a claim like any other, and it stays unverified until someone confirms it with the customer. Marking it that way stops it being used as proof before it is one.