AI hallucinations in marketing are statements a model presents with confidence but has invented: a statistic, a product claim, a quote, a price or a policy. You catch them by treating every factual sentence in AI copy as unverified until it is traced to a source you hold, then running a fixed pre-publish checklist with a named person signing off before anything ships.
Key takeaways
- A hallucination is a fluent, confident falsehood, and marketing copy is full of the specific facts models are most likely to get wrong.
- Regulators in the UK and Singapore are clear that the advertiser, not the tool, is responsible for every claim in an ad.
- The fix is a process: tier copy by risk, verify every claim against a source you hold, and log a human sign-off.
- Grounding AI in your own approved sources, and letting it say "unknown", removes most hallucinations before QA ever sees them.
What is an AI hallucination in marketing copy?
An AI hallucination is a plausible statement that is not true, delivered with the same confidence as everything else the model writes. OpenAI's September 2025 research describes it as a model confidently generating an answer that isn't true, and argues that standard training and evaluation reward guessing over admitting uncertainty.
The same paper explains why arbitrary, low-frequency facts are the hardest for a model to get right: they cannot be predicted from patterns. Marketing copy is built from exactly those facts, such as prices, launch dates, award names, survey figures and terms and conditions.
In marketing copy, hallucinations usually fall into six types:
- Invented numbers: a market statistic, a growth rate or a "9 out of 10 customers" line with no study behind it.
- Fabricated sources: a report, author or URL that does not exist, or a real report cited for a figure it never contained.
- Wrong product facts: an outdated price, a feature borrowed from a competitor, a specification from last year's model.
- Unsupported claims: "clinically proven", "the fastest" or "number one in Singapore", written because they sound like advertising.
- Made-up voices: testimonials, reviews or expert quotes attributed to people who never said them.
- Policy errors: promotion dates, eligibility rules or refund terms that differ from the live terms.
Who is responsible when AI copy gets a claim wrong?
The brand that publishes it. Advertising regulators treat AI copy like any other copy: the rules on misleading claims apply however the ad was made.
In the UK, the ASA's June 2025 guidance for marketers using AI says marketers are ultimately responsible for their ads and must be able to substantiate their claims. Its June 2026 note on AI and deepfakes goes further: in the Stripe & Stare ruling, an ad built with automated tools remained the advertiser's primary responsibility, and advertisers are told they must always sense-check outputs.
Singapore points the same way. The Competition and Consumer Commission of Singapore's Guide on Quality-Related Claims (October 2025) sets five principles: claims should be true and accurate, clear, meaningful, accompanied by material information and supportable by credible evidence. Consumers can report misleading ads to the Advertising Standards Authority of Singapore, which publishes the Singapore Code of Advertising Practice.
Tribunals have ruled on hallucinated customer information too. In Moffatt v. Air Canada (February 2024), a British Columbia tribunal held the airline responsible for a chatbot that described its bereavement-fare policy incorrectly, and rejected the argument that the bot was responsible for its own words. What an AI says on your channels, you said.
Why are marketing teams more exposed in 2026?
Because volume has outrun review. Teams publish more AI-assisted copy across more channels, and checking has not scaled with it.
SOMIN's own workspace data shows marketers feel this. In SOMIN's audience-research workspace, a 30-day whitespace study of 220 posts (September 2026) surfaced the tension theme "Navigating the AI Marketing Maze" with a sample size of 91. Its summary names AI-generated content that requires substantial human revision, and over-reliance on AI tools without adequate oversight leading to misallocated budgets.
The ads-optimisation workspace says the same from the paid side. Its September 2026 whitespace study of 156 posts includes the theme "Navigating the Challenges of AI and Marketing Effectiveness" (sample 23), which links superficial adoption of AI tools without proper human oversight to falling content quality. And in the social-listening workspace's October 2026 study of 411 posts, the largest theme (sample 99) covers AI in marketing and social media, including concern over consumer trust eroded by AI-generated content.
A hallucination that reaches the feed spends exactly the trust these teams are worried about.
How do you check AI copy before it ships?
Run every piece through the same checklist, scaled to its risk. The table below is the pre-publish checklist we recommend for AI-assisted marketing copy; the right-hand column is what counts as a pass.
| Check | What to look for | Passes when |
|---|---|---|
| 1. Numbers and statistics | Any figure, percentage, ranking or date | Each figure links to a source someone on the team has opened, and the source says it |
| 2. Product and price facts | Prices, specs, availability, ingredients, offers | It matches the current product master or price list, not the model's memory |
| 3. Performance and superiority claims | "Best", "No.1", "proven", "x% faster" | Evidence is on file before publishing, not found afterwards |
| 4. Competitor comparisons | Named or implied rivals | The comparison is like-for-like, current and fair |
| 5. Quotes, reviews and testimonials | Any named person or customer voice | It is real, attributable and permissioned, with nothing paraphrased into a quote |
| 6. Citations and links | Report titles, URLs, authors | The link resolves and the page says what the copy claims |
| 7. Terms, policies and promotions | Dates, eligibility, refunds, T&Cs | It matches the live terms, word for word where it matters |
| 8. Local detail | Holidays, regulators, currencies, spelling | It has been checked for the market it runs in |
| 9. AI images and video | Product results, people, endorsements | Nothing exaggerates what the product does or implies an endorsement that was not given |
| 10. Sign-off | Who approved it | A named person, a date and the version are logged |
Tier the copy by risk
Not every caption needs ten checks. Tier one is anything with a price, a performance claim, a health, finance or environmental statement, or a named competitor: the full list and a second reviewer. Tier two is everyday copy that contains facts: checks 1, 2, 6 and 10. Tier three is pure tone, such as a headline variant: a voice read and sign-off.
Read for confidence, not fluency
Hallucinations hide in the smoothest sentences. Highlight every sentence that asserts something checkable and ask of each: where did this come from? If the answer is "the model", it does not ship until a source replaces it.
How do you stop hallucinations before QA sees them?
Ground the model in sources you control and give it permission not to know. Most hallucinations disappear when the model writes from supplied facts instead of from memory.
- Give it an approved claims library. Substantiated claims, current prices and proof points, with an instruction to use nothing else for factual statements.
- Ask for a source per claim. Each factual sentence in the draft is marked with where it came from, so reviewers check a list rather than hunt.
- Allow "unknown". OpenAI's research notes that models can abstain when uncertain, so reward a gap marked [CHECK] over a confident guess.
- Retrieve, don't recall. Tools that read your own documents or live category data before writing invent fewer specifics.
- Keep a gate where it matters. Anything that spends, sends or publishes should wait for a person.
This is how SOMIN's AI agents are built. Every recommendation traces back to the ads and engagement signals that produced it, so a reviewer can open the evidence instead of trusting it. SOMIN's proposition records in the audience-research workspace put it plainly: concept tags trace to the specific posts that evidence them, and every AI conclusion can be checked against the posts behind it. On the execution side, the Paid Media agent pushes ads built paused until someone approves the spend, and the CRM agent needs two human approvals before an email goes out.
Who should own AI copy QA?
One named person per channel, with the authority to stop a post. Shared ownership usually means nobody checks the statistic in paragraph three.
From a QA engineering point of view, AI copy deserves the same habits as software: a test plan (the checklist), a defect log of every hallucination caught, and a regression check whenever prompts or models change. After a month, the defect log shows which prompts, products and claim types produce errors, so you tighten the claims library instead of adding reviewers.
FAQ
Can you stop AI from hallucinating completely?
No. Grounding models in approved sources and letting them abstain reduces hallucinations a great deal, but OpenAI's own research says accuracy will never reach 100%. Human review of factual claims stays necessary.
Is a brand liable for false claims in AI-generated ads?
Yes. The UK's ASA says marketers are ultimately responsible for their ads however they were made, and Singapore's CCS expects claims to be true and supportable by credible evidence. A Canadian tribunal also held Air Canada responsible for what its chatbot told a customer.
How do I fact-check AI-written marketing content quickly?
Highlight every sentence that asserts something checkable, confirm each against a source you hold, and fix or cut anything without one. A claims library makes this a list check rather than research.
Do I have to label ads as AI-generated?
In the UK there is no blanket requirement: the ASA's test is whether the audience would be misled without a disclosure, and a label does not fix a misleading claim. Check platform rules and local regulators for each market you run in.
Start this week with two things: write the claims library for your top five products, and put the ten-point checklist in front of whoever presses publish. Then log every hallucination you catch for a month. That log will tell you whether you need better prompts, better sources or a firmer gate.
