The ASA, the CMA and Claims Made by Machines
In short UK advertising rules hold the advertiser responsible regardless of who or what drafted the claim; walks through the substantiation, comparative-claim and testimonial rules most likely to be breached by unchecked AI copy.
A model will write “clinically proven”, “the UK’s number one” or “loved by thousands of customers” in under two seconds. It has no idea whether any of it is true. You do, or you’re supposed to. Under UK rules, that’s where the responsibility sits.
This post covers the parts of the ASA rules on AI advertising that unchecked machine copy is most likely to break: substantiation, comparative claims and testimonials. It also covers where the CMA comes in, because the ASA isn’t the only body that can cause you a bad quarter.
Who is responsible? You are.
The CAP Code (the rulebook the ASA enforces) doesn’t have a carve-out for machine authorship. Rule 3.1 says marketing communications must not materially mislead. Rule 3.7 says you must hold documentary evidence for any objective claim before you publish it. The advertiser is the one who answers to the ASA, whether the line came from a freelancer, an intern, ChatGPT, Claude or Jasper.
“The tool wrote it” isn’t a defence, and it isn’t mitigation either. The ASA rules on what the ad says and how consumers are likely to read it. How it was drafted doesn’t enter into it.
The same logic applies to the CMA. Under the Digital Markets, Competition and Consumers Act 2024, which gave the CMA direct fining powers for consumer law breaches from April 2025, the penalty ceiling is 10% of global annual turnover. The CMA’s remit covers misleading actions, fake reviews and drip pricing. A small business is unlikely to see that maximum. But the point for a one-to-five-person team is that the regulator’s route to you has got shorter, and the person named on the correspondence will be whoever approved the copy.
Substantiation: the rule most AI copy breaks
Rule 3.7 is the big one. Language models are fluent in the vocabulary of proof (“proven”, “tested”, “independent studies show”) without any of the proof.
Here’s a realistic case. You ask an assistant for a landing page intro for a project management tool and get this back:
Teams using FlowBoard save an average of 6 hours a week and
deliver projects 32% faster. Clinically tested. The UK's
most trusted planning software.Four claims, none from your data. Check each against what you’d need:
| Claim | Type | Evidence the ASA would expect |
|---|---|---|
| “6 hours a week” | Objective, quantified | Methodology, sample size, who was measured, when |
| “32% faster” | Objective, comparative (faster than what?) | Baseline, and a like-for-like comparison |
| “Clinically tested” | Implies medical-grade trial | Nothing relevant exists; delete |
| “Most trusted” | Superlative | Independent survey or ranking that supports it |
The invented numbers are the dangerous bit. “32%” looks like something a researcher produced. The model produced it because percentages are a common pattern in marketing text. If you ship it and get a complaint, you’ll be asked for the study. There isn’t one.
If you have real data, the claim has to match it. Say a customer survey of 212 users found a median self-reported saving of 2.4 hours a week. The claim you can make is something like “In a 2026 survey of 212 FlowBoard users, the median reported time saving was 2.4 hours a week”, with the survey details held on file. That’s less punchy than 6 hours. It’s also defensible, and defensible is the standard.
A prompt-level fix that works
Don’t ask the model to write claims. Give it a closed list of approved ones and tell it to use nothing else:
Use ONLY the claims in the APPROVED CLAIMS block. Do not add
statistics, awards, rankings, study references or superlatives.
If the copy needs a proof point you don't have, insert
[CLAIM NEEDED: describe] and stop.
APPROVED CLAIMS
1. Median reported saving 2.4 hrs/week (survey, n=212, Mar 2026)
2. Rated 4.6/5 on Trustpilot (1,380 reviews, checked 1 Oct 2026)This won’t make the model perfectly obedient, so you still read every line. But it turns invention into visible placeholders, which are easy to catch, and invented stats are not.
Comparative claims: naming a rival is the easy part
Comparative advertising is permitted under the CAP Code (section 3, rules 3.32 to 3.36 and the related guidance), provided it isn’t misleading, compares like with like, and objectively compares a material, relevant, verifiable and representative feature. It must not take unfair advantage of a competitor’s reputation or discredit them.
Models trip on this in three ways.
- Stale facts. An assistant will tell you a competitor’s plan costs £15 a month. That may have been true two years ago. If your page says “half the price of Rival” and Rival’s price changed in June, you’re now making a false comparison. Verify against the live pricing page and save a dated capture.
- Unlike-for-like. “Cheaper than Asana” compared against what tier? Your £12 plan against their £24 plan with twice the features isn’t a fair comparison.
- Hallucinated features. The model says a competitor “has no reporting dashboard”. It does. You’ve now made a false and disparaging statement, which is the kind of thing that moves from an ASA complaint to a solicitor’s letter.
A working check for any comparative line: write down the competitor, the tier, the date you checked and the URL. If you can’t fill all four boxes, the line doesn’t run.
Testimonials and reviews: the quietest risk
Section 3 of the CAP Code also covers endorsements and testimonials: you must hold written permission from anyone quoted, and the quote must be genuine and representative. The CMA’s rules on fake reviews are blunter, and under the 2024 Act publishing or commissioning fake reviews is on the list of banned practices.
AI makes this easy to get wrong without any bad intent. Ask a model for “three customer testimonials for the case study page” and it’ll write them, in plausible voices, with plausible first names and job titles. Even labelled as placeholders, they have a way of surviving into the published version.
Rules to adopt on a team of any size:
- Models never write quotes attributed to a real or implied person. Quote only from recorded or written sources you hold.
- Tidying a genuine quote for length is fine. Changing its meaning, or merging two customers’ comments into one, is not.
- Do not use AI to generate or bulk-post reviews, and do not ask staff or friends to do it either.
- Keep the permission email alongside the quote in the same folder.
Other things that catch people out
A few quick ones, because they come up in the same review.
“Results not typical” doesn’t cure a headline. If the headline implies a result most customers won’t get, a small-print disclaimer won’t fix the overall impression.
Health and green claims. Nutrition and health claims are controlled by the CAP Code’s food and health sections and by pre-approved lists. Environmental claims such as “carbon neutral” or “eco-friendly” need specific evidence under the CMA’s Green Claims Code. Models produce these phrases without a second thought.
Disclosure. If an influencer or affiliate posts AI-assisted content for you, it still needs to be clearly identifiable as an ad (#ad up front, not buried at the end of a hashtag cluster). AI assistance itself doesn’t have to be labelled under the CAP Code, though the ASA has made clear that the ad’s content is judged however it was made. The wider governance picture, including policies and record-keeping, is in Governance, Disclosure and UK Compliance.
A review step that takes ten minutes
You don’t need a legal department. You need a repeatable pass, and it should happen after generation and before scheduling.
- Highlight every sentence containing a number, a superlative (“best”, “leading”, “first”), a comparison, a testimonial or a health or environmental term.
- For each, find the evidence file. If there isn’t one, cut the line or rewrite it to what you can show.
- Log it: claim, evidence link, date checked, who approved.
- Set a re-check date. Prices and rankings drift, so quarterly is a sensible default for evergreen pages.
On a typical 1,200-word blog post, this usually flags between 5 and 12 sentences. At roughly a minute each, that’s your ten minutes. Teams that skip it usually find out the expensive way, via a competitor complaint to the ASA, which is public: rulings are published by name on the ASA website, and a ruling against your brand stays searchable.
The tools will keep getting better at sounding right. Whether it is right remains your job, and the regulator has already said whose it is.