Skincare Product Claims: What AI Copy Must Not Say

A skincare product claim is safe when it describes how the product looks, feels or cleans, and you can prove it; it breaks the rules when it says the product treats, cures or prevents a skin condition. To stop AI copy crossing that line, give it a banned-phrase list, then check every draft against your evidence file before publishing.

The trouble is that AI writes confident, persuasive copy by default, and persuasive skincare copy drifts towards medicine very quickly. Ask for "benefits" and you will get "reduces inflammation", "heals breakouts" and "clinically proven". None of that is malicious, but each phrase can turn a cosmetic into something regulators treat as a medicine, or make a promise you cannot evidence. The fix is a routine, not a smarter prompt.

Follow me on Instagram@sagnikteaches

Rules differ between markets, and they change, so I will not quote a rulebook here. Read your own regulator's guidance for each place you sell, and ask a qualified adviser for anything borderline. What I can give you is the workflow: a place for the rules to live, a prompt that uses them, a checklist that catches what the prompt misses, and a log that proves you looked.

Connect on LinkedInSagnik Bhattacharya

The line between a cosmetic and a medicine

In most markets a product is judged by what it is intended to do, and your marketing is the strongest evidence of intent. A cosmetic is applied to the outside of the body to clean, perfume, change the look of, protect, or keep it in good condition. A medicinal or drug claim says the product treats, prevents, diagnoses or cures a disease or alters how the body works. The same cream can sit on either side depending on the words around it. Some regulators also weigh what consumers expect the product to do and whether it contains an ingredient with a well-known therapeutic use, so the formula and the product's reputation can count as well as the copy.

Subscribe on YouTube@sagnikteaches

A quick way to sort a sentence is to ask what the customer would be doing if the claim were true. If they would be improving how their skin looks or feels, you are in cosmetic territory. If they would be treating a named condition, or changing something beneath the skin's surface, you have moved towards a medicinal claim.

Notice that the safe column is not vague. "Softens the look of fine lines" still sells. The skill is writing copy that is specific about appearance and sensation, and silent about physiology and disease.

Why AI copy fails in predictable ways

A language model has read an enormous amount of skincare marketing, including plenty written by brands that did not follow the rules. When you ask for punchy copy, it reproduces the average of that material. Four patterns show up again and again.

• Medical verbs. Heal, repair, treat, cure, prevent, detoxify, reverse, regenerate, rebuild, fight, kill.
• Named conditions. Eczema, psoriasis, acne, rosacea, dermatitis, infections, scarring, hyperpigmentation described as a disorder.
• Invented authority. "Clinically proven", "dermatologist recommended", "scientifically tested" with nothing behind them.
• Absolutes. "Works for all skin types", "100% safe", "no side effects", "permanent results", "instantly erases".

It also copies the ingredient claims it finds online. Ask about retinol and you will get statements about cell turnover, which describe a biological process. Ask about niacinamide and you may get "strengthens the skin barrier". These are claims about what the ingredient does in studies, not about what your product, at your concentration, has been shown to do.

Step one: write down what you can prove

Before you touch a prompt, build the evidence file. This is a single spreadsheet, one row per product, that lists every claim you want to make and what stands behind it. If you cannot fill the evidence column, the claim does not go in the copy.

Substantiation means having proof on file before you say it, not finding some afterwards. The proof must match the claim: a test on twenty volunteers for a "visibly smoother after four weeks" claim, a lab report for "SPF" wording, a supplier certificate for "organic". A study on an ingredient alone rarely supports a claim about your finished product.

Here is a filled-in extract for an imaginary brand with three products. All figures are illustrative.

Notice the perception panel row. A panel of customers who say their skin looks brighter supports a statement about what they said, not a statement that the product brightens skin in general. Quote the panel size and the result accurately, and keep the original questionnaire.

Step two: build the banned and flagged phrase list

You need two lists. The banned list holds words you never use. The flagged list holds words that are fine in some contexts and need a person to look. Keep both in a shared document, because they are what you paste into the AI and what the reviewer checks against.

A starter banned list

• Treat, treats, treatment (of any skin condition), cure, heal, heals, remedy, therapy, therapeutic
• Prevent, prevents, protects against disease, anti-inflammatory, antibacterial, antifungal
• Any named condition: eczema, psoriasis, acne, rosacea, dermatitis, shingles, warts, cold sores
• Repair, regenerate, rebuild, stimulate, boost (when followed by collagen, elastin, cell turnover or immunity)
• Detox, toxins, purify the body
• Doctor-recommended, clinically proven, medically tested, with no file behind them
• Permanent, guaranteed results, works on everyone, no side effects, 100% safe

A starter flagged list

• Natural, pure, clean, chemical-free (rarely meaningful, often misleading)
• Hypoallergenic, non-irritating, sensitive-skin safe
• Anti-ageing, anti-wrinkle (check how your market treats these; "reduces the look of" is usually safer)
• Dermatologist tested or approved
• Percentages and time claims ("in 7 days", "92% saw")
• Before-and-after photos and testimonials
• Sun protection of any kind
• Free-from claims ("paraben-free"), which can imply other products are unsafe

Keep these lists short enough to read. A list of 300 words nobody reads is worse than 40 that everyone knows.

Step three: put the rules inside the prompt

A prompt cannot guarantee compliance, but it removes most of the obvious failures. Put the rules in a Project in ChatGPT or Claude so every conversation starts with them, rather than pasting them in each time. If you have not set that up, giving AI your business context shows how.

You write product copy for a skincare brand. Rules:
1. Describe how the product looks, feels, smells and cleans. Never say it treats, prevents, cures or repairs anything.
2. Never name a skin condition or disease.
3. Use only claims from the EVIDENCE table below, in the exact allowed wording.
4. Do not use any word on the BANNED list. If a word on the FLAGGED list seems needed, use it and add it to a "Needs review" section.
5. If the brief asks for something the evidence table does not support, do not write it. Say which part you left out and why.
6. After the copy, add a table: each claim in the copy, and the evidence row that supports it.

BANNED: [paste list]
FLAGGED: [paste list]
EVIDENCE: [paste the rows for this product]

Brief: product page intro, 90 words, for the 50 ml barrier cream. Tone: warm, plain, no hype.

Rule 5 matters most. Without it, the model quietly fills gaps with plausible-sounding benefits. With it, you get a refusal you can act on.

What the output looks like

Here is an illustrative response to that prompt, trimmed.

Meet the cream for skin that feels tight by mid-afternoon. It sinks in quickly, leaves no sticky film, and keeps skin feeling soft and moisturised for up to 24 hours. Fragrance-free and light enough for day or night.

Claims used:
- "Moisturises for up to 24 hours" - evidence row 1 (hydration test)
- "Fragrance-free" - evidence row 3 (ingredient list)

Left out: you asked for "great for eczema". No evidence is on file and naming a condition is on the banned list.

Needs review: none.

That draft is usable. Even so, check it by hand: "keeps skin feeling soft" is a sensation claim, and "light enough for day or night" is a fair description. The prompt did its job when the refusal appeared.

Step four: the checklist a person runs on every draft

Treat the AI's own claims table as a starting point, not a verdict. A model can mark a claim as "supported" when the evidence row says something slightly different. The person doing the check should work through the following in order, and it should take about five minutes per page.

1. Search for banned words. Use the find tool, not your eyes. Include variants: "heals", "healing", "healed".
2. Search for named conditions. Also check image captions, alt text and file names, because those are copy too.
3. Match every claim to a row. If the sentence says "24 hours" and the row says "up to 24 hours", fix the sentence.
4. Check numbers against the report. Sample size, duration and the exact percentage. Rounding 82% up to "over 85%" is a false claim.
5. Check implied claims. A photo of a clear-skinned model beside the word "breakouts" implies treatment of acne even if the sentence does not.
6. Check comparisons. "Better than" another named brand needs a head-to-head test and carries extra risk. "Free from" claims should not imply competitors are unsafe.
7. Check the ingredient list matches. A "vegan" or "fragrance-free" page must agree with the current label, not last year's.
8. Check the flagged words. For each, decide: evidence on file, reworded, or removed.
9. Record the decision. Initials, date, page address.

Where the same copy appears: a map of surfaces

Claims travel. The clean product page you approved on Monday is copied into a marketplace listing on Tuesday, a social caption on Wednesday and an email on Thursday, each edited by someone with a slightly different idea of what is acceptable. Run the check on every surface, not just the website.

• Product pages and collection pages. Include the short summary shown in search results.
• Marketplace listings. These often have their own category rules on top. See how small sellers use AI for listings and reviews.
• Emails and text messages. Subject lines are claims too: "Say goodbye to breakouts" is a treatment promise in four words.
• Social posts, captions and video scripts. Spoken claims count, so check the script, not just the caption.
• Influencer briefs. If you pay or gift, you are generally responsible for what they say. Send them the approved wording, and the banned list.
• Customer service replies. A chatbot that tells a customer the serum "will help your eczema" has just made a medicinal claim in private.
• Packaging and leaflets. The slowest to fix once printed.

The customer-service point is easy to miss. If you use an AI chatbot, load the same banned list into its instructions, and test it by asking leading questions such as "Will this help my psoriasis?" The correct answer is a polite refusal to give medical advice and a suggestion to see a doctor.

A four-product brand clears its site in a weekend

Take an imaginary brand with a cleanser, a barrier cream, a vitamin C serum and an eye cream, selling through its own store plus one marketplace. The owner has 38 pages of copy, much of it drafted with AI over the past year. This is an illustration, not a real case.

Saturday morning. The owner builds the evidence file. It takes two hours, because the test reports are in four different inboxes. At the end there are 31 claims wanted, 19 with evidence and 12 without.

Saturday afternoon. She exports all page text to a spreadsheet, and asks the AI, with the banned list loaded, to flag every sentence containing a banned word or a flagged word, and to output the page, the sentence and the reason. The AI returns 64 hits. A manual search finds three more that it missed, including "calms" inside a longer phrase. That gap is why the human step exists.

Sunday morning. She works through the hits. Of the 67, 41 are rewritten by the AI using the allowed wording column, 15 are deleted, and 11 are kept because the evidence exists. The 12 unsupported claims are dropped from all copy.

Sunday afternoon. Check log completed: 38 pages, initials, date. The owner sets a rule that any copy change, however small, goes through the five-minute checklist, and puts a recurring reminder in her diary to revisit the top pages every six months.

Total time: roughly eleven hours across the weekend. The claims-check step now adds about five minutes to each new page.

Five more cases worth rehearsing

A supplier's marketing sheet becomes your copy

Your ingredient supplier sends a PDF saying their peptide "stimulates collagen synthesis by 70%". You paste it into the AI and ask for a product description. The model uses the figure. The 70% came from a lab test on isolated cells at a different concentration, so it does not describe your finished product. Rule: supplier data may inform the formulation, but it is not your substantiation. Mark such sheets as background only in your prompt.

An AI-written subject line

For a spring campaign the model suggests "Banish winter eczema for good". It names a condition, uses a cure verb and promises permanence. The check catches all three. The rewritten version, "Soft skin, even when the heating's on", describes a feeling the cream can honestly deliver.

A glowing review

A customer writes "This cleared up my dermatitis in a week". You would like it on the product page. Do not use it. If you wish to show reviews, choose ones that describe feel, scent, texture and look, and keep a policy that you never display reviews naming conditions. Do not edit a review to remove the condition either, because altering what the customer said is misleading.

A new market

You begin selling into a second market. A phrase you used freely at home is a regulated claim there, or the reverse. Before you translate or localise any copy with AI, add the new market's rules to your banned and flagged lists, and run the full sweep again. Translation can change the strength of a claim, so check translated copy too, ideally with a native speaker.

The "just a small tweak" edit

A team member asks the AI to "make this page punchier". The model replaces "helps skin feel comfortable" with "relieves irritation". Nobody notices, because the change was not on anyone's list of things to review. Fix: any edit, even a tone change, re-triggers the checklist. Make the rule about the page changing, not about the intention.

When the AI says a claim is fine

Do not ask the model "is this claim compliant?" and treat a yes as an answer. It will often say yes, because it cannot see your evidence or your markets' current rules, and it tends to agree with the framing of the question. Ask it instead to do tasks it is good at: find medical-sounding wording, list implied claims, rewrite a sentence to describe appearance only, compare copy with an evidence table you have pasted in.

A useful second pass is to ask the model to play a sceptical reviewer.

Act as a sceptical reviewer of skincare marketing. Read the copy below. List every statement that:
(a) names or implies a skin condition or disease,
(b) says the product changes how the body works,
(c) gives a number, a percentage or a time frame,
(d) uses an authority claim such as tested, recommended or proven.
For each, quote the sentence and say which evidence a reader would expect to see. Do not say whether the copy is compliant.

COPY: [paste]

The last line matters. You want an inventory of exposure, not an opinion on legality. The inventory then goes to the person who owns the evidence file.

Keeping a record that shows you looked

If a regulator, a marketplace or a competitor ever questions a claim, the first thing they will ask is what you had when you published it. A simple log answers that. Keep one row per publication: date, page or message, product, the claims used, evidence references, reviewer and outcome. Store the evidence documents with dates, because a test report from five years ago on a formula you have since changed does not support today's product.

It also helps to log changes to the formula. A reformulated cream that now contains a different humectant may no longer match its hydration test. When the formula changes, mark every dependent claim as "retest or remove" until you have proof again.

For a wider view of making sure each AI-written statement can be traced to a source, how to back up every claim AI writes in your marketing covers the general routine, and the eco-friendly claims on skincare packaging tutorial covers the green-wording side, which has its own rules.

Who signs off, and how fast it should be

The routine fails when it is slower than the campaign calendar. Aim for a same-day turnaround on routine pages, and a clear rule on who can approve. For a small brand it usually works like this.

• Writer or AI operator produces the draft and the claims table.
• Checker, a different person, runs the nine-step list. If you are a solo founder, leave the draft for a day and check it as if someone else wrote it; fresh eyes catch more.
• Product owner confirms the evidence file is current for any new claim.
• Adviser reviews the borderline cases in a batch, rather than one at a time, to keep costs down.

The batching point deserves attention. Keep a running list of questions for your adviser: "Can we say 'visibly calms' about redness?" Send ten at once. The answers then feed straight back into your banned and flagged lists, so you pay to learn each rule only once. For more on building review into AI work without slowing it, see setting up human review for AI work.

Signs your routine is working, and signs it is not

Look at three things a month after you start. First, how many drafts the checker rejects. If the number is zero, your prompt is working or your checker is not looking; sample five approved pages and test them yourself. If it is above a third, the rules in the prompt are probably too thin. Second, how many refusals the AI makes under rule 5. A healthy number is a handful per product; none suggests the evidence table is not being read. Third, how long the check takes. If it creeps beyond ten minutes, shorten the lists and fix the common failure at the source.

One practical measure is a monthly spot check. Pick five live pages at random, run the full list, and record what you find. Anything you find goes straight back into the banned or flagged list, so the routine learns from its misses.

Finally, remember this is guidance for building a process, not legal advice. Rules on cosmetics, sun protection, "free from" wording and testimonials differ by market and change over time. Treat your own regulator's published guidance as the source of truth, and let a qualified adviser decide on anything you cannot settle with the evidence in your file.

Questions skincare owners ask about AI copy and claims

Can AI tell me whether a skincare claim is legal?

No. AI can flag wording that sounds medical and suggest softer phrasing, but it cannot know the rules in your selling markets or what evidence you hold. Use it as a first filter that catches obvious problems, then have a person who knows your formulations and a qualified adviser decide on borderline wording.

Do customer reviews count as my claims?

They can. Advertising guidance generally says an endorsement cannot make a claim you could not legally make yourself, so a review you feature on a product page or in an advert is treated much like your own words. A reviewer saying a serum cleared their eczema is a medical claim you are now repeating. Curate which reviews you display, and run them through the same checklist as your own copy.

Is 'dermatologist tested' a safe phrase?

Only if it is true and you can show it. Keep the report, the test dates and who ran them. The phrase tells customers a professional tested the product, and it is often read as meaning dermatologist approved or recommended, which is a different and stronger claim. Say exactly what was done.

Should I ban the word 'natural' from AI prompts?

Treat it as a flagged word rather than banning it. 'Natural' has no single agreed meaning, so it can mislead when the formula contains synthetic ingredients. If you use it, say what you mean, such as the share of ingredients from plants, and keep the calculation on file.

How often should I rerun the check on old pages?

Rerun it whenever the formula, the packaging or the page changes, and when you start selling in a new market. A full sweep of your top twenty pages every six months is a sensible extra habit, because AI edits made by staff over time can quietly add new wording.

Further reads

• Health Food Store Marketing Ideas: AI Posts Without Health Claims — Another sector where health wording gets posts into trouble.
• Does Writing Etsy Listings With AI Hurt Your Search Ranking? — Marketplace listings carry the same claims risk.
• AI Marketing Best Practices for a Local Business: A Pre-Post Checklist — A pre-post checklist you can adapt for every channel.
• How to Write Website Articles With AI That Still Rank on Google — Write longer product-led pages without drifting into medical claims.
• How to Review an AI Tool After 90 Days: Keep, Fix or Cancel — Judge whether the claims-check step is earning its time.
• Octane AI Pricing for Skincare Quizzes: Credits, Plans, Payback — What a skincare quiz costs on Octane AI once credits, overage and AI features are counted, and how many extra orders it needs to pay back.
• AI for Beauty Brands: Product Copy, Photos and Customer Questions — A worked product page for a small beauty brand: AI drafts the copy, photos and replies, and a claims check keeps every line provable.
• AI for Cosmetic Formulation: What It Can Draft and a Chemist Must Check — What AI can draft for a small cosmetics brand (briefs, ingredient lists, batch sheets) and what a qualified chemist or safety assessor must still check.
• Cosmetic Label Requirements: A US, UK and EU Checklist Using AI — A region-by-region cosmetic label checklist, a prompt that makes AI list what is present or missing, and the manual checks no chatbot can sign off.
• Cosmetic Product Photography With AI: Textures and Honest Colour — Which AI edits are safe on lipstick, foundation and cream photos, which change the shade, and how to check colour before it costs you returns.
• INCI Ingredient List Check With AI: Names, Not Safety Sign-Off — A checklist for indie skincare brands: what AI can check on an INCI list, what it cannot, and what to give your safety assessor.
• Skincare Replenishment Emails: Time Refill Reminders From Orders — Work out when each skincare product runs out from pack size and order history, then build a tested refill-reminder flow in Klaviyo.
• AI for Skincare Products: Six Jobs a Small Brand Can Delegate — Six brand jobs for a small skincare label to hand to AI, each with a prompt, an example and the check that keeps claims safe.
• Can AI Write a Cosmetic Safety Report? What an Assessor Signs — What AI can prepare for a cosmetic product safety report, what only a qualified assessor can sign, and how to brief one so the job goes faster.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: regulators' published definitions of cosmetic and drug products and their common criteria for substantiating cosmetic claims (read the versions for each market you sell into), published endorsement guidance on testimonials and influencers, and the ChatGPT and Claude help pages on Projects.