What can actually get you in trouble when you publish AI content?

A pre-publish compliance pass is the review you run on marketing content that a generative model wrote or materially shaped — a landing page, an ad, an email, a product page — before it goes live. Most marketers arrive asking one question: do I have to say this was made with AI? Across the documents gathered here, that is the narrowest obligation in play. The ones that carry weight are older than generative AI and indifferent to who typed the words: hold evidence for your objective claims before you publish them, and do not put words in the mouth of a customer who does not exist.

This page is not legal advice. It reports what specific published documents require, with a link on each, so you can run the pass yourself and recognise where to stop and involve a lawyer. Obligations differ by jurisdiction, industry and platform; scope here is US federal (FTC), the EU AI Act, UK self-regulation (ASA/CAP) and a few platform policies — not US state law, not sector regulators, not your client contracts. Where a document is silent, this page says the documents reviewed contain no such requirement, which is not the same as saying something is permitted.

When should you use this — and when should you not?

Run this on any piece where a model produced or reshaped copy, imagery or claims and the piece ships under your brand. It sits downstream of a general quality check like ai-draft-pre-publish-check and answers a different question: not "is this good", but "can we stand behind every assertion in it".

Escalate instead when you work in a regulated vertical (health, supplements, financial products, legal services, gambling, advertising to children); when the piece claims superiority over a named competitor; when it is a political, electoral or social-issue ad; when the output depicts a real, identifiable person photorealistically; or when you build and sell the AI system rather than use it — the EU's machine-readable marking duty falls on providers, not on the marketer using the tool (European Commission). If nothing in the piece makes an objective claim, skip the evidence matrix in Step 2 but not the fabrication sweep in Step 3.

What do you need before you start?

Input	Why it is needed
The final draft: headline, subheads, body, captions, alt text, meta title and description, CTA text	Claims hide outside body copy more often than inside it
The channels it ships to	Organic search, paid search and paid social differ
The jurisdictions your audience sits in	The US, EU and UK diverge sharply
Your evidence store: studies with method and sample, internal data with date ranges and definitions, certifications with issuer and expiry, customer permissions	Advertisers must have "a reasonable basis for advertising claims before they are disseminated" (FTC)
A provenance note: which model, what it produced, how deeply a human edited	The EU exemption for humanly reviewed text turns on exactly this (European Commission)
The name of whoever signs off	Every document reviewed assumes an accountable person, not a tool

How do you do it, step by step?

One rule governs every step: the model finds and formats candidate problems, and never decides that a claim is supported. Asked "is this claim substantiated?", a model produces confident, well-formatted support for things that do not exist. Verification stays outside the model, in a named human's hands.

Step 1 — what does this piece actually claim?

Separate objective claims from opinion and obvious puffery, mechanically, before any judgement.

You are extracting claims for a pre-publication review. Do not evaluate truth.
Do not improve or add claims. Extract only.

DRAFT:
[PASTE FULL DRAFT — headline, subheads, body, image captions, alt text,
meta title, meta description, CTA text]

Extract every statement a reader could reasonably read as an objective,
verifiable assertion, including implied claims — takeaways a reasonable reader
would form even though the text does not say them outright.

Output: | # | Verbatim text | Express or implied | Claim type | Evidence needed |
Claim type: performance | comparative | quantitative | testimonial | endorsement
| credential | pricing | safety | environmental | ai-capability | other

Rules:
- Two claims in one sentence produce two rows.
- For an implied claim, write the takeaway in your own words prefixed
  "[IMPLIED]" and quote the text that generates it.
- Flag any statistic, date, price, ranking, award, named study or named person.
- Do not state whether any claim is true or supported.
- End with the row count and the count of flagged specifics.

A full landing page usually yields fifteen to forty rows; fewer than about eight means the model under-extracted, so rerun asking specifically for implied claims. They matter because the requirement covers "express and implied claims, however conveyed, that make objective assertions" (FTC), and implication is what AI copy manufactures: "trusted by industry leaders" beside a logo strip, "clinically" used as an intensifier.

Step 2 — what evidence do you actually hold?

Go claim by claim against your evidence store. Human-led; the model formats and challenges.

Find MISMATCHES between claims and held evidence. Do not justify claims.

CLAIM INVENTORY: [PASTE STEP 1 TABLE]

EVIDENCE WE HOLD (this list is exhaustive — treat anything not listed as NOT
HELD; do not infer, recall, or use outside knowledge):
[LIST each item: name, date, who produced it, sample, what it measured]

Output: | # | Claim | Evidence ID | Verdict | Gap | Safe rewrite |

Verdict is one of:
  SUPPORTED       listed evidence supports the claim as worded
  OVERSTATED      evidence is narrower than the claim (population, conditions,
                  timeframe or magnitude)
  UNSUPPORTED     no listed evidence addresses this claim
  LEVEL-MISMATCH  wording asserts a level of proof we do not have
                  ("studies show", "clinically proven", "#1", "guaranteed")

HARD RULES:
- You have no knowledge outside the evidence list. Not covered there means
  UNSUPPORTED. Never fill a gap from memory.
- Evidence that is dated, from another product version, or that measured a
  different metric is OVERSTATED, not SUPPORTED.
- Give a narrower "Safe rewrite" for every OVERSTATED row.

End with counts per verdict.

LEVEL-MISMATCH is the verdict worth watching. Where an ad makes an express substantiation claim — "tests prove", "doctors recommend", "studies show" — "the Commission expects the firm to have at least the advertised level of substantiation" (FTC). Models reach for those phrases as rhetoric, so the phrase creates the obligation even when the underlying fact is fine.

Step 3 — what did the model make up?

This is the step that does not exist in a pre-AI checklist.

Adversarial review. Assume this draft was AI-assisted and that some specifics
are fabricated. Surface everything that must be verified externally before
publication.

DRAFT: [PASTE FULL DRAFT]

List every item in these categories, with verbatim text and location:
1. statistics, percentages, counts, prices, growth rates, rankings
2. named studies, reports, publications, institutions, analysts
3. quotes, testimonials, reviews, "one customer told us", composite customers
4. awards, certifications, standards, partner status, "certified", "official"
5. superlatives and comparatives: "fastest", "only", "leading", "#1"
6. dates, versions, prices, availability
7. claims about what our AI does and how accurate it is
8. regulatory assertions: "compliant with", "approved by", "meets X"

Output: | Item | Category | Verbatim text | Location | Verification needed |

Then answer: which of these did you generate or reconstruct rather than find in
my input, and which people, companies or publications named here cannot be
confirmed from my input alone?

Do NOT verify anything. Do NOT supply citations. Output the to-verify list only.

Every row goes to a human and a primary source; the method for the numeric rows is spot-fabricated-stats. Category three bites hardest. The FTC's rule on consumer reviews and testimonials makes it an unfair or deceptive act to write, create or sell a testimonial that materially misrepresents "that the reviewer or testimonialist exists", that they used the product, or the nature of their experience — and extends that to disseminating testimonials the business "knew or should have known" misrepresented those things (16 CFR Part 465). An invented customer quote sits inside that, and so does a composite assembled from real feedback.

Step 4 — what must be disclosed, and where?

A lookup, not a judgement. Have the model classify the piece and stop there: does it contain generated or manipulated image, audio or video depicting a realistic person or event, and is that person real and identifiable; is any generated text published to inform the public rather than to promote your product; did a named human substantively review it, and does someone hold editorial responsibility; is it a social-issue, electoral or political ad. Require the deciding text quoted for each answer, and forbid recommendations. Then read the table yourself.

Situation	What the document says	Source
US federal, generic AI-written copy	The FTC materials reviewed contain no general AI-disclosure requirement; the duties attach to substantiation and to misrepresented endorsements, not to AI authorship.	FTC
US federal, AI-generated testimonial or insider endorsement	Creating or disseminating testimonials that materially misrepresent that the reviewer exists, used the product, or had the described experience is prohibited. An officer or manager reviewing their own business needs "a clear and conspicuous disclosure of the officer's or manager's material relationship to the business" (§465.5). Under the endorsement Guides the duty to disclose arises where a connection "might materially affect the weight or credibility of the endorsement, and that connection is not reasonably expected by the audience" (§255.5) — not for every connection.	16 CFR 465, 16 CFR 255
EU, ordinary product-marketing copy	The deployer text-labelling duty attaches to text published to inform the public on matters of public interest — the Commission lists politics, public administration, justice, fundamental rights, public security, public health, environmental protection, consumer safety and "any economic, financial, political, scientific, or cultural developments that may be relevant subject of public debate". Whether commercial copy ever falls inside that list is not resolved by any source reviewed here.	European Commission
EU, text a human has edited	Text that "has undergone human review or editorial control" does not need to be labelled, and the Commission defines those as two separate routes: human review is "the deliberate examination of the substance of the content by one or more natural persons possessing relevant knowledge and professional judgement", while editorial control is exercised by "a responsible editorial entity (e.g. an editor-in-chief)" having "the authority to approve, alter or reject the substance of the text based on substantive grounds". Editorial responsibility is a third term: a person holds "the ultimate legal responsibility over the publication of the content". "Superficial, solely formal, or procedural checks (e.g. spell-checking or grammatical correction) are not considered to be human review or editorial control."	European Commission
EU, machine-readable marking	Addressed to providers, who "must ensure that AI-generated or manipulated content are marked in a machine-readable format" — not to the marketer using a commercial tool. Article 50 applies "as from 2 August 2026", with a grace period to "2 December 2026" for marking systems already on the market.	European Commission
UK	"There is no blanket legal requirement in the UK to disclose the use of AI in ads." The frequently quoted expectation that transparency applies where AI use "features prominently in an ad and is unlikely to be obvious to consumers" is not the regulator's rule: the ASA article reports it as one of twelve guiding principles produced in 2023 by two industry bodies, ISBA and the IPA.	ASA/CAP
Google organic search	No requirement. "Sharing information about how a piece of content was created can help give your readers more context", added "in a way that makes sense for your audience".	Google
Meta social-issue, electoral or political ads	Disclosure required where the ad contains photorealistic image, video or realistic audio "created or edited using third-party generative AI tools" in the specified ways; undisclosed ads are rejected.	Meta

Two limits, stated rather than papered over. The third row is the most consequential gap here: no source reviewed draws the line between commercial marketing copy and text informing the public on matters of public interest, and a sustainability claim or a health-adjacent product page plausibly brushes the Commission's list. This page does not resolve that, and neither should you alone. Second, US state statutes on AI disclosure and synthetic media were not researched, so treat the US rows as FTC coverage only.

Step 5 — will the platforms accept it?

Platform rules are contractual and ranking consequences, separate from regulators, and they are where per-page thinking fails. Google's spam policies define scaled content abuse as "when many pages are generated for the primary purpose of manipulating search rankings and not helping users", the first listed example being "using generative AI tools or other similar tools to generate many pages without adding value for users" (Google, updated 2026-05-15). The trigger is volume and intent, not authorship: the separate guidance on AI content sets the standard at meeting the Search Essentials and spam policies rather than at avoiding AI (Google).

So ask about the portfolio, not the page: how many similar pieces you shipped in the last ninety days, whether they came from one template, and what this one adds that is not already on your site or on the pages already ranking. If the honest answer is "nothing", it fails here however clean Steps 1 to 4 were. For paid placements, read the ad policy of the surface — Google Ads prohibits claims that "entice the user with an improbable result (even if this result is possible)" and states that "manipulating media to deceive, defraud, or mislead others is not allowed" (Google Ads). The section reviewed contains no clause specific to AI-generated content; check the policies for your ad type rather than reading that silence as clearance.

Step 6 — what record survives publication?

Produce a dated artifact and file it with the asset, not in a chat transcript: every claim published and the evidence held for it, every claim cut or narrowed and why, every externally verified item with source, verifier and date, the disclosure decision per jurisdiction and channel with its reason, every unresolved item, and a named sign-off. This is what shows a reasonable basis existed before dissemination (FTC) when someone asks six months later. Add a review date: nothing in the pass re-triggers when a study is superseded or a certification expires.

What does a good result look like?

Three rows from a finished verdict table, showing what the pass does to real copy:

Line in the draft	Verdict	What happened
"Studies show teams cut review time by 40%."	LEVEL-MISMATCH	We hold one internal sample of twelve accounts from Q1, not a study. Rewritten to "in our own Q1 sample of twelve accounts, median review time fell 40%", dataset filed as evidence.
"Trusted by industry leaders."	UNSUPPORTED, implied	The implied claim is endorsement by the logos beside it. Two of the four had no permission on file. Line and logos cut.
"It paid for itself in a month" — a marketing director at a mid-size retailer	Blocked	No identified person said this; the model composed it from a support ticket. Replaced with a named customer's approved quote.

The other half of a good result is the artifact: a colleague who was not in the room can open one dated file and reconstruct why every surviving line survived.

How do you know the output is good?

• Every claim in the inventory carries a SUPPORTED verdict backed by a named, dated evidence artifact a third party could be shown; no OVERSTATED, UNSUPPORTED or LEVEL-MISMATCH rows remain.
• Every item on the verification worklist has a named human verifier, a primary-source reference and a date; no item is marked verified by the AI system that produced the draft.
• Every testimonial, review or first-person customer quote traces to a real identified person with a retained permission record; anything that cannot is cut.
• The classification table is complete for every jurisdiction and channel on the shipping list, and each row resolves to an explicit decision naming the document relied on — including rows that resolve to "no requirement found in the documents reviewed".
• No platform criterion is marked FAIL, and the question "what does this page add that is not already available" has an answer other than "nothing".
• A dated sign-off record exists before the publish date, names the accountable human, and lists zero unresolved items.

Hand the checks to someone who did not write the piece; each is answerable yes or no from the record alone. Three things decide whether the pass was real rather than performed.

Evidence has to be an artifact, not an assertion. "The model said we have support for this" is not support. For each surviving claim someone should be able to produce a named, dated document — a study with its sample and method, an internal dataset with a date range and a definition of what it measured, a certificate with issuer and expiry — and hand it to a stranger. Where that hand-off is impossible, the row is not supported however the table is coloured.

Nothing is verified by the system that wrote it. Read down the verifier column: if a row reads "verified" and the verifier is the model, that row is unverified. This is the check most worth auditing, because it is the one that passes silently.

"No requirement found" is a decision and needs its source too. Recording which document you read, and on what date, makes such a row defensible rather than an assumption — particularly where obligations start on a known date, as the EU transparency rules do (European Commission).

For a repeatable numeric read on draft quality alongside this pass, use a different instrument: score-ai-output.

Where does this usually break?

1. The model verifies its own output. The highest-frequency failure by a distance. Asked to find a citation for a claim it wrote, a model produces a plausible one that does not exist. Steps 2 and 3 are built so it can only flag and format.
2. A label is treated as a cure for a weak claim. ASA/CAP puts it directly: "disclosure alone is very unlikely to mitigate the harm caused by a fundamentally misleading message" (ASA/CAP).
3. Implied claims are missed. Extraction that catches only literal sentences misses most of the exposure, because the requirement reaches implied claims "however conveyed" (FTC).
4. The EU rules get read backwards. Two mistakes travel together: assuming the AI Act requires labelling all AI marketing copy, and assuming the machine-readable marking duty is yours. The text duty is limited to public-interest text and falls away where the text has had human review or editorial control; marking sits with providers (European Commission). Over-reading teaches a team to treat the framework as noise; under-reading leaves the unsettled boundary unexamined.
5. Per-page compliance with portfolio-level violation. A team can pass on every page and still ship four hundred near-identical pages a month — the pattern the scaled content abuse policy describes (Google). Only the Step 5 portfolio questions catch it.
6. Jurisdiction drift after publication. A page cleared for a US audience gets translated or paid-promoted into the EU and the Step 4 classification no longer describes reality. Re-open the record when the channel or jurisdiction list changes, not only when the copy does.
7. The provenance note is wrong. It is self-reported, and self-reporting here is soft: in a vendor-commissioned survey conducted by Savanta for Optimizely among 2,003 marketing leaders across seven markets in May–June 2026, 30% said they frequently or always pass off AI-generated work as their own (Optimizely). One vendor source, no published questionnaire or margin of error — a reason to check your own provenance line, not a measured rate.
8. The record lives in a chat window. Transcripts rotate out and cannot be retrieved per asset. If the basis existed but cannot be produced, operationally it did not exist.

What else do people ask?

What counts as "human review or editorial control"?

The Commission's guidance gives a floor, not a threshold, and it names two routes rather than one test. Human review is "the deliberate examination of the substance of the content by one or more natural persons possessing relevant knowledge and professional judgement"; editorial control is what "a responsible editorial entity (e.g. an editor-in-chief)" exercises with "the authority to approve, alter or reject the substance of the text based on substantive grounds". Editorial responsibility is separate again — a person holding "the ultimate legal responsibility over the publication of the content" (European Commission). No source reviewed sets a minimum edit fraction or a record format, so keep the Step 6 record without treating its existence as proof the exemption applies. Editing standards are a separate craft — edit-ai-draft, leaning on brand-voice-doc-for-ai.