Who Is Liable for AI-Generated Ads? The Business That Runs Them

The business that publishes an AI-generated ad answers for every claim in it, whoever or whatever drafted the words. Singapore's consumer protection law, the advertising industry's own code, the US Federal Trade Commission (FTC) and EU consumer law test the claim, not the author, and each can ask the advertiser for its evidence. For marketing and compliance leads publishing AI copy at volume, the control that works is a gate that checks every factual claim against evidence on file before anything ships.

Brand voice drift is embarrassing. An unsubstantiated claim is a legal problem. On 23 March 2026 the Advertising Standards Authority of Singapore (ASAS) reported that seven of the 379 pieces of feedback it received in 2025 involved generative AI, more than the three it received in 2023 and 2024 combined. The numbers are small. The direction of the guidance is not.

Who is liable when AI writes a misleading ad?

The business that publishes the ad is liable for an AI-written claim. ASAS put it plainly on 23 March 2026: "False and misleading advertisement content is unacceptable, regardless of the tools used or the disclosures added." Singapore's consumer law, the US FTC and EU consumer law reach the same place by different routes.

This is general information, not legal advice, from a firm that builds claim controls. Singapore's binding law comes first. The Consumer Protection (Fair Trading) Act 2003 (the CPFTA, enforced by the Competition and Consumer Commission of Singapore, or CCS) makes it an unfair practice for a supplier, in a consumer transaction, "to do or say anything, or omit to do or say anything, if as a result a consumer might reasonably be deceived or misled" or "to make a false claim" (section 4). A supplier includes any business that promotes goods or services. Section 4 tests what was said, never who or what drafted it.

The industry code casts a wider net. The Singapore Code of Advertising Practice (SCAP, 3rd edition, February 2008, the version ASAS still publishes as current) says "Responsibility for observing the Code rests primarily with the advertiser," and that it "also applies to any advertising agency or medium involved". Part II 3.4 reads almost as if written for AI: content that "originates wholly or in part from other sources is not an excuse for non-compliance". SCAP is voluntary self-regulation, not law. Its teeth are practical: ASAS can ask media owners to support its decisions and can publish offenders' names.

In the US, the FTC's Endorsement Guides (16 CFR 255.1(d)) say "An advertiser may be liable for a deceptive endorsement even when the endorser is not liable." Agencies and endorsers can share the exposure. The advertiser never drops out of it.

The tool vendor is leaving the frame

Rytr, an AI writing tool whose users could generate fake reviews, was one of the FTC's Operation AI Comply cases in September 2024. On 22 December 2025 the FTC reopened and set aside that order. Christopher Mufarrige, Director of its Bureau of Consumer Protection, said: "Condemning a technology or service simply because it potentially could be used in a problematic manner is inconsistent with the law and ordered liberty."

The rule binding the publisher did not move. Under 16 CFR 465.2, the fake-review rule the FTC announced on 14 August 2024, a business may not disseminate a testimonial it "knew or should have known" misrepresents that the testimonialist exists. Our reading: the US regulator has stepped back from the tool and kept its grip on whoever publishes. An indemnity does not change that.

Does an "AI-generated" label protect you?

No. An AI disclosure answers "who made this", while advertising law asks "is it true". ASAS said in March 2026 that tools and disclosures do not "absolve marketers of the responsibility" to comply, and that "indiscriminate and vague disclosures of AI usage are not useful, and they hinder rather than aid consumer discernment." A labelled false claim is still a false claim.

Article 50 of Regulation (EU) 2024/1689 applies from 2 August 2026 and is a transparency rule. Paragraph 2 requires providers of generative AI systems (the model and tool makers) to mark outputs as machine-generated. Regulation (EU) 2026/1744, the Digital Omnibus on AI, in force since 27 July 2026, gives providers whose systems were on the market before 2 August 2026 until 2 December 2026 to comply with that paragraph.

Paragraph 4 is the one that reaches you. A deployer of an AI system that generates a deep fake image, audio or video "shall disclose that the content has been artificially generated or manipulated." The matching text duty covers only text "published with the purpose of informing the public on matters of public interest", and falls away where a person holds editorial responsibility. Ordinary ad copy is outside it.

Here is the trap. Article 3(60) defines a deep fake as content that "resembles existing persons, objects, places, entities or events". A synthetic "customer" who resembles nobody may need no EU label, and still mislead. Article 50(6) says it is "without prejudice to other transparency obligations", and the Unfair Commercial Practices Directive still governs your claims. Its Article 12 lets courts and authorities require the trader "to furnish evidence as to the accuracy of factual claims" where that appears appropriate, and treat a claim as inaccurate if the evidence is not produced. The burden can land on you.

ASAS's own example shows why. A telemedicine company used generative AI to make a video of a man giving a first-person testimonial; complainants thought he was real, and the advertiser ceased the ad. SCAP covered this in 2008: testimonials "based on fictitious characters" should not be "framed so as to give the impression that real people are involved." A corner tag would not have cured that impression.

Why AI-written copy carries more unchecked claims

Language models produce plausible specifics with no source behind them: a percentage, an award, a "No. 1", a customer quote. Plausibility is not evidence, and evidence is what every regime asks for. The FTC's substantiation policy (1984) requires that advertisers and ad agencies have "a reasonable basis for advertising claims before they are disseminated."

SCAP Part III 1.1 requires claims about "objectively ascertainable facts" to be "capable of substantiation", held "ready for immediate production without delay to ASAS upon request". Ready means on file now. Not reconstructed after the complaint.

A simple test for what counts as a claim: if a reader could check it, it needs evidence: a number, a ranking or comparison, a certification or award, a price or deadline, a person's words, a physical attribute. Everything else is opinion and must read as opinion. Screen every batch for four patterns:

• Borrowed authority. The model adds "award-winning", "certified" or "trusted by leading brands" because similar copy said so. In November 2025 the CCS warned a mattress seller, in a case with no AI involved, over badges implying official standards that do not exist and a 4.5-star rating built mainly on reviews the business wrote itself.
• Phantom numbers. A made-up line such as "cuts admin time by 40%" appears because the prompt said "make it punchy". The CPFTA's Second Schedule lists representing that services have "performance characteristics ... or benefits that they do not have" as an unfair practice.
• Opinion sliding into fact. "Customers love our fastest-ever delivery" mixes a feeling with a measurable claim. SCAP Part III 4.1 allows opinion only where it is clearly opinion and does not mislead about facts.
• Translation drift. A safe English line becomes a firmer promise in a language nobody on the team reads.

What evidence does each type of AI-written claim need?

Different claims carry different legal weight, so route them differently. The table maps the claim types AI copy produces most often to the evidence to hold before publishing and who signs. The evidence and approver columns are our design recommendation; the last column names the provision behind it.

Financial, health and investment claims have their own regulators and stricter sector advertising rules. Give them a separate track with sector sign-off.

The price row is where AI schedules cause trouble. The CCS Guidelines on Price Transparency, in effect since 1 November 2020, say time-limited discounts must state the period "clearly and accurately", and the CPFTA lists a "stated period" discount that runs substantially longer as an unfair practice (Second Schedule para 20).

The environmental row has a date for anyone selling into Europe. Directive (EU) 2024/825 tells Member States to apply its rules from 27 September 2026 through national law, and adds "a generic environmental claim for which the trader is not able to demonstrate recognised excellent environmental performance" to the EU list of practices banned in all circumstances. "Eco-friendly" is exactly what a model adds unprompted.

How a claims-evidence gate works

A claims-evidence gate is a pre-publication check that stops any AI-written claim without evidence on file. It pulls each factual claim out of the copy, matches it to an evidence register and routes it to a named approver by risk. The CCS Guide on Quality-Related Claims (October 2025, guidance under the CPFTA) says suppliers should take "reasonable steps to verify" a claim before communicating it.

The evidence register is the part most teams lack. Give each approved claim one row: claim ID, exact approved wording, permitted variants, source document, date of the evidence, markets, owner, approver tier and review date. It turns substantiation into a lookup and gives the AI something to draft from. The gate still checks the output, because models add facts nobody gave them; any fact without a claim ID is a defect.

The review date matters as much as the evidence. CCS asks suppliers to "periodically review" claims (guide para 3.9), and SCAP bars testimonials that are "obsolete or otherwise no longer applicable". A promotion that ended on Friday is a false claim on Saturday if a scheduled variant still runs it. That is the most likely failure in a working gate: nobody wires the end date to the ad scheduler.

Illustrative example: a retailer's AI tool produces 400 ad variants in a month. Extraction finds 1,200 claim instances, which collapse to 30 distinct claims. Twenty-six match register rows that are approved and current, so they pass. Four do not: "award-winning" (no award on file), "Singapore's most trusted" (no survey), "delivery in 2 hours" (the service level is same day) and a customer quote nobody can trace. Those four are rewritten as opinion or cut. The compliance lead reviews four claims, not 400 drafts. At an assumed 10 minutes a claim, that is 40 minutes of sign-off instead of days of line edits. The numbers are made up; the ratio is the point.

CCS also says that when it investigates it will consider "whether the supplier has made genuine efforts and taken reasonable steps to verify the accuracy of any information that it relied on", weighed against "the business's scale and resources" (para 2.3). A gate log of claim, source and approver records those steps. It is not a safe harbour. It is evidence you acted before the complaint arrived.

Who owns which part of the gate?

In a claims-evidence gate, marketing approves claims instead of drafts and compliance owns the stop rule. The CTO builds what makes that rule enforceable. CCS weighs a business's "scale and resources" (October 2025 guide) when judging whether its checks were reasonable, so a small team can run a lighter version with the same logic.

Marketing lead. You approve claims, not drafts, and brief the tool to build from register claims only.

CTO or technical lead. Claim extraction can be a model pass with a strict output schema, backed by pattern rules for digits, superlatives, currency and words like "certified" or "free". What matters is enforcement: the publishing step refuses any variant whose claim IDs are not all approved and current, and it re-checks review dates at send time, not at approval time.

Compliance and legal. You own the stop rule, the high-risk tiers and the log. Recent FTC matters are claims cases, not AI-copy cases. On 24 March 2026 the FTC filed a proposed settlement with Air AI over earnings and performance claims for an AI business-opportunity service: an $18 million judgment, suspended except for $50,000, subject to court approval. On 27 August 2026 it finalised orders requiring Cox Media Group, MindSift and 1010 Digital Works to pay a combined $930,000 over an "Active Listening" AI ad-targeting service that did not use voice data at all. Both punish a claim nobody could stand behind. Our earlier piece on brand voice drift and agentic marketing ROI covers the tone side; the gate covers the legal side.

Owner or MD. One test: if CCS asked tomorrow for the evidence behind your three biggest claims, the team should produce it that afternoon.

Where Origin Pi stands

Regulators test the claim and ignore the tool. ASAS logged seven generative AI cases in 2025 and says it takes "a consistent approach that is independent of the tools used". So the useful response is narrow. Treat every factual claim as a governed record with a source and a review date, and let the AI write freely around it.

We design and build that layer for marketing teams. The agent drafts from an approved evidence register, a claim extractor checks every output, and a confirm step routes anything new to the right approver with bounded permissions, so the agent cannot publish what nobody signed. Every decision lands in an audit trail you can hand to a regulator. It is the same pattern behind our marketing AI work: agents propose, humans approve.

Your first move takes an afternoon. Pull your five most-used ad claims, find the document behind each one, and write down who would sign it. Any claim without a document is the one to fix before your next AI campaign runs. For running those campaigns, see performance marketing at leapbuzz, our marketing consultancy brand.