Does a Singapore brand have to label its AI ads?

The question comes up in the first meeting now, usually phrased as a formality: do we have to say it was made with AI? For a brand sitting in Singapore the honest answer is that Singapore is the wrong place to look it up. There is no labelling law here. There is one across the strait, and another north of it, and your media plan almost certainly crosses both.

This is not a legal opinion and we are not lawyers. It is the map we work from when a campaign is going to run in more than one market, which is most of them, and the reason we ask where a film is going before we ask what it should look like.

What Singapore actually requires

The government set its position out in Parliament on 10 September 2026, answering a question on unauthorised AI‑generated likeness in advertising. Three frameworks were named, and none of them is a labelling mandate.

The Singapore Code of Advertising Practice, administered by ASAS, holds that advertising must be legal, decent, honest and truthful. Two parts of that bear directly on this work: advertisers must disclose the use of AI where not disclosing it could mislead, and must not depict a person or their property without consent. ASAS acts on complaints and can require an ad to be amended or withdrawn. An AI ad is treated as an ad.

Separately, Codes of Practice under the Online Criminal Harms Act were issued on 17 August 2026, aimed at ads that impersonate real people. Designated services have to verify who their advertisers are and take down suspected scam ads. And the Online Safety Commission, under the Online Safety (Relief and Accountability) Act, currently covers five harms — intimate image abuse, image‑based child abuse, doxxing, harassment and stalking. Inauthentic material abuse is named as a future category, to be commenced in due course. No date has been given.

“Disclose where not disclosing would mislead” is a test about the audience, not a sticker you apply at the end.

That distinction matters more than it sounds. A mandate tells you what to put on screen. A misleading‑ness test asks what the viewer was led to believe, which means the answer depends on the film. A fully synthetic product shot of a real product is not obviously misleading. A person who appears to be a real customer giving a real account of a real experience is a different thing, and that is the case Singapore has already seen.

The complaints are small and pointing one way

ASAS received 379 pieces of advertising feedback in 2025. Seven concerned generative AI. That is a small number and worth saying plainly rather than inflating — but it was more than the 2023 and 2024 totals combined.

The example that matters is a telemedicine company that used generative AI to produce a first‑person testimonial video which appeared to come from a real person. Complaints followed. The advertiser pulled the ad and ran an internal review. Nobody was fined. The ad still came down, which for a brand is usually the expensive part.

Then the campaign leaves the country

Here is where the Singapore answer stops being the useful one. The same film, on the same media plan, lands in markets that do have mandates, and have had them for over a year.

China has required labels on synthetic content since 1 September 2025, under the Measures for Labeling Artificial Intelligence‑Generated Content issued by the Cyberspace Administration of China on 14 March 2025, alongside the mandatory national standard GB 45438‑2025. It asks for two things at once. An explicit label — something a person can see or hear — on content that could mislead or confuse. And an implicit label: metadata carried inside the file identifying the service provider and a content ID.

South Korea's AI Basic Act, Act No. 20676, came into force on 22 January 2026. Article 31(2) requires that generative AI output be labelled as such, and the Enforcement Decree asks for a machine‑readable marker accompanied by at least one notice a human can actually read or hear. It is worth being precise about the penalty, because it is widely reported loosely: failing to label is not itself directly subject to an administrative fine. The fine provisions cover a different list, and a labelling failure reaches one only indirectly, by way of a corrective order that then goes unfollowed. The duty is real. The headline number attached to it usually is not.

India, Vietnam, Thailand and Taiwan have introduced disclosure duties of their own, on their own timetables. The detail varies enough that the only safe way to treat the region is market by market, and that a regional summary — including this one — is a starting point rather than an answer. We have written separately on what the platforms themselves require, which is a different set of rules again and moves faster than the law does.

What that means for the way a film gets made

Three things follow, and all of them are production decisions rather than legal ones.

The widest market sets the rule, not the home one. If a campaign touches China or Korea, the labelling question is already answered, and it was answered before the brief was written. Deciding it in Singapore and discovering it in Shanghai is the expensive order to do this in.

Metadata is not something you add afterwards. An implicit label lives inside the file. That is a decision about the pipeline — what tools touched the work, what provenance survived the round trip through a retouch and a grade and an export. A visible label can be added on the last day. The thing inside the file usually cannot.

Consent for a likeness has to exist on paper. Singapore's own code already says a person or their property cannot be depicted without consent, and that is the clause with teeth here — no AI mandate required. It applies whether the face was filmed, licensed or generated to resemble someone. The adjacent question — which models a brand is cleared to use at all — is settled before any of this becomes relevant.

The part nobody legislates

Every rule above is about disclosure, and disclosure is a creative problem long before it is a compliance one. A label dropped into the corner of a finished film reads as an apology. A film built in the knowledge that it will carry one reads as a choice. The work we are proudest of would survive the label, because none of it is pretending to be documentary evidence of something that did not happen.

That is the test we would apply before any of the above: if the audience were told exactly how this was made, does the film still work? If it does, the regulation is paperwork. If it does not, the regulation found a problem that was already there.

Where this came from

Singapore's position is taken from the MDDI reply in Parliament of 10 September 2026. The complaint figures and the telemedicine case are reported by Marketing‑Interactive. China's measures are summarised by Covington's Inside Privacy. The wider regional picture, which is where the India, Vietnam, Thailand and Taiwan positions come from, is BowerGroupAsia's Indo‑Pacific overview. Dates and article numbers were checked against those sources in November 2026 and will move.