Singapore AI Advertising Rules: What ASAS Expects From AI Video

Singapore AI Advertising Rules Start With a Self-Regulatory Code

Singapore AI advertising rules do not live in a dedicated AI statute. They live inside the Singapore Code of Advertising Practice, which the Advertising Standards Authority of Singapore administers as industry self-regulation, and which holds that every advertisement should be legal, decent, honest and truthful. For a brand video team, the operative test is therefore whether the finished spot misleads a consumer, not whether it carries an AI badge.

That distinction changes where your review effort goes. A codified labelling mandate gives you a binary task: apply the label, keep the record, move on. A four-word honesty premise gives you a judgement task on every frame, because the question is what the audience reasonably takes away from the ad.

Teams running multi-market work tend to arrive in Singapore with a labelling reflex built elsewhere. Compare it with {{link}}, where the enforcement conversation has been about who stays accountable for a generated depiction, and the pattern becomes clear: the label is the cheapest part of compliance and rarely the part that gets you into trouble.

The practical consequence is that your AI pipeline inherits the same evidentiary burden as a live-action shoot. If the ad depicts a product doing something, someone has to be able to show it does that. If it depicts a person endorsing something, the endorsement has to be real.

Compare it with the UK's tightening rules on AI-generated ads, where the enforcement conversation has been about who stays accountable for a generated depiction, and the pattern becomes clear: the label is the cheapest part of compliance and rarely the part that gets you into trouble.

What ASAS Has Actually Flagged

The volume is small and the direction is what matters. ASAS received 379 pieces of advertisement feedback in 2025, of which seven concerned advertisements involving generative AI. That was more than double the three pieces received across 2023 and 2024 combined. In the first half of 2026 the AI-related count was two, roughly flat against the prior year.

Read those numbers correctly. Seven complaints out of 379 is not a crisis, and anyone treating it as proof of a crackdown is overreading. ASAS has already called on industry bodies to develop guidelines on the use of AI in advertisements, explicitly not to restrict the practice but to ensure the resulting advertisements do not mislead consumers.

That framing tells you where the binding standard will actually come from. In a self-regulatory system, sector guides written by trade bodies become the reference a complaint gets measured against.

For a production team the near-term action is documentary rather than creative. Keep the prompt history, the model and version used, the reference assets, the approval trail and the substantiation file for every claim the video makes. None of that is required by a statute today. All of it is what you will be asked for if a spot is challenged, and reconstructing it six months after delivery is close to impossible.

Why Disclosure Alone Does Not Clear the Bar

ASAS has been unusually direct on this point. Its chairman has said that disclosure of the use of AI alone does not whitewash the issue, and has described indiscriminate and vague disclosures of AI use as not useful. A blanket badge on the end card is not a shield. If the ad misleads, the badge travels with the misleading ad.

The failure modes are specific and worth naming. Generated product geometry that does not match the shipping SKU. Invented interface screens that show features not in the release. A generated performance that reads as a customer testimonial when no customer was involved. Each of these is a substantiation problem dressed as a rendering problem, which is why {{link}} matters more than the disclosure wording.

Better disclosure is specific disclosure. Say what was synthesised and where, in language a viewer can act on: the presenter is synthetic, the environment is generated, the product footage is real.

Machine-readable provenance is the complement, not the substitute. The C2PA Content Credentials standard gives publishers, creators and consumers an open technical way to establish the origin and edits of a digital asset, functioning as a history that travels with the file.

Each of these is a substantiation problem dressed as a rendering problem, which is why substantiating what a generated product shot depicts matters more than the disclosure wording.

The Economics Driving the Surge

The demand curve explains the regulatory attention. One Singapore AI content agency reported enquiries growing four to five times against the late-2025-to-April-2026 period. An AI livestream avatar provider reported enquiries rising 20 to 30 per cent every month since the start of 2026. This is not an experimental pilot phase; it is a procurement shift.

The avatar arithmetic is blunt. An ordinary human livestream host in Singapore costs roughly S$3,000 to S$4,000 a month and delivers about four to five hours a day. An avatar runs about S$458 a month and can stream continuously for twenty-four hours. On a pure cost-per-broadcast-hour basis that is not a close comparison, which is exactly why the format is spreading through commerce categories before anyone has settled the disclosure question. If you are weighing that trade, the operating detail sits in {{link}}.

Generated video is not automatically cheap, though. The same agency puts a simple one-minute video for a smaller business at around S$1,000, while a large multinational wanting complex frame-by-frame editing pays closer to S$20,000. The saving shows up in volume and iteration, not in a lower unit price for finished, brand-grade work.

Volume is where compliance quietly breaks. Agoda reported that generative AI took its video adaptations from 64 to 720 per month. That is an eleven-fold increase in assets flowing toward audiences, and human review capacity does not multiply on the same curve.

If you are weighing that trade, the operating detail sits in putting a synthetic presenter in front of a brand.

Where Singapore Brands Draw Their Own Red Lines

The more instructive material is not the regulator's; it is the operators'. Circles.Life, a Singapore telco using AI across copywriting, image creation and video since 2025, puts every AI-made asset through the same quality control as any other campaign material: review for accuracy, brand consistency, creative and technical quality, audience suitability, and copyright and usage compliance before approval. AI does not make final creative decisions and does not publish independently.

That team also published its reversal conditions, which is rarer and more useful than a policy statement. It would scale back further if AI-assisted work stopped meeting quality or brand standards, introduced factual or cultural inaccuracies, or if customer feedback showed the work was reducing relevance or trust.

The brand-equity risk is unevenly distributed. A Nanyang Technological University researcher studying digital advertising has argued that the backlash lands hardest on large, high-end brands, because consumers read a generated ad as lacking effort and therefore read the brand as less authentic. A luxury house sells an aspirational lifestyle whose credibility rests on evident craft, so visible automation reads as a withdrawal of investment in the relationship. This is the same dynamic described in {{link}}.

High-trust categories should treat some formats as out of scope entirely. A tuition centre filming customer testimonials, a clinic showing outcomes, a financial adviser demonstrating a product: in each case the persuasive value comes from the viewer believing a real person is speaking. Synthesising that performance does not just risk a complaint, it removes the reason the asset worked.

This is the same dynamic described in the quality ceiling on cheap synthetic advertising.

A Pre-Flight Checklist for Singapore Campaigns

Run five gates before a generated spot ships into a Singapore buy. First, a claim register: list every factual assertion the video makes, including implied ones, and attach the substantiation. Second, a depiction audit: confirm every product, interface, price and result on screen matches something that exists and is available. Third, a presenter check: if a human likeness appears, record its origin, the consent chain and whether it is synthetic.

Fourth, write the disclosure to be specific rather than decorative, naming which elements were generated, and place it where the viewer meets the claim rather than only on the end card. Fifth, retain the provenance record: prompts, model and version, reference assets, approvals and the export chain, stored with the asset rather than in someone's local project folder. Teams running the same creative across several jurisdictions should reconcile this against {{link}} instead of maintaining five disconnected review habits.

Add one localisation gate that generation makes newly necessary. Singapore campaigns routinely run across several languages and cultural registers, and generated dialogue, on-screen lettering and setting details are where cultural inaccuracy enters unnoticed. A native-speaker review of every localised variant, not just the master, catches the errors that a model confidently produces and an English-first reviewer never sees.

None of this is heavy process, and it is deliberately independent of the regulatory state of play. Singapore is likely to get sector guidance from industry bodies rather than a single AI advertising statute, which means the standard will arrive in pieces and at different speeds by category. A team that can already show what it generated, what it claimed and who approved it will clear whichever version lands. A team that cannot will be assembling that evidence under deadline, with the campaign already live.

Teams running the same creative across several jurisdictions should reconcile this against one cross-market disclosure matrix instead of maintaining five disconnected review habits.