A deliberate non-decision
On August 5, 2026, Singapore's Ministry of Digital Development and Information (MDDI) gave Parliament a carefully hedged answer to a straightforward question. MP Alex Yeo had asked whether the government would require indicative labelling of AI-generated or digitally manipulated content shown to Singapore users — the same obligation the European Union imposes under Article 50 of its AI Act, which itself had come into force just three days earlier, on August 2, 2026. MDDI's answer was, in effect: not yet. It pointed instead to two existing instruments — the Infocomm Media Development Authority's (IMDA) Codes of Practice for Online Safety, which require designated platforms to "mitigate Singapore users' exposure to harmful content, including those generated by AI," and the voluntary Transparency Guidelines for Generative AI Chatbots that IMDA published on July 20, 2026. The Ministry said it would keep "monitoring international developments," including the maturity of technical standards like watermarking, before deciding whether a mandatory labelling regime is warranted.
What the EU actually requires
Article 50 is broader than a single deepfake rule. It obliges providers to mark AI-generated or manipulated output in machine-readable form, requires chatbot operators to disclose that users are talking to a machine "unless this is obvious," and forces clear disclosure whenever deepfake content is first shown to a viewer. AI-generated text published on matters of public interest — politics, health, public administration — must be labelled unless a human has exercised editorial control over it. Non-compliance can draw fines of up to €15 million or 3% of global turnover, whichever is higher, with the marking duty for content already in circulation extending to December 2, 2026. It is, by design, a horizontal rule: it applies to the technology (AI generation) rather than to a specific category of harm.
Singapore's alternative: regulate the harm, not the technology
Singapore's existing framework runs the opposite direction. The Online Safety Code for social media services, in force since July 18, 2023, and its 2025 counterpart for app distribution services, target defined harm categories — sexual content, violence, self-harm, cyberbullying, content facilitating vice — and leave it to designated platforms to build the systems that catch AI-generated instances of those harms. The GenAI chatbot guidelines add a second layer: a voluntary "Chatbot Info Card" disclosing a deployer's capabilities, limitations, safeguards and data practices, which IMDA has billed as among the first frameworks of its kind globally.
A second, more recent reply fills in the picture. On September 10, 2026, MDDI told Parliament that enhanced Codes of Practice under the Online Criminal Harms Act — issued August 17, 2026 — now require designated online services to verify advertiser identities and promptly remove suspected scam advertisements, addressing AI-generated fake endorsements directly at the fraud layer. Separately, the Online Safety Commission currently grants relief for five defined harms — intimate image abuse, image-based child abuse, doxxing, harassment, stalking — with "inauthentic material abuse" recognised as a future sixth category under the Online Safety (Relief and Accountability) Act, but not yet commenced. Layered together, this is a government building a lattice of harm-specific instruments rather than one blanket content-labelling mandate.
The case for the EU's approach, stated fairly
The strongest argument for mandatory labelling is that it is a floor, not a ceiling: a universal machine-readable tag doesn't require regulators to anticipate every harm AI content might cause, because the disclosure travels with the content regardless of context. It also gives ordinary users — not just platforms with harm-detection systems — a way to tell synthetic media from authentic media at the point of consumption, which matters most exactly when a deepfake is novel enough that no existing harm category has caught up to it. Given Singapore's own recent history — a 2024 law banning digitally manipulated depictions of election candidates, and a documented AI-generated disinformation campaign impersonating Prime Minister Lawrence Wong in early 2026 — the case that synthetic content is already a live problem, not a hypothetical one, is not hard to make.
Why the targeted route is still the better bet
But a horizontal labelling mandate is a blunt instrument for a fast-moving technology. Watermarking and provenance standards remain immature and contested — the EU's own Code of Practice on AI-content transparency was still in draft form as of the Commission's December 2025 publication — which means mandating a technical standard now risks locking in an approach regulators may need to unwind within a year or two. The EU's carve-out for human-edited text also shows how much of the rule's bite depends on definitional line-drawing that will generate disputes rather than clarity. Singapore's harm-first approach, by contrast, lets enforcement track where AI-generated content actually does damage — scam ads, election deepfakes, child abuse material — without taxing the much larger volume of low-risk synthetic content (marketing images, satire, routine chatbot replies) with a compliance-grade labelling obligation. For a smaller regulator without the EU's enforcement bandwidth across 27 member states, concentrating capacity on defined harms, while continuing to watch how Article 50 plays out in practice, is the more proportionate call — provided MDDI actually revisits it once the EU's implementation experience matures, rather than treating "monitoring" as a permanent placeholder.
The open question
MDDI's reply leaves real ambiguity: it names no trigger, no review date, and no threshold at which "monitoring international developments" converts into action. That is defensible as a starting posture. It becomes a problem only if Singapore is still monitoring in 2028 while the EU's regime has already reshaped global platform defaults — at which point the absence of a decision becomes a decision by default.