A synthetic-image carve-out closes a real gap
On August 17, 2026, provisions of Singapore's Criminal Law (Miscellaneous Amendments) Act 2025 took effect making it a standalone offence to produce an intimate image of another person without consent — including images that are entirely AI-generated, with no underlying photo or video ever taken of the victim. The Ministry of Home Affairs confirmed the commencement date on August 14, and the general offence carries up to two years' imprisonment, a fine, or both; where the image depicts a person under 14, the offence carries mandatory imprisonment of up to two years plus a fine or caning (MHA, via AsiaOne, Aug 17, 2026).
The gap being closed is a real one. Singapore's existing intimate-image offence, Penal Code section 377BE, was written for images "altered" from a real photo or recording — the classic face-swap deepfake. Generative models can now produce a convincing intimate image of a named person from a text prompt alone, with no source photo manipulated at any point. Prosecutors briefed on the earlier law worried that defence counsel could argue such images fell outside the statute's wording entirely. The amendment (via clauses to section 377BE) closes that textual loophole directly.
The steelman: this is a narrow, well-targeted fix
The case for the law is straightforward and worth taking seriously before critiquing it. Non-consensual intimate imagery causes acute, documented harm regardless of whether the underlying image is "real" — the humiliation, reputational damage, and coercive leverage it hands abusers do not depend on authenticity. Senior Minister of State Sim Ann told Parliament the amendments aim "both to reduce the roadblocks to enforcement and to reduce the supply of such materials, which contributes to a vicious cycle of abuse" (MHA Second Reading speech). That is a legitimate and narrow rationale — the law targets production of a specific, identifiable harm category, not AI image generation broadly, and it tracks comparable moves already made or underway in South Korea and the UK, whose 2025 Data (Use and Access) Act created its own offence for creating sexually explicit deepfakes without consent (UK government press release).
Where the design earns scrutiny
Two features of the Singapore approach deserve closer attention. First, the offence turns on production, not distribution — a person can be criminally liable the moment an image is generated, before it is shared with anyone. That is defensible for genuinely synthetic material aimed at a real, identifiable victim, but it puts real weight on how "depicts another person" gets interpreted in borderline cases — stylized composites, satire, or images where resemblance is contested. The statute does not appear to carve out a public-interest or satire exception the way defamation law sometimes does, and MHA's public materials do not address how the offence applies to ambiguous resemblance.
Second, this criminal track runs alongside a separate civil-regulatory one. Singapore's Online Safety (Relief and Accountability) Act 2025 (OSRAA) took effect June 29, 2026, standing up an Online Safety Commission empowered to order platforms, group administrators, and app stores to remove content, suspend accounts, or disclose a perpetrator's identity to a victim — with intimate image abuse named as one of the Commission's five initial priority harm categories from day one (Ministry of Law, June 2026). Larger platforms face binding response-time requirements — as fast as six hours for the most serious harm categories — with non-compliance risking penalties independent of any criminal prosecution (Baker McKenzie, June 2026).
Stacking a criminal offence on top of an administrative takedown regime with its own compliance clock is not unreasonable given the harm, but it does mean platforms and individuals in Singapore now face two independent enforcement tracks — one court-supervised with the ordinary burden of proof, one administrative and fast-moving — for overlapping conduct. That combination raises the practical stakes of getting classification right: an image wrongly flagged as intimate-image abuse could trigger a six-hour platform takedown well before any court has tested whether it meets the criminal statute's definition at all.
The proportionate path forward
None of this argues against the core move. Criminalizing non-consensual synthetic intimate imagery closes a genuine enforcement gap, and pairing it with a fast civil takedown channel gives victims recourse that a slow criminal court process alone cannot. The open question is procedural, not principled: as the Online Safety Commission builds a track record on intimate-image cases, Singapore should publish data on takedown volumes, false-positive rates, and how often OSC-ordered removals precede rather than follow criminal charges. A six-hour response window is only proportionate if the classification behind it is reliably accurate — and that is precisely the kind of claim regulators should be expected to substantiate with evidence, not assert by design.