What the first two months show
Singapore's Online Safety Commission (OSC) began operations on 29 June 2026 under the Online Safety (Relief and Accountability) Act 2025. On 22 September it reported more than 500 reports for the period to 31 August, and said it had resolved more than 90 per cent of eligible reports, according to AsiaOne's account of the briefing. It did not say how many of the 500-plus were eligible, so the 90 per cent figure has no published denominator.
The mix of cases matters more than the headline. Doxxing was 44 per cent of reports and online harassment 38 per cent. Intimate image abuse was about 8 per cent, online stalking about 7 per cent, and image-based child abuse about 3 per cent. The commission is mostly handling everyday harassment and doxxing, not the headline-grabbing categories.
The deepfake case
The most instructive case sits in the smallest category. A victim had faced online sexual harassment since 2022, including AI-generated deepfake nude images and videos. The OSC estimated that more than 1,000 such posts had been made across multiple platforms. They were removed after the OSC issued its first Access Disabling Direction.
This is the strongest argument for the new regime. For years, victims in this position had to report post by post, platform by platform, and could then litigate at their own cost. A single direction from a regulator that can reach the poster, page administrators and platforms is a real improvement in speed and cost. Outcomes like this are what a narrow, harm-specific statute should deliver.
Why deepfakes are still outside the formal scope
The Act covers 13 categories of harm, but the Ministry of Law's announcement says the OSC will first focus on five: intimate image abuse, image-based child abuse, doxxing, online harassment and online stalking. The rest, which the reporting we reviewed says include impersonation and inauthentic material, are to be "progressively implemented." We could not read the ministry's annex listing them, so we are relying on press accounts for which categories are deferred.
The strongest case for moving faster is straightforward. Generative tools have made fabricated sexual imagery cheap, and a harm does not become less real because the pixels are synthetic. A victim whose case does not neatly fit "intimate image abuse" may find the fast route closed. Waiting costs real people real time.
The deepfake case above also shows the gap is narrower than it looks. The sexualised, harassing deepfakes were handled through the harassment and intimate-image categories. The OSC dealt with the conduct and the harm, and did not need a separate deepfake category. The remaining gap is mainly non-sexual synthetic material: impersonation, fake endorsements, and fabricated statements attributed to real people.
The criminal law is already ahead of the civil route
Singapore did not start from zero. In a written reply on 5 February 2025, the Ministry of Home Affairs said possessing or accessing sexually explicit deepfakes was already an offence under the Penal Code, punishable by up to two years' imprisonment, with heavier penalties of up to 10 years for production and five for possession where minors are depicted. It also said Penal Code amendments would follow that year. We did not verify the final text or commencement of those amendments, and readers should check Singapore Statutes Online before relying on them.
That layering is sensible. Criminal law deters and punishes. The OSC's administrative route removes content quickly. Civil remedies compensate. Each does a different job, and the OSC's job is speed.
What proportionate extension looks like
The risk in the next phase is not that Singapore regulates deepfakes. It is that it regulates them too broadly. Three design choices will decide whether the extension is good policy.
First, target harm, not technique. A rule keyed to harassment, sexual exploitation, fraud or impersonation of a real person works. A rule keyed to "AI-generated" content catches satire, parody, art and accessibility tools, and pushes platforms to over-remove. The category list should stay defined by what happens to the victim.
Second, keep the takedown power narrow and reviewable. The government's July 2026 parliamentary answer is a good template. It said directions may extend to identical copies on the platform, that the Commissioner will not mandate particular technologies, and that directions apply only to access by Singapore users. Outcome-based duties leave room for platforms to innovate. They also avoid forcing blunt filtering systems on every service.
Third, publish the numbers. The OSC's 90 per cent figure is encouraging but incomplete. Before widening scope to contested speech such as impersonation, the commission should publish eligibility rates, time to resolution, appeals and reversals, and the share of directions that were challenged. Intimate image abuse is an easy case, because consent is the whole question. Impersonation and "inauthentic material" are harder, because they sit next to parody and political speech.
Bottom line
Two months of data support the OSC's design. It is dealing quickly with concrete, identifiable harms, and it has handled a 1,000-post deepfake campaign without a bespoke deepfake law. That is a reason to be patient about the deferred categories, not hasty. Singapore should extend the commission's reach to non-sexual deepfakes only where a specific victim and a specific harm can be identified, and it should publish enough data for outsiders to check that the speed is not coming at the expense of error.