Meta’s Oversight Board says a label on synthetic media is not an adequate response when a deepfake is used to spread hate or target a public figure. In two decisions published on September 17, 2026, the Board criticised the company’s handling of manipulated videos and called for clearer escalation, stronger enforcement and better reporting tools.
The immediate case concerned a fabricated video of a UK local politician appearing to make hostile remarks about refugees. Meta had left it on Facebook. The Board overturned that decision, finding that the video should be removed under Meta’s hate-speech rules because it attributed serious criminality to refugees as a protected group.
The ruling matters because the Board treats the synthetic production of the video as part of the harm, not merely a disclosure problem. A viewer who encounters a convincing false clip in a feed may absorb the alleged statement before noticing a label, sharing context or correction. The video can then be copied, reframed and redistributed beyond the original post.
From label to response
The Board did not argue that labels are useless. It argued that they cannot substitute for a moderation system that assesses the content, its target and its likely effect. Its recommendations include broader application of high-risk labels for harmful deepfakes, stronger penalties for repeat sharing and systems that prioritise reports involving AI manipulation.
That distinction is central. Provenance signals can tell a platform or viewer something about how an asset was made or changed. They do not, by themselves, decide whether the asset is lawful, misleading in context or likely to cause harm. A platform still has to make and explain a moderation decision.
In a second case, the Board reviewed the treatment of an AI-manipulated video targeting a female campaigner. It found Meta’s existing approach inadequate, including the burden of requiring the target to self-report and the narrow interpretation of manipulated imagery. The Board recommended that Meta use AI-manipulation signals in its severity assessment and improve detection of copies. This is a recommendation about enforcement, rather than evidence that the recommended system is already operating.
Why the cases set a higher bar
Both decisions focus on distribution as well as detection. A label is most useful when it reaches a person before they rely on the clip. But a response to a harmful deepfake also has to work when the provenance metadata is missing, the asset has been recompressed or the original post has already been amplified.
For Meta, the decisions are recommendations rather than an automatic rule change. The Board can require action on the cases it decides; its broader policy recommendations are answered publicly by the company. The practical test is therefore whether the reporting paths, ranking systems and enforcement teams change in a way that lets people surface urgent impersonation cases before the harm compounds.
For readers, the lesson is narrower but useful: a synthetic-media label should prompt questions, not settle them. Identify the original uploader, seek independent reporting or an official source, and keep a record of the link and context before reporting a suspected harmful impersonation. Our companion analysis examines what provenance labels can establish—and what they cannot.