AINEWS
Meta's own quasi independent review body, not a court or regulator, issued a binding ruling and prescribed nine policy changes.
Meta's own quasi-independent oversight board, funded by the company but with binding authority over it, has ordered Meta to take down two AI-generated deepfakes that targeted UK women in public life, and in the same ruling called the company's safeguards "consistently and fundamentally inadequate." The 17 September decision is the first time Meta's own review body has put a structural gap between its AI-labeling regime and its hateful-conduct rules on the public record with concrete policy changes attached.
The first case is a fabricated clip of a Scottish Labour councillor. In the original video, the woman appears to say refugees in her town "rape our women" and "groom our kids"; Meta's systems had flagged the audio as likely AI-generated. Even after the board raised the case directly, Meta concluded the clip did not violate its community standards and did not warrant an AI label. The board disagreed. The content attacks refugees as a group, alleging serious criminality and predatory sexual behavior, which is the threshold for removal under Meta's Hateful Conduct policy. The fact that the councillor is a public figure, who would otherwise receive less protection, does not change the analysis: the hateful-conduct rule turns on the group being defamed, not the individual being targeted.
The second case, an AI-generated video of a Muslim female campaign volunteer shown doing absurd exercises and eating junk food while offering health advice, sat on Meta's platforms for months. Engagement was modest, around 5,000 views, 50-plus comments, 20-plus reactions, and 10-plus shares, but the board concluded the content was a deliberate harassment artifact aimed at a private individual, not satire. A majority found the post should at minimum have carried a "high risk AI" label. Meta's rules at present do not require such a label in enough cases, the board wrote, and the company's labeling framework is too narrow to capture how synthetic video is actually used to attack women in public discourse.
Meta's AI-labeling regime, built to make synthetic media visible to users, worked as a procedural shield against a substantive hateful-conduct reading. In the Scotland case, the system flagged the audio as likely synthetic; Meta used that to argue the content was a labeling question, not a hate-policy question. The board's ruling closes the gap. A post that meets the hateful-conduct standard must be removed under that standard, regardless of whether a label is added, and Meta's own policy should say so explicitly.
The dissents are where the analytical lever sits. One member argued the Scotland video targets the politician's alleged views rather than refugees generally, and that the appropriate label is "AI info" rather than removal. A second argued an "AI info" label would suffice for the Muslim volunteer's video. A third agreed the videos target women in public life but said the harms were severe enough to justify outright removal, a stronger position than the majority's labeling-plus-removal split. The dissenters are not on the same side, which is the point: the case turns on where the line sits between labeling and removal, and the board is openly divided about it.
The ruling is binding, not advisory. The board made nine policy recommendations, including expanding when "high risk AI" labels can be applied, increasing penalties for users who repeatedly share labeled synthetic content, and adding transparency about how often AI labels are actually attached to flagged posts. Meta has 60 days to respond in writing. The board's co-chair, Pamela San Martin, framed the decision as part of a pattern in which women, especially women of color and Muslim women, are disproportionately targeted by synthetic harassment. The next milestone is not a public statement; it is whether Meta's 60-day response adopts, modifies, or rejects the nine prescriptions.