Meta's Oversight Board ordered the company to take two AI-generated videos off Facebook, one of a Labour councillor in Scotland and one of a young Muslim campaigner, and said Meta's existing defenses against synthetic media are "consistently and fundamentally insufficient." The board, which Meta created in 2020 as an independent arbiter of content on its platforms, published the rulings in a blog post on Thursday along with nine recommended policy changes. Meta has been asked for comment.
The first video showed the councillor, who is not named, appearing to say that refugees are welcome even if they rape local women, since white people do the same. The clip carried a visible signature of AI production: the audio did not fully match the movements of his face. The board found the post should have been removed for violating Meta's hate speech rules, because it attributed criminal and predatory sexual behavior to refugees as a group rather than describing the conduct of particular people. It should also have carried a "high-risk AI" label.
The part of that sequence worth dwelling on happened before the ruling. The board raised the video with Meta directly. Meta looked at it and concluded that it broke no content rules and required no AI label. This was not a moderation queue buckling under volume. It was the company reading its own policy against a clip it had been told to examine, and deciding there was nothing to do.
The second decision concerns an AI-generated video mocking a young Muslim woman from Europe who campaigned for better menstrual health education and for reducing the stigma around conditions affecting women and girls from ethnic minorities. The clip showed her giving health advice while performing absurd exercises and eating junk food, none of which she did. The board found it violated Meta's bullying and harassment rules and recommended removal. Fake videos and images mocking the campaigner have drawn tens of millions of views online, including on Meta's platforms.
Pamela San Martín, the board's co-chair, said deepfakes are increasingly used against politicians and ordinary people to harass women and silence them in public debate. The cases, she said, point to a wider problem: women who speak publicly on social questions face disproportionate harassment and disinformation. Meta and other social networks need stricter rules against the spread of deepfakes.
Most of the board's nine recommendations are about distribution rather than deletion. A majority of members wants Meta to widen the set of cases where the "high-risk AI" label applies, take further steps to reduce the spread of misleading AI-generated content, increase penalties for accounts that repeatedly circulate it, and publish more data on how labels are being applied. Two proposals are specific: demote posts carrying the "high-risk AI" label so they appear less often in feeds, and add friction to viewing AI-generated content, such as a warning shown before the clip that the user must dismiss with a separate tap. The board also proposed redefining "unwanted altered images" so the category covers deepfakes of private individuals credited with words or actions that were never theirs.
Taken together, those recommendations read as an admission. A body that has just ordered two takedowns is spending most of its remedy list on throttling, labeling and warning screens, which is what you propose once you have concluded that removal cannot keep pace with supply. The arithmetic inside the same ruling makes the case: two videos adjudicated, tens of millions of views on the campaign against a single woman. The board is not trying to clear the field. It is trying to make the field less rewarding to play on.
There is also a gap between the two halves of what the board did. It ordered two videos removed. It recommended the nine changes. The binding portion covers two posts on one platform; the portion that would change how Facebook handles the next ten thousand is advice.
Which leaves the councillor's video as the instructive case. Meta was handed a deepfake of an elected official by the body it built to second-guess itself, and decided the clip was acceptable and did not even warrant a label. The board can now tell Meta what its policy ought to say. This week established that telling Meta what its policy already says does not get a video taken down.