Meta has been told by its own oversight body that it failed, twice, to handle AI deepfakes properly. The decision, published Thursday, orders the company to take down a fake video of a Scottish Labour councillor and a manipulated clip of a young Muslim campaigner, and it comes with nine recommended policy changes.
The board, a quasi-independent panel whose rulings bind the company, described Meta’s defences against AI-generated deception as “consistently and fundamentally inadequate” in a statement carried by the Guardian. In the councillor case, an AI-generated clip showed the unnamed local politician making inflammatory remarks about refugees that were never said. The video stayed live. When the board raised it directly with Meta, the company decided it broke no rules and did not even warrant an AI label. The board disagreed on both counts, ruling that the post violated Meta’s hateful conduct policy by accusing refugees as a group of sexual violence, and that it should have carried a “high risk AI” label.
Pamela San Martin, one of the board’s co-chairs, put the cases in a wider frame. “From politicians to private citizens, AI-generated deepfakes are increasingly being used to harass and silence women from engaging in public discourse,” she said in the board’s statement.
The second ruling concerned an AI-manipulated video of a young woman in Europe who had volunteered for a menstrual health campaign. The clip falsely showed her giving health advice while doing absurd exercises and eating junk food, part of a wave of mocked-up posts that drew tens of millions of views. The board found it breached Meta’s bullying and harassment rules.
What interests policy people is the remedy list. The board wants Meta to widen the circumstances in which “high risk” labels apply, to demote labelled content in feeds, to raise penalties for accounts that repeatedly spread deceptive AI material, and to publish more data on when the labels are actually applied. It also suggested friction: a warning screen that a user must click through before viewing suspected AI-generated content.
Read together, the recommendations amount to a complaint that Meta’s current toolkit was built for a slower era. Detection cues like out-of-sync audio were plain in the councillor video, the board noted, yet the company’s review process still cleared it. The board was created in 2020 to act as a referee on the platform’s hardest content calls, and deepfakes are the newest category it has had to build case law for.
For UK politicians and local officials, the ruling lands at an awkward moment, with AI video tools cheaper and better every month and general elections in other countries already marred by synthetic audio. The board’s verdict is that waiting for platforms to police this themselves has not worked. Meta has been approached for comment and has not publicly responded to the board’s rulings.
This article draws on reporting by the Guardian and the Oversight Board’s published decisions of 17 September 2026.