Oversight Board overturns Meta deepfake calls. It asks for nine changes

Meta’s Oversight Board ruled on 17 September that the company got two cases wrong. Both videos were made with AI and both stayed on Facebook. The board said Meta deepfake policy is too weak for the volume of synthetic video people now post.

Fake refugee remarks put in a Scottish councillor’s mouth

The first clip showed a Labour councillor in Scotland saying things about refugees that she never said. The lip movement did not match the words. The post passed 5,000 views and Meta left it up. The board overturned that call. It said the video broke the hateful conduct rules and should have carried a “High Risk AI” label.

Clips mocking a menstrual health campaigner drew tens of millions of views

The second case came from a young Muslim woman in Europe who runs a campaign about menstrual health. Fake clips placed her in scenes that never happened. They spread well past Facebook. The board ruled that the videos broke the bullying and harassment policy. Co-chair Pamela San Martin said deepfakes are used “to harass and silence women” in public life.

Six recommendations in one case, three in the other

The board wants Meta to use the “High Risk AI” label in more situations. It wants warning screens that people have to click through. It wants repeat sharers to face harder penalties, and it wants the ranking system to show this kind of video to fewer people. Meta has to take the two posts down. The nine recommendations are advice, not orders.

Why ad teams on Facebook should read the label rules now

Brands make a lot of video with generative tools. A wider “High Risk AI” label would touch more of that work. The label itself is not the problem. The reach limits sitting next to it are. European teams have already had to settle who has to label AI content under the new transparency rules, and some have started buying licensed faces for AI ads to stay clear of trouble. Platform liability is moving too: in August a Frankfurt court held Meta liable for scam ads that used a real brand.

Meta has 60 days to answer, so a reply is due around mid-November. The question that answer settles is narrow: does Meta widen the label, or does it only take the two posts down and leave the rest of its rules as they are?