Meta must remove two AI-generated deepfake videos from Facebook. The technology company’s own Oversight Board has ruled as much. The independent body also criticised Meta for current measures against misleading AI content do not provide sufficient protection. The two videos – reported by The Guardian – showed a British Labour politician and a young Muslim woman in situations that had never taken place.
One of the videos showed a Scottish Labour councillor supposedly making provocative statements about refugees. According to the Oversight Board, the video displayed clear signs of AI manipulation. For instance, the audio and lip movements were not fully in sync.
Meta had initially not removed the video, nor added an AI label. The Oversight Board ruled that the video did in fact violate Meta’s rules on hate speech. The images linked refugees as a group to criminal and sexually predatory behaviour. According to the Board, Meta should therefore have removed the video and applied a ‘high risk AI’ label to it.
A second ruling concerned an AI-manipulated video of a young Muslim woman in Europe. She campaigns for better education on menstrual health. In the video, she appeared to be exercising in a ridiculous way and eating unhealthy food.
According to the Board, the images were part of a broader series of AI-manipulated videos and photos mocking the woman. Some of the material received tens of millions of views, including on Meta’s platforms. The Board ruled that the video violated the policy on bullying and harassment and ordered its removal.
The Board also warns of a broader pattern. According to the body, women who speak out publicly are increasingly becoming targets of AI-manipulated images and disinformation.
The Oversight Board described Meta’s existing safeguards against deepfakes as “consistently and fundamentally inadequate”. The Board therefore issued 12 policy recommendations across the two cases.
In the first case, the Board proposed nine changes. These include broader use of ‘high risk AI’ labels, tougher sanctions for accounts that repeatedly spread misleading AI content, and greater transparency about the use of AI labels.
The Board also wants Meta to adjust its algorithms. According to the recommendation, labelled ‘high risk’ content should appear lower down in users’ feeds. This is intended to reduce the spread of misleading AI content.
The Board also proposed showing a warning screen before users can view certain AI-generated content. In addition, according to the recommendations, Meta must broaden the definition of ‘undesired manipulated images’. This would mean that deepfakes of private individuals who are wrongly attributed with words or actions would also fall more clearly under the rules.
Business AM
Ltd
