Meta’s Oversight Board has issued a sharp new warning about the company’s treatment of AI-generated videos, arguing that its rules are not keeping pace with material used to impersonate, ridicule and harass real people. The criticism comes from two decisions involving realistic synthetic videos: one impersonating a Scottish politician, and another manipulating footage of a volunteer who had spoken publicly about menstrual-health education.
The board’s central finding is broader than either individual post. It says Meta’s policies are “consistently and fundamentally inadequate” for the rapid growth of AI-made content on its platforms. In practical terms, the cases expose three connected failures: a disputed video was not removed when initially reported, neither video received an AI label at the time, and one report plus its appeal were automatically closed before the board’s intervention led to a removal.
That is a consequential combination. A deepfake is not merely an edited clip or an obvious parody. It is synthetic or digitally manipulated media made to depict a person doing or saying something they did not do or say. When it convincingly borrows someone’s likeness or voice, the harm can arrive before a viewer ever learns the video is false. A later moderation decision may limit further spread, but it cannot completely undo the reputational, emotional or political damage caused by an initially believable fabrication.
Two cases, one repeated moderation problem
The first case concerned a realistic synthetic video of a Scottish politician. The clip made it appear that she had made an inflammatory statement about refugees and sexual violence. The politician whose voice was replicated described the impact as traumatic.
Meta told the board that it had not taken down the video when it was reported because the post had not been flagged by one of the company’s trusted-partner organizations and did not appear to interfere with voting or other processes. The content also was not labeled as AI-generated. The apparent reason was procedural: the person who uploaded it had not chosen to disclose that it had been created with AI.
The board disagreed with that outcome. It found the video should have been removed under Meta’s hateful-conduct rules and should have carried a prominent AI label. That distinction matters because the board is not describing a harmless disclosure gap. It is saying that the platform’s assessment of the post itself was insufficient, while its labeling system also failed to give viewers meaningful context.
The second decision involved footage based on a television interview with a volunteer promoting menstrual-health education. Accounts across social media later digitally manipulated the interview footage in a campaign of ridicule, and those videos spread widely. Meta automatically closed a report and a subsequent appeal concerning the video in question. It was not labeled as AI-generated at that point.
Meta ultimately removed the clip under its bullying rules, but only after the Oversight Board became involved. The sequence is central to the board’s criticism: an initial report can be the moment when someone targeted by a manipulated video most needs a quick, intelligible response. An automatic closure, especially where the material is never properly labeled, can leave the person with little visible route to correction while the altered media remains available.
Why labels and removals are different tools
AI labeling and content removal address different questions. A label tells viewers that a post has been generated or materially altered using AI. Removal takes the post off the platform for violating a rule. One is context; the other is enforcement. Neither automatically substitutes for the other.
In some cases, a clear label can be useful context for content that is allowed to stay online. But a label does not neutralize a post that independently breaks rules on harassment, bullying or hateful conduct. Conversely, a removal policy cannot be the only answer if ordinary users routinely encounter altered material that is allowed but indistinguishable from authentic footage.
The two decisions point to a particularly difficult issue for systems that rely heavily on uploader disclosure. If a person creating or posting deceptive material simply declines to identify it as AI-made, then the people most motivated to obscure its origin may be the least likely to trigger a voluntary label. The board’s recommendations therefore include prioritizing potentially AI-generated material for review after it is reported.
That recommendation does not mean every reported video is automatically fake. It means reports should help the platform identify posts that warrant closer examination, particularly when a real person says their image or voice has been fabricated or manipulated. The distinction is important: a report is an allegation that needs assessment, not proof. Yet treating it as irrelevant because it did not come through a designated partner risks missing precisely the sort of personal, fast-moving abuse illustrated by these cases.
What “prominent” labeling means in practice
Meta’s existing AI indicator typically appears as a small “AI Info” banner. The board’s emphasis on prominence signals a concern that a technically present notice may still be too easy to miss or too vague to help a viewer understand the post. A label that is detached from the video, shown only after interaction, or phrased without explaining the relevant alteration may not meaningfully counter a convincing false impression.
Prominence is not just a design preference. For a synthetic clip that appears to show a politician making a hostile statement, viewers need context at the point of viewing, before sharing or reacting. For a manipulated interview used to humiliate a health-education volunteer, a generic AI notice may also fail to convey the key fact: the person’s real interview was altered and repurposed to ridicule her.
The board’s decisions do not establish that every AI-made post involving a person must be removed. They instead underline that platforms need policies capable of distinguishing between permitted synthetic expression and materially deceptive or abusive impersonation. That work involves reviewing what has been altered, how the alteration is presented, whether it targets an identifiable person, and whether it violates existing conduct rules.
Public participation is part of the harm
The cases also focus attention on who bears the cost when deepfake safeguards fail. Oversight Board co-chair Pamela San Martin said AI-generated deepfakes are increasingly used to harass and silence women participating in public discourse, from politicians to private citizens. The board identified a troubling pattern in which women speaking publicly on issues are disproportionately exposed to harassment and misinformation.
That observation makes the issue larger than the accuracy of any one video. Manipulated content can discourage people from seeking office, discussing public policy, joining advocacy, appearing in interviews or sharing expertise on subjects such as health education. The harm is not limited to audiences believing a false clip. It can also be the pressure placed on targets who must spend time disproving fabricated material, reporting copies, facing abuse and deciding whether continued public participation is worth the risk.
For users, the cases are a reminder to be careful with media that appears unusually inflammatory, humiliating or tailored to provoke outrage. A realistic voice and face are no longer reliable proof that a video is authentic. Looking for clear context, avoiding impulsive shares, and reporting suspected deceptive impersonation can reduce amplification. But the board’s critique is ultimately directed at platform systems, not at users being expected to solve a detection problem alone.
What happens next
The Oversight Board makes policy recommendations to Meta and has previously pressed the company to strengthen its AI-content approach. Meta has made few changes to its AI labels and has not agreed to implement many of the board’s recommendations from other recent labeling cases. The company did not provide a response to the latest criticism, but it has 60 days to respond to these recommendations.
The immediate recommendation to prioritize likely AI content for review after reports could be significant if acted upon. It would shift attention toward a basic moderation reality: people harmed by a fake may be the first to know it is fake, even when automated systems and formal partner channels have not caught it. Whether Meta changes its processes remains unresolved.
These cases arrive amid wider concern about the reliability and misuse of AI systems. For another example of how AI-related failures can range from inaccurate outputs to more serious operational problems, see our coverage of six documented AI misalignment incidents. Deepfakes are a separate issue, but both debates point to the same practical requirement: systems that affect people need safeguards that work under real-world pressure, not only when users follow the ideal path.
Meta now faces a clear test. A useful response would need to address both speed and clarity: faster review when reports credibly identify AI manipulation, and labels that are noticeable enough to inform viewers before deception travels further. The board’s latest decisions argue that the current approach falls short on both counts.
