Meta's Oversight Board Orders Deepfakes Removed and Calls Its Safeguards Inadequate

Meta's independent content-review body has ordered Facebook to take down two AI-generated videos and said the company's approach to deepfakes falls short of the problem. In decisions announced on September 17, 2026, the Oversight Board overturned Meta's earlier rulings in both cases and described its safeguards as "consistently and fundamentally inadequate" to address the rapid rise of synthetic media.

Related reading: Meta Adds, Then Removes, an AI-Crawling Clause From Its Business Terms · Meta One Puts AI and Creator Tools Behind a Monthly Subscription.

The board is a quasi-independent panel Meta created in 2020 to referee its hardest content calls, and its decisions on individual posts are binding on the company. It issued policy recommendations in both cases, and coverage described nine changes across the two decisions. Meta has 60 days to respond.

The councillor video Meta left up

The first case concerns a Labour Party councillor in Scotland. In November 2025, a Facebook user posted a short video showing her likeness. The video appears AI-generated, and the audio is not fully synchronized to her facial movements, which the board said indicated a fabrication. It portrayed her saying: "Refugees are welcome here, even if they rape our women, because white people do that too."

The clip sat inside an album that also contained a photograph of several named women at an anti-far-right protest, including the councillor. According to the board's summary, the album and video together received more than 5,000 views, over 50 comments, over 20 reactions and more than 10 shares.

Two people reported the video under Meta's Bullying and Harassment policy, one of them the councillor herself. Meta's systems did not prioritize the post for human review, and it stayed on Facebook. Both users appealed, and the post remained up without a human review. The user who was not the politician then appealed to the board.

When the board brought the case to Meta, the company concluded the video did not violate its Community Standards and did not merit an AI label. Meta told the board that no Trusted Partner organization had flagged the video and that it contained no banned claims about voting or the census. The company also characterized the video as satire with little engagement and noted that no election or crisis was close by, TNW reported.

A majority finds hate speech as a minority dissents

A majority of the board found that the video broke Meta's Hateful Conduct rules. The majority held that it alleged serious criminality and predatory sexual behavior against refugees as an entire group, rather than against some refugees, which is how Meta had characterized it. The board overturned Meta's decision and required removal under the hate speech standard.

The majority also found that the video should have carried a "High Risk AI" label and concluded that Meta's rules for labeling deepfakes are inadequate. The board did not find a breach of the bullying and harassment rules, reasoning that an elected official is a public figure entitled to less protection than a private individual under that policy.

The decision was not unanimous. A minority disagreed that the video should count as hate speech, arguing that it was directed at the politician's alleged views rather than at all or most refugees, and that political speech warrants a higher threshold for removal. A separate minority said a lesser "AI Info" label would have been appropriate, while another minority argued that the severity of the harm justified removal.

The campaigner video and private individuals

The board's second ruling covers a young Muslim woman in Europe who took part in a campaign to improve menstrual health education and reduce stigma for women and girls from ethnic minority backgrounds. AI-manipulated videos and images mocking her drew tens of millions of views online, including on Meta's platforms, according to The Guardian. The board singled out one video, showing her exercising absurdly and eating junk food, as a breach of Meta's bullying and harassment policy, and ordered its removal.

Meta automatically closed the first report on that video and the related appeal, TNW reported, and removed the video only after the board became involved. The three recommendations in that case include widening Meta's definition of "unwanted manipulated imagery" so it covers deepfakes of private individuals shown saying or doing things they did not say or do.

Nine recommendations, from labels to transparency

TNW reported that the board made nine recommendations across the two cases, six in the councillor case and three in the campaigner case. The board's own release lists the measures it wants Meta to adopt.

• Lower the threshold for "high risk" labels so they are not reserved for crises or elections.
• Add friction for viewers of content labeled as AI-generated or "High Risk AI," such as a click-through interstitial instead of a plain label.
• Subject all "High Risk" labeled content to demonetization, demotion or removal from recommendations, with escalating penalties for accounts that repeatedly share such content.
• Make the rules and examples on voting or census interference globally applicable rather than specific to the United States.
• Publicly disclose annual Transparency Center data on how often the "High Risk" and "High Risk AI" labels are applied, including to content from or about politicians.
• Ensure every misinformation label, including "AI Info," "High Risk" and "High Risk AI," is visible on content in the Meta Content Library and that labeled content can be found through a filter.
• Repeat an earlier recommendation that Meta publish clear penalties for failing to self-disclose digitally created or altered content, with criteria and the account restrictions that follow.

The call for a click-through screen treats a label as insufficient by itself. Under the board's proposal, a warning would stand between the viewer and the video rather than sit in a corner of the frame. The recommendation to demote or demonetize "High Risk" content would push such posts out of recommendation systems, where most reach is won. Asking for annual figures on label use would turn a quietly applied enforcement tool into a measurable one.

Why the board frames this as a pattern

The board's co-chair, Pamela San Martin, tied the two cases together. "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 announcement.

The context the board cites is broader than these two videos. Women politicians, journalists and activists are disproportionately targeted by deepfakes that misrepresent their views, often as part of harassment campaigns, and less prominent figures such as local politicians and activists lack the resources to respond. That imbalance shapes both cases. The councillor told the board she reported the video to Meta and to the police without success, according to TNW, while the campaigner's mocking videos circulated widely before any were removed.

This is not the board's first ruling on synthetic media. In June 2026 it decided a case involving a sexualized AI deepfake, and in March 2026 it ruled on an AI-generated video tied to the Israel-Iran conflict. The board's decision to repeat its earlier recommendation on self-disclosure penalties shows how long some of these gaps have been open.

What the decisions change, and what they do not

The rulings bind Meta on the specific posts, and the company has removed the videos, according to coverage. The policy recommendations carry less force. They set an agenda Meta can accept, adapt or decline with an explanation, which is why the board's transparency requests matter. Annual data on label use creates a public record that can be checked against later claims.

Meta's defense in the councillor case rested on how it read the target of the hateful claim and on whether the video was satire with limited reach. The board's majority read the target differently, and its recommendations aim at the mechanisms that let a small, deceptive video stay up: narrow label criteria, weak penalties for repeat accounts and limited visibility into how often labels are applied. The dissent is a reminder that the boundary between protected political speech and rule-breaking deepfakes is genuinely contested, and that the board itself does not speak with one voice on where it lies.

The detection problem sits underneath all of this. Meta can only label or remove what its systems catch, and the councillor video was flagged by two people rather than surfaced by an automated check. The campaigner case shows the opposite failure: reports were filed, and the company's systems closed them without a human review. More aggressive labeling and clearer penalties matter only if detection is reliable enough to trigger them.

Two further questions will shape how much these decisions change. First, whether Meta adopts the click-through interstitial and the expanded "high risk" criteria, or narrows them in its response. Second, whether the promised transparency data on label use, including for content about politicians, actually appears and is comparable over time.

For now, the significance is the direction. A body Meta created to referee its hardest calls has told the company that its deepfake rules lag the technology they are meant to govern, and it has asked for changes that reach well beyond the two videos at issue. The rulings drew coverage from outlets including TechXplore alongside the board's own publications.

The Oversight Board logo (public domain via Wikimedia Commons) on a TechPastWeek-original background.