The Oversight Board pulled a councillor's deepfake after Meta's systems skipped human review

Two people reported the councillor video, and one of them was the councillor. Meta's systems kept it out of human review, and both appeals failed ^[9]. She said she had also reported it to the police, without success ^[10]. The case reached a human decision because the other person who reported it appealed to the Oversight Board ^[9].

Meta's defence was that nothing in its rulebook caught the video. No Trusted Partner organisation had flagged it, and it contained no banned claims about voting or the census ^[12]. Meta said the video was satirical and drew little engagement, with no election or crisis close by ^[13]. Even after the board took the case, Meta held that the video sat within its rules and required no AI label ^[11], and it told the board it had not penalised the user for failing to disclose AI use ^[14].

Little engagement measures distribution. The video and the album it sat in, which included a photo of several named women at an anti-far-right protest, drew more than 5,000 views ^[8]. In the second case, AI-manipulated videos and images mocking the campaigner had tens of millions of views online, including on Meta's platforms, according to the Guardian ^[26]. The councillor told the board that the words in the fabricated video ran contrary to everything she stands for ^[10].

"From politicians to private citizens, AI-generated deepfakes are increasingly being used to harass and silence women from engaging in public discourse," said Pamela San Martin, a co-chair of the board ^[5].

The removals are settled. Board decisions bind Meta ^[4], and it overturned Meta in both cases, the Guardian reported ^[3]. The rest arrived as requests. Meta has 60 days to respond, Engadget reported ^[24]. Sixty days from 17 September falls on 16 November ^[32].

Two of the requests would change what a Facebook user sees. The board asked Meta to put a click-through screen, instead of a label, on content marked AI-generated or "High Risk AI" ^[21], and to lower the threshold for high risk labels beyond crises and elections ^[20]. The pair sets how often someone scrolling has to tap past a screen to reach a post. On the threshold Meta applies now, the board's majority found this video should have carried the High Risk AI label ^[17]. Other requests would demote or demonetise High Risk content, penalise accounts that share it repeatedly ^[22], and have Meta publish yearly data on how often it applies the labels ^[23].

The cases separate on two axes a team can apply to its own queue: whether the subject is a public figure, and whether an outside partner can escalate for them. The councillor counted as a public figure, so the board found no breach of the bullying and harassment rules ^[18]. The campaigner was a private individual, and the board asked Meta to widen "unwanted manipulated imagery" to cover deepfakes of private individuals ^[29]. A private individual with nobody to escalate for them gets the automated path, and in both of these cases that path left the content up. For a team running the same queue, the measurable version is the share of reports it upheld only after someone complained from outside the product.