Product1 publisher3 min readPublished
Meta left a politician deepfake up because no trusted partner had flagged it
In the Oversight Board's two new decisions, a takedown waited on a partner organization's flag and an AI label waited on the uploader's own disclosure. Meta has 60 days to answer the recommendations, one of which asks for faster review.
The Product Desk · Product desk
What happened
- The Oversight Board published two decisions on AI video and said Meta's rules are "consistently and fundamentally inadequate to address the rapid rise of AI-generated content on its platforms."
- Meta took that video down for breaking its bullying rules only after the Oversight Board got involved in the case.
- Meta now has 60 days to respond to the board's recommendations, one of which asks the company to prioritize reported content that is likely AI-generated for review.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint The conditions Meta named for removal are held by other parties, so a private person reporting a video of herself is on the slowest route available to her.
- decision Any team copying Meta's disclosure design is choosing to make the label's existence depend on the uploader, the one party with a reason to skip the checkbox.
- exposure When automation closes a report and its appeal, the cost of that wrong call is paid by the person in the video while the edit keeps spreading.
- precedent On Meta's record with earlier AI labeling recommendations, the most likely outcome of this round is another published decision the company absorbs without changing the label.
A volunteer gave a television interview about menstrual health education. Accounts then, in the Oversight Board's words, "digitally manipulated the interview footage to ridicule her," and the edited videos went viral [8]. Someone reported the post to Meta. The report closed automatically, the appeal that followed closed the same way, and the video was never labeled as AI-generated [9]. It came down for breaking Meta's bullying rules after the board got involved [10].
The other decision spells out what Meta was waiting for. The company told the board it did not remove a realistic synthetic video of a Scottish politician when the video was reported, because no "trusted partner" organization had flagged it and because it did not appear to be interfering with voting or other processes [5]. The board said the video broke Meta's hateful conduct rules, and that it should have been removed and prominently labeled as AI [7]. The politician whose voice was replicated said the effect of the video was "quite traumatic" [4].
The label had a dependency of its own. According to the board, the video went unlabeled apparently because the person who posted it did not choose to disclose it was made with AI [6]. Meta's AI labels typically consist of a small banner reading "AI Info," and the company has made few changes to them [13].
Count the preconditions across the two cases: a trusted-partner flag, an elections signal, the uploader's own disclosure, and an automated queue that handled both the report and the appeal. None of the four produced a removal or a label, and the one removal on the record followed the board's involvement [17].
Board co-chair Pamela San Martin said in a statement that "from politicians to private citizens, AI-generated deepfakes are increasingly being used to harass and silence women from engaging in public discourse" [11]. She also said the two cases "demonstrate a broader, troubling pattern in which women who engage publicly on issues are disproportionately subjected to harassment and misinformation" [12].
Meta has 60 days to respond to the latest recommendations, including one that it prioritize likely AI content for review when it is reported [15]. The company has not agreed to implement many of the board's recommendations on AI labeling in other recent cases [14], and it did not respond to Engadget's request for comment on the board's latest statement [16].
If you run a reporting queue, the useful count is which of last quarter's removals were triggered by the affected person's report and which came from a partner escalation, a press inquiry, or an outside body. Where the affected person's report is a minority of triggers, automation is routing complaints and the decision is being made somewhere else. Prioritizing reported content that looks AI-generated, as the board asks Meta to do [15], buys the subject a faster path and spends reviewer hours on a bucket that also holds satire and edits the uploader already disclosed.
What to watch
- Meta's reply within 60 days, and whether it accepts prioritizing reported content that is likely AI-generated for review.
- Whether the "AI Info" banner changes in prominence or stops depending on the uploader's self-disclosure.
- Whether Meta narrows the cases in which its automation can close a report and an appeal without human review.