Meta oversight board calls company's deepfake rule 'incoherent'

Meta's Oversight Board has determined a Facebook video wrongfully suggesting that US President Joe Biden is a paedophile does not violate the company's current rules while deeming those rules "incoherent" and too narrowly focused on AI-generated content.

The board, which is funded by Meta but run independently, took on the Biden video case in October in response to a user complaint about an altered seven-second video of the president posted on Meta's flagship social network.

Its ruling is the first to address Meta's "manipulated media" policy, which bars certain types of doctored videos, amid rising concerns about the potential use of new AI technologies to sway elections this year.

The policy "is lacking in persuasive justification, is incoherent and confusing to users, and fails to clearly specify the harms it is seeking to prevent", the board said.

The board suggested Meta updates the rule to cover both audio and video content, regardless of whether AI was used, and to apply labels identifying it as manipulated.

It stopped short of calling for the policy to apply to photographs, cautioning that doing so may make the policy too difficult to enforce at Meta's scale.

Meta, which also owns Instagram and WhatsApp, informed the board in the course of the review that it was planning to update the policy "to respond to the evolution of new and increasingly realistic AI", according to the ruling.

The company said in a statement that it was reviewing the ruling and would respond publicly within 60 days.

The clip, on Facebook, manipulated real footage of Biden exchanging "I Voted" stickers with his granddaughter during the 2022 US midterm elections and kissing her on the cheek.

Versions of the same altered video clip had started going viral as far back as January 2023, the board said.

In its ruling, the Oversight Board said Meta was right to leave the video up under its current policy, which bars misleadingly altered videos only if they were produced by artificial intelligence or if they make people appear to say words they never actually said.

The board said non-AI altered content "is prevalent and not necessarily any less misleading" than content generated by AI tools.

It said the policy also should apply to audio-only content as well as videos depicting people doing things they never actually did.

Enforcement, it added, should consist of applying labels to the content rather than Meta's current approach of removing the posts from its platforms.