Moderation of deepfakes in Meta isn’t good enough, Oversight Board Says

Meta’s methods for identifying deepfakes “are not robust or comprehensive enough” to address how quickly disinformation spreads during armed conflicts such as the Iran war. That’s according to Meta’s Oversight Board — a quasi-independent body that guides the company’s content moderation practices — which is now calling on Meta to fix how it works Surfaces and labels AI-generated content across Facebook, Instagram and Threads.

The call to action stems from An AI fake video investigation About alleged damage to buildings in Israel that was shared on Meta social platforms last year, but the council says its recommendations are particularly relevant now, given the “massive military escalation” across the Middle East this week. The board says in its announcement that access to accurate and reliable information is vital to people’s safety amid the growing risks of using artificial intelligence tools to spread misinformation.

“The Board’s findings highlight that Meta’s current system for properly labeling AI content is overly reliant on self-disclosure of AI use and escalated review and does not meet the realities of today’s online environment,” Meta’s Oversight Board said. “The case also highlights the challenges of spreading such content across platforms, as the content appears to have originated on TikTok before it appeared on Facebook, Instagram and X.”

The board’s recommended steps include pushing Meta to improve its existing rules on misinformation to address deceptive deepfakes, and creating a new, separate community standard for AI-generated content. Meta is also required to develop better AI detection tools, be transparent about penalties for AI policy violations, and expand AI content labeling efforts. The latter includes ensuring that “High Risk AI” labels are added to synthetic images and videos frequently, and improving C2PA certification (also known as content credentials) so that information about AI-generated content is “clearly visible and accessible to users.”