Key themes
The growing threat: With generative AI, deepfakes, and synthetic media becoming more sophisticated, falsified audio and video now target politicians, influence elections, create non-consensual sexual imagery, and undermine evidence from war zones. Technical advances are erasing visible "tells" (blinking eyes, six-fingered hands) that once made detection easier. The real danger is not just deepfakes but the ability to plausibly dismiss real media as faked, eroding shared ground for democratic discourse.
Why detection alone isn't enough: Deepfake detection tools have fundamental limitations—they often work on only one type of manipulation, fail on low-quality social media content, produce confidence scores that are difficult to interpret, and don't distinguish AI manipulation from manual editing. Moreover, if detection tools are made public, those designing deceptions will test against them and adapt. Even at 85–95% effectiveness in ideal conditions, current tools are not in the hands of those who need them most: journalists, community leaders, election officials, and human-rights defenders globally.
Content provenance and disclosure: Standards like C2PA enable invisible watermarking and cryptographically signed metadata to travel with media across distribution, recording a transparent "recipe" of how AI and humans were involved in creation and editing. This approach is already emerging organically on TikTok and similar platforms. However, it must respect privacy and not force citizen journalists, satirists, or activists in repressive contexts to disclose identity—the goal is transparency about how AI was used, not transparency about who used it.
Structural accountability: Without a pipeline of responsibility from foundation models and open-source projects through deployment into systems and platforms, AI risks repeating the failures of social media platforms. Governments must ensure transparency, accountability, and liability throughout this chain.
Concrete cases
Gregory shares real examples from WITNESS's deepfakes rapid-response task force:
- Sudan: Machine-learning detection algorithms (trained on over a million examples of synthetic speech) proved audio was authentic, not deepfaked as claimed.
- West Africa: Detection was inconclusive due to audio quality and noise artifacts.
- India: A politician loudly denied leaked audio was real, but experts using his voice samples concluded it was at least partially authentic.
These cases illustrate that even forensic experts cannot always rapidly and conclusively separate true from false.
Broader context
Gregory's organization, WITNESS, has coordinated global efforts including the "Prepare, Don't Panic" initiative, which helps frontline defenders (human-rights advocates, journalists) navigate synthetic media and fortify their content's credibility. The talk frames this not as a panic-driven crisis but a solvable governance challenge if structural solutions are implemented now.
Notes
This talk is a comprehensive overview of the deepfake threat landscape and policy solutions from a leading human-rights technologist. Gregory avoids panic-driven framing (rejecting "deepfake is the root of all misinformation") while taking the threat seriously and centering the needs of vulnerable frontline voices (journalists, activists, election officials). The talk's emphasis on content provenance as a process standard (how AI was used, not who used it) is particularly important for balancing transparency with protection of vulnerable creators. His point that "any detection tip you know now is already outdated" underscores why detection-centric solutions alone will fail, motivating the shift to provenance and governance.