Key points
-
Capability demonstrations: Face-swapping (Tom Cruise impersonations), real-time voice transfer across faces, and facial expression transfer using captured video models. The technology is advancing from offline to real-time processing.
-
The authenticity crisis: When deepfakes become sufficiently convincing, society will no longer be able to trust visual evidence. Existing examples include fake images of political figures (e.g., Trump arrests), non-consensual deepfake pornography, and manipulated video.
-
Proposed solutions: Public awareness campaigns, new verification institutions, and crucially, property rights over individuals' biometric data so they control how synthetic versions of themselves are created and used.
-
Positive applications:
- Entertainment: Enabling actors to persist indefinitely (Mission Impossible, etc.)
- Education: A single inspiring teacher could teach in multiple languages and classrooms simultaneously
- Personal connection: People could interact with synthetic versions of deceased loved ones, preserving memory and emotional connection
-
Cultural exchange: Reduced language barriers through real-time voice and appearance translation
-
The control problem: Graham argues that individuals must own the data used to train the algorithms that generate their synthetic selves, not cede it through corporate terms of service. He demonstrates this by creating a photorealistic avatar of himself and submitting it for copyright registration.
Notes
The talk balances demonstrated technical capability with honest acknowledgment of harm and the need for new governance frameworks. Graham's proposal for property rights over biometric data is a concrete policy direction rather than hypothetical hand-waving. The live demonstrations (face-swap, voice transfer, facial expression transfer) show the technology at 2023 maturity; the speed of advancement suggests these capabilities have only accelerated since the talk.