As generative AI makes synthetic images, audio, and video nearly indistinguishable from real, a hard problem moves center stage: how do we tell what's genuine? Deepfakes — convincing fake media of real people — are the sharp edge of this, and the fight against them is a genuine arms race.
Why detection is losing ground
The intuitive fix is detection: train classifiers to spot fakes. It works, for a while. But it's inherently reactive — detectors learn the artifacts of current generators, and the next generation of generators eliminates those artifacts, so detectors must be retrained, always a step behind. As generation quality approaches perfect, the telltale signs detectors rely on shrink toward zero.
Detection asks "does this look fake?" But generators are trained precisely to not look fake. It's a race the detectors are structurally set up to lose.
The shift to provenance
Because detecting fakes gets harder, attention is shifting to proving what's real. Instead of asking "is this fake?", provenance systems cryptographically sign content at the moment of capture or creation, so authentic media carries a verifiable trail of where it came from and how it was edited. It flips the problem: rather than catch every fake, establish trust for the genuine.
Why it's ultimately a social problem
No technical fix fully solves this. Watermarking can be stripped, detection can be evaded, and provenance only helps if platforms and people adopt and check it. The deeper challenges are social and institutional: media literacy, platform policies, legal frameworks, and norms around labeling. The technology to make convincing fakes is out; the response is a mix of provenance standards, detection where it helps, and — most importantly — a public that learns not to trust media at face value. It's one of the defining trust challenges of the era.
