Loading...
In March 2023, a synthetic image of Pope Francis wearing a white Balenciaga puffer jacket went viral, racking up 28 million views on Twitter in 48 hours. The image was generated by Ren, a Chicago construction worker, using Midjourney v5 in roughly 90 seconds. It cost him nothing. Debunking it cost significantly more: AFP, Reuters, and Bellingcat collectively spent an estimated 40 analyst-hours verifying it was fake, tracing the metadata, and publishing corrections that reached perhaps 3 million people—a tenth of the original audience. This asymmetry accelerated. By late 2024, the AI-detection startup Sensity reported that synthetic media online had grown 550% year-over-year, reaching 500,000 new deepfake videos per month. Platforms like X and TikTok were flooded. A single person with a ...
Popular framing: Deepfakes are bad because some people use them to lie.
Structural analysis: The cost asymmetry — 90 seconds to generate, 40 analyst-hours to debunk — lets cheap synthetic media outpace verification by orders of magnitude; Gresham's law applies to information, with the bad driving out the good. Trust in the entire signal class collapses as an externality: people stop believing real images not because fakes are perfect, but because volume makes every signal suspect.
The popular frame treats each viral deepfake as a discrete enforcement problem to be solved at the content layer, which consistently misses that the system incentives make overproduction mathematically rational. As long as creation costs approach zero and verification costs remain fixed or rise, no content moderation regime can close the asymmetry — the correct interventions target the cost structure itself (liability assignment, provenance standards with economic teeth, public investment in verification as a common good) rather than individual pieces of content.
