Fighting Deepfakes

Science Magazine (AAAS)
Science Magazine (AAAS)May 15, 2026

Why It Matters

Accurate physics in AI‑generated media will make deepfakes harder to spot, amplifying misinformation risks and demanding stronger detection and policy frameworks.

Key Takeaways

  • AI images often ignore correct shadows, reflections, and scale.
  • Humans rarely detect physical inaccuracies in synthetic visuals.
  • Current models lack built‑in physics constraints for realism.
  • Future AI‑CGI hybrids may enforce physics, raising misuse risks.
  • OpenAI currently lacks incentive to prioritize physics‑accurate generation.

Summary

The video examines how AI‑generated images frequently violate basic physical rules—shadows, reflections, perspective—making deepfakes harder to detect yet less reliable.

Researchers cite perceptual studies showing viewers overlook such errors because human vision evolved to accept coherent scenes without scrutinizing vanishing points. Current generative models prioritize pixel loss over physics, allowing them to ignore these constraints.

One speaker notes, “We didn’t evolve to ask, are the vanishing points wrong?” and predicts a future “physics‑preserving CGI meets AI” system, while warning that malicious actors could exploit it. He also criticizes OpenAI for lacking incentives to address the issue now.

If AI tools begin embedding accurate physics, verification will become tougher, raising stakes for misinformation, legal liability, and brand safety. Companies will need new detection standards and possibly regulatory guidance.

Original Description

In the latest Science Podcast, Contributing Correspondent Kai Kupferschmidt talks with Hany Farid, the godfather of digital forensics, about the never-ending battle against fake imagery and why Farid is not giving up.
#Deepfakes #Science #SciencePodcast

Comments

Want to join the conversation?

Loading comments...