Marketplace Tech A test to weed out AI-generated deepfake images
Jun 4, 2025
Matt Groh, a lead researcher at Northwestern University specializing in AI detection, discusses a new test designed to help people identify deepfake images. He shares insights from a study that reveals the average person can correctly spot 5 out of 6 deepfakes. The conversation explores the challenges of detection, using analogies like 'Where's Waldo' to highlight familiar cues that can either assist or confuse spotting fakes. Groh emphasizes the importance of skepticism and critical thinking in navigating a world increasingly filled with fabricated media.
AI Snips
Chapters
Transcript
Episode notes
Scoring the Deepfake Test
- Nova Safo took the Northwestern deepfake spotting test and scored five out of six images correctly.
- This was the average score found by Northwestern's study on spotting AI-generated images.
Taxonomy of Deepfake Flaws
- Matt Groh defines key categories to spot deepfakes: anatomical, stylistic, functional, physics, and sociocultural inconsistencies.
- These taxonomies help systematically identify AI-generated image issues.
Easiest Deepfake Clues
- Anatomical implausibilities are often the easiest deepfake inconsistencies to spot because humans recognize typical human forms well.
- Some errors like extra limbs can be subtle and require focused attention to detect.