Researchers Double Deepfake Detection Accuracy to 80% With 1-Hour Training

3 articles · Updated · bbc.co.uk · Jul 11

Accuracy rose from about 40% to 80% after participants were trained to distinguish AI-generated faces from real ones, with a few people reaching nearly 100%.
Six cues drove the improvement: symmetry, proportionality, attractiveness, distinctiveness, expressiveness and memorability, replacing older tells like extra fingers that newer AI often avoids.
Confidence also increased after training, countering earlier findings that people were often overconfident while still misidentifying fake faces.
StyleGAN3-generated images were used in the tests, and researchers said AI remains weaker at recreating non-white, older and younger faces because training data skews toward young white people.
The work matters as deepfake fraud grows: Deloitte estimates US losses could hit £40bn next year, after cases including a £25m transfer triggered by a fake boss on a video call.

Sources

Left100%

bbc.co.uk4h ago

Researchers Boost Deepfake Detection Accuracy from 40% to 80% with Training

National Institutes of Health (NIH) (.gov)4h ago

The public mental representations of deepfake technology: An in-depth qualitative exploration through Quora text data analysis - PMC

BBC.com4h ago

Researchers Train Public to Spot AI Deepfakes, Boosting Accuracy to 80%

6 Sources

Will human training to spot deepfakes become obsolete as generative AI technology rapidly outpaces our perception?

As we learn to spot fake faces, is the next great threat undetectable AI-generated disaster footage designed to cause mass panic?

Can new laws stop deepfake harm, or is mandatory digital watermarking for all AI-generated content the only real solution?

Doubling Deepfake Detection Accuracy: Empowering Humans and Hybrid AI to Combat Fraud and Misinformation

Overview

This report highlights a major breakthrough showing that people can greatly improve their ability to spot AI-generated deepfakes through targeted training. By treating deepfake detection as a skill that can be learned, structured programs help individuals recognize subtle clues that separate real from fake faces. This marks a shift from relying only on automated tools to empowering humans as a strong line of defense. The findings suggest that with the right training, human accuracy in identifying deepfakes can be doubled, offering a promising way to combat the growing threat of AI-driven misinformation.

...