A short training session made US government intelligence analysts meaningfully better at telling real photos from AI fakes.
Researchers ran a randomized within-subject experiment with 32 intelligence analysts in 2024, collecting 2,544 image-level judgments. Before training, analysts hit 72% overall accuracy — better than a coin flip, but not good enough for operational use. After a 30-minute expert-led session covering patterns across 7 real and 50 AI-generated images, overall accuracy rose by 9 percentage points. The biggest gain came from correctly identifying real images as real, which improved by 14.2 percentage points — a meaningful fix for a specific blind spot.
Most detection research focuses on catching fakes; this study reveals the underappreciated flip side: people also struggle to trust authentic images. That asymmetry matters for intelligence work, where falsely flagging a real photo as synthetic can be just as damaging as missing a fake. The finding also pushes back against the fatalistic view that human judgment is too far behind AI image generators to be worth training.
The sample is small — 32 analysts is a thin basis for broad conclusions — and it is worth noting that a 30-minute briefing is far easier to dismiss than a sustained media literacy curriculum. Whether these gains hold weeks later, or scale outside a government analyst context, remains untested.