AI vision models flunk simple looking-puzzles that people solve in seconds
Researchers built 17 picture puzzles that need careful, back-and-forth looking, like counting patches on cloth or tracing a rope through crossings. Three human testers got 96% right; the best AI vision model got only 11%.
Why it matters
Reading medical scans, checking factory parts, and inspecting satellite photos all need the same skill: scanning an image, keeping track of what you have seen, and going back to check. This study shows today's leading AI systems still miss that skill, even with extra thinking time or the ability to write their own image-checking code. It is an early warning for anyone hoping to hand these jobs to AI soon.
Who's behind it: Jiarui Zhang and colleagues, University of Southern California.
Summary by the Lemma AI · how we grade