AI Matters
Why it matters
New benchmark reveals major flaws in AI visual perception
The Decoder New benchmark confirms AI models still perform poorly at visual perception
Moonshot AI's new PerceptionBench showed that no leading multimodal model achieved 60 percent accuracy in basic visual perception tasks.
Why it matters
Users relying on multimodal AI for visual tasks should remain cautious, as models frequently fail at basic image reading before reasoning.
Latest brief · Aug 15, 2026