Deep learning models rank simple images as more likely than complex ones, even when trained only on complexity

Researchers discovered that neural networks consistently assign higher probability to simpler images than complex ones, regardless of what data they were trained on. This suggests the networks have an inherent bias toward simplicity baked into their architecture—which means they're fundamentally misjudging how typical or realistic different images are, a problem that affects any system relying on these probability estimates for decision-making.

This isn't a minor technical quirk. If a medical imaging system or fraud detector is built on probability estimates from a deep network, and that network systematically under-rates complex patterns as 'unlikely,' you've built a system that's confidently wrong in ways that scale consistently across different architectures and training approaches. The finding that even networks trained on only the hardest samples still prefer simple ones suggests this is a structural property of how these models work, not a fixable training artifact—which means the bias is much harder to patch than previously assumed.