arXiv submissions in high-energy theory have nearly doubled in recent months, raising alarms about AI-generated academic papers. Data from arXiv hep-th submissions shows a dramatic increase from stable yearly numbers around 600-700 to over 1,100 submissions in early 2026. The surge appears concentrated in recent months, with February 2026 submissions already exceeding previous years' monthly totals.

Physicist Peter Woit notes this trend follows Sabine Hossenfelder's prediction that AI agents could flood academic literature with mediocre papers. The quality concerns are already manifesting - researchers report a sharp increase in poorly generated papers mixing standard equations without meaningful connections. One paper combined claims about "topological fractal complexity" with trivial Python plots, while others showed coherent introductions but vanished logic in crucial sections.

Some researchers have stopped reading arXiv entirely, retreating to private channels like Slack or Discord for meaningful discussions. The democratization narrative appears flawed - while skilled physicists can use AI productively with proper guidance, others produce only incremental work and non-experts generate nonsense. This suggests the flood of AI-generated papers may actually degrade the research ecosystem rather than expand participation.