Third of new arXiv papers machine-written

Academic publishing is undergoing a massive, silent transition to LLM-assisted writing that traditional detector metrics fail to capture without rigorous baselining.

• –Grounding the study's threshold against a pre-LLM baseline successfully mitigates the "detector false positive" excuse, showing a true signal rather than tool noise.
• –The discrepancy between fields like computer science (65%) and math (0.7%) is likely a mix of actual adoption rates and detector blind spots on notation-heavy prose.
• –Because the detector is a lower bound (not finding all AI text) and doesn't distinguish between heavy editing and full generation, the actual role of AI in text creation is likely even higher.