A measurable share of submitted research papers are now substantially AI-written, and the institutions built to guarantee trustworthy knowledge are scrambling to respond. This episode unpacks what's actually verified, what's being overstated, and why the detection tools meant to fix the problem may be creating new ones.

AI-generated (NotebookLM) audio overview. Source: HexLocal in-house research — Research - AI Flooding Science (peer review integrity under automated authorship) (Dr. Priya Nair). Primary external sources include NeurIPS official conference reporting, Pangram AI detection analysis of ICLR 2025 peer reviews, and Liang et al. (2023) on AI-text detector bias against non-native English writers.

- NeurIPS's own data shows 28.2% of its 2026 Position Paper Track submissions substantially used AI for writing — a real figure, but one that applies to a single specialized track, not the conference as a whole

- Journals and conferences including ICML, ICLR, ACL/EMNLP, Nature/Springer, and Science have all moved to update authorship rules and enforcement in a short window

- The detection layer is the structural weak point: AI-text detectors are unreliable enough to produce false positives, with documented bias against non-native English writers

- NeurIPS's own desk-rejections were publicly disputed as miscalibrated, exposing the gap between deploying a detection tool and trusting its verdicts

- The "21% of ICLR peer reviews are fully AI-generated" figure comes from a commercial vendor, not ICLR itself — a case study in how directional data hardens into fact

- The deeper question isn't just about bad actors: AI as a writing accelerant and AI as a flood are two different problems, and the institutions responding may not be distinguishing between them