2026-07-21 / Signal #4
One-third of recent arXiv papers now read as AI-written
“One in three new science papers reads like AI wrote it. The detector completely fails on math - and that tells us something stranger about the future of knowledge.”
Why It Matters
Concrete measurement artifact showing AI flooding scientific culture/institutions; meta-weirdness that the detector itself breaks on math papers adds intelligible story about changing researcher identity (author vs. curator/editor of machine prose). Builder utility for detecting slop; not generic hype.
Evidence
unslop.run blog (recent, ~226 HN points) with calibrated detector study: rise from ~0.4% pre-ChatGPT to ~32% (peaking ~39% in early 2026), CS at ~65%, math anomalously low (~0.7%) due to notation confounding the model; full-text analysis of 12k+ papers.
Signal Read
Source Trail
Daily scan: 2026-07-21
- No public source URL captured yet.