Op-ed: up to 30% of STEM PhD abstracts show AI use, and AI detectors are rarely wrong
Afinetheorem · x · 2026-11-05
University of Toronto professors Kevin A. Bryan and Dror Shvadron argue in The Globe and Mail that AI-text detection is far more reliable than skeptics claim:
- Haitian-Québécois author Thélyson Orélien's prize-winning book was accused of being largely AI-written; publisher Boréal cut ties, while he insists not one sentence was machine-written.
- A new study finds up to 30% of recent STEM PhD dissertation abstracts show AI assistance; Harvard op-eds, a Commonwealth Prize story in Granta, an Akutagawa-winning novel, and a NYT book review were all at least partly AI-written.
- Detectors work like Shakespeare stylometry: models such as Claude and ChatGPT leave statistical fingerprints — "lists of three," negative parallelism — that classifiers like Pangram 4 and GPTZero reliably spot.
- Their conclusion: excuses denying AI authorship are almost always wrong.
Related event: 30% of STEM PhD Dissertation Abstracts Suspected to Be AI-Written(2 posts)→
More from AGI Musings
- Chamath: long-horizon AI tasks are 'still a joke,' warns of disillusionment trough — 2C_ornot2C · 2026-11-05
- AI optimism is waning: essay dissects why enthusiasm falls short of expectations — WolframRvnwlf · 2026-11-05
- Do we need another Transformer-sized breakthrough? Direct weight editing may suffice for AGI — haider1 · 2026-11-05
- 'Compute Is the New Oil': If Nations Buy American AI Outputs, the Dollar Is Backed by Compute — TinfoilTricorn · 2026-11-05
- Ben Goertzel calls for a different kind of AI: compassion over pause debates — bengoertzel · 2026-11-05
- AI search may be eating the ad web: Google paid clicks slow from 13% to 8% — ElphieJ · 2026-11-05