AI novelty checks reject 96% of ideas human scientists later publish, UIUC paper finds
rohanpaul_ai · x · 2026-11-01
A University of Illinois paper shows that novelty checks in AI-scientist pipelines, which decide which generated ideas get compute, judge whole proposals at once and thus discard most ideas that build on prior contributions. A deployed check caught all 380 disguised copies but rejected 96% of ideas researchers later published.
- The authors built ClaimNov, which splits a proposal into claims, matches each against prior papers, and gives the judge claim-level evidence
- ClaimNov catches 97% of disguised papers with only a 1.5% false-alarm rate, keeping 91 of 100 novel proposals containing one spliced prior contribution
- Practical takeaway: score claims separately and keep proposals that build on a small number of known results instead of discarding them wholesale.
Related event: UIUC Paper Finds AI Novelty Checks Reject 96% of Real Research Ideas(2 posts)→
More from AGI Musings
- Only a few hundred people truly know B-rep/CAD geometry kernels, and most are retiring — rms80 · 2026-11-01
- AI 'brain drain' irony: brightest kids now curing cancer instead of building exchanges — OwariDa · 2026-11-01
- Ackman: Brain-computer interfaces will restore sight in years; warns neoclouds are leveraged plays — elonmusk · 2026-11-01
- After witnessing two crashes in a month, blogger predicts human driving banned by early 2030s — Dr_Singularity · 2026-11-01
- Stop treating AGI as a precursor of ASI — they measure different things — Realistic_Stomach848 · 2026-11-01
- Rationing vs. abundance: framing intelligence as a capability to be engineered — TinfoilTricorn · 2026-11-01