ICLR 2027 hits 60,000+ submissions, nearly 90% possibly AI-written
DominiqueCAPaul · x · 2026-09-20
adrilemercier sparked debate: ICLR 2027 has received over 60,000 submissions, up more than 3x from under 20,000 at ICLR 2026 — and nearly 90% may be primarily AI-generated, with authors potentially not understanding their own papers.
- The workflow is trivially easy: a vague idea, a Claude Code session asking for a SoTA synthesis, a bit of question refinement, and the model writes a mediocre-but-plausible paper.
- DominiqueCAPaul argues the deeper issue: publishing is treated as an achievement regardless of content, impact, or usefulness, while good blogs now carry more signal.
Note: figures are unverified third-party claims.
More from AGI Musings
- Nature commentary: universities must stop protecting academic bullies or lose more scientists — skoularidou · 2026-09-20
- AI Debater: Distinguish First-Principles ASI Risk From Today's Regulatory Noise — JOBhakdi · 2026-09-20
- As Agents Start Choosing Products, GEO Replaces SEO as the New battleground — nikvassev · 2026-09-20
- Former AI safety skeptic: the 'preposterous' predictions are already coming true — dan_s_becker · 2026-09-20
- "Four Rivers of Computing": a personal syllabus for stealing pioneers' moves — threepointone · 2026-09-20
- Millennium problems become 'low-hanging fruit' once AI solves them, author laments — haider1 · 2026-09-20