Journal editor tests authors: 3 of 7 can't answer basic questions about their own AI-assisted papers
Dr_Atoosa · x · 2026-09-18
Nihar Shah, editor-in-chief of the open machine learning journal TMLR, revealed a troubling trend in a blog post: a growing number of researchers submit AI-assisted papers they don't fully understand themselves.
- The share of submissions to TMLR desk-rejected without external review has jumped from 6% in 2023 to 53% now.
- Shah interviewed 10 authors of desk-rejected papers (7 responded): 3 couldn't answer "basic questions"; 3 handled high-level ideas but stumbled on technical details; the only one who answered everything had a major flaw in the paper.
- Shah stresses authors are responsible for the accuracy and integrity of their papers, and is building an AI system to question authors about their understanding during submission.
Related event: TMLR Desk Rejects Surge to 53% as Authors Fail to Grasp Their Own AI Papers(4 posts)→
More from Safety
- Rumor: Mistral hacked, samples of unreleased tools leaked, codebase reportedly for sale — qtnx_ · 2026-09-19
- LLM answers are being gamed: GEO spam comments show the gray market is real — soumitrashukla9 · 2026-09-19
- 10% catastrophic risk claims plus misalignment reports leave users struggling to keep up — _arohan_ · 2026-09-19
- 10% catastrophic risk, evasive models, wet labs: even insiders find the narrative bizarre — _arohan_ · 2026-09-19
- ChatGPT quietly profiles you; 15 prompts to view and wipe its memory data — HeyAmit_ · 2026-09-19
- Gov. Newsom orders California to draft stronger AI safety rules, including possible kill switches — Polymarket · 2026-09-19