Columbia Prof John Hewitt Bans AI-Written First Drafts, Sets 'First Follow-Up' Rule for Lab
johnhewtt · x · 2026-09-03
John Hewitt, NLP researcher at Columbia, published his lab's policy on AI in writing and communication, arguing that as agents take on more research tasks, explicit norms are needed to prevent skill atrophy and shallow thinking.
The four core rules:
- Write first drafts without AI: drafts carry the valuable thinking; AI-assisted first drafts anchor too much.
- Be ready for the first follow-up: members must know their code, experiments, and presentations well enough to answer a labmate's reasonable first follow-up question without consulting an AI assistant.
- Obey Grice's maxims of Quantity and Quality even when agents don't: agents don't control information amount/quality; anyone stating a claim must personally understand it well enough to stand behind it.
- Own your technical terms: agents love coining technical-seeming terms; strong communication uses a small set of precisely defined terms consistently.
The policy covers papers, lab meetings, 1:1s, and Slack updates, and may evolve. Hewitt says he uses AI agents heavily (especially for coding) but worries skills like tensor-mucking may become obsolete while clear communication and thinking never will.
More from AGI Musings
- Referees increasingly allege papers were AI-written, but claims often unsubstantiated — RexDouglass · 2026-09-03
- New model generation will make novel RL discoveries, blurring training and work — andreisavu · 2026-09-03
- Five worrying AI trends in combination: models harder to monitor and autonomy accelerating — RobbWiller · 2026-09-03
- Ilya warns neocloud security is weak; X user outlines 4-step scheme to "steal" frontier model weights — Sam_Witteveen · 2026-09-03
- Debate: why should AI stay constrained by human notions of self and individualism? — yeastsplainer · 2026-09-03
- Why the 'AGI will care for us like pets' analogy fails — danfaggella · 2026-09-03