NEJM AI study: Physicians spend hundreds of hours editing AI drafts; commonest edits cost the most
kdpsinghlab · x · 2026-08-28
A team at UC San Diego Health (led by Lily Poursoltan) published a study in NEJM AI quantifying how much time physicians spend editing AI-drafted replies to patient messages.
Their two-step framework:
- Used AI to compare AI drafts against what physicians actually sent, categorizing the edits.
- Fit a mixed-effects model to estimate the time cost of each edit type, adjusting for message complexity via the first 10 principal components of neural embeddings of the original messages.
They identified 15 distinct edit categories. The largest per-message time increases came from:
- Interpreting radiology results: +70.1%
- Clarifying/ruling out diagnoses: +63.9%
- Interpreting lab results: +60.8%
However, the slowest edits aren't the costliest: high-frequency drafts about scheduling, lifestyle advice, and referral coordination imposed the greatest cumulative burden—hundreds of physician-hours per 1,000 physicians in just over a year. The work offers health systems a practical framework for monitoring and fixing human-in-the-loop AI.
More from AGI Musings
- Were economists 'dead wrong' about AI? The GPT-as-GPT debate resurfaces — soumitrashukla9 · 2026-08-28
- Goertzel: decentralized watermarking could beat World's Orb for proof of humanity — bengoertzel · 2026-08-28
- Anthropic's 'Model Welfare' Is Making Claude Worse as an Assistant — Nouni2 · 2026-08-28
- Why isn't growth 30% if AI and robots do all the work? — TheKanter · 2026-08-28
- Terence Tao warns of AI-generated proofs that no human understands — liuzhuang1234 · 2026-08-28
- Reverse Approach: Watermarking Humans Instead of AI Content — AkindaGood_programer · 2026-08-28