The release of ChatGPT as an app marked a turning point in how researchers use AI. Although many scientists are comfortable using AI for technical tasks, there remains unease with AI in scientific writing. That discomfort matters: writing is a primary interface between science and the community that supports it. If AI can help researchers communicate more clearly and efficiently—without compromising rigor—avoiding it may be more harmful than helpful.
Here we describe the use of large language models (LLMs) to accelerate manuscript drafting while preserving human ownership of the science and accountability. Our case study is a multidisciplinary paper currently under review that presents an AI system for automated sediment grain size analysis from streambeds. We did not prompt the LLM to write the paper from scratch. Instead, human authors first completed the research, generated figures, and produced initial text. We then iteratively engaged an LLM to draft topical sentences to convey a compelling story based on scientific outcomes and draft paragraph content informed by initial human text.
A safeguard was human verification and editing as part of iterations. Figures with captions were provided to help the LLM understand the technical science and improve paragraph content; a domain expert reviewed and corrected LLM interpretations and refined prompts, and the LLM then regenerated text. We also provided human identified keywords to constrain the LLM-generated storyline. Paragraphs were drafted one at a time. The LLM was told to not modify figures or tables and to not suggest references to avoid hallucinations. Use of the LLM was discontinued once a full draft was developed. Final revisions, references, and compliance sections were human-completed.
This human–LLM loop produced a high-quality initial draft in roughly eight hours, dramatically faster than our experiences in human-only writing. Critically, the LLM was constrained by human scientific insights and human generated figures/data; the LLM helped generate text constrained by human technical outcomes and expertise. We argue that openly sharing workflows and experiences turns individual LLM-engaged writing efforts into replicates of a community-wide experiment—accelerating convergence on robust norms as LLMs evolve.