A new editorial published September 8, 2026 in PNAS Nexus proposes strict limits on how generative AI should be used in scientific research. Authors Charles Branas and Bruce Levine argue that while AI tools can handle routine tasks, the core intellectual work of science-forming research questions and making final judgments-must remain human-led.
The editorial draws a sharp line between acceptable and unacceptable uses of AI. Using these tools for nongenerative tasks like data classification or copy editing is broadly acceptable. But deploying AI to generate scientific ideas or seed research questions crosses into territory that demands scrutiny and full disclosure in manuscripts.
Where AI should not tread
The authors are explicit about what should be off-limits. "The use of generative AI for the core intellectual, creative, ethical, interpretive or accountability-bearing work of research-where AI replaces human interpretation and judgment-should be disallowed," they write. This prohibition extends to generating initial research ideas, research questions, ethical determinations, risk assessments, and policy recommendations.
Their concern is not just about accuracy. It is about the nature of discovery itself. Generative AI is trained to produce probable outputs based on known patterns and existing data. Scientific breakthroughs, however, often require departing from prior assumptions and breaking free of typical patterns. Serendipity counts in science.
The risk of homogenized thinking
If researchers across institutions rely on the same tools with similar inputs, the authors warn, science loses its happy accidents and human insights. They argue that science should not be unthinkingly optimized as if it were a manufacturing workflow producing standardized outputs. Research requires a wide diversity of inputs, approaches, and ideas.
The editorial, titled "AI and authorship: Norms and uses to preserve human-led science," makes one final demand: "The hard problem of the research question and the final evidentiary judgment in every scientific publication should be produced by human minds."
For researchers navigating the expanding role of Generative AI and LLM tools in their workflow, the editorial offers a framework rather than a blanket rejection. The distinction is functional: automate the mechanical, but protect the intellectual core. More broadly, these proposed norms arrive as the AI for Science & Research community grapples with where to draw the line between assistance and authorship.
Why this matters for science and research professionals
This editorial signals where journal policies and funding expectations may be heading. If you use generative AI in your research pipeline, the boundary between acceptable tool use and intellectual replacement is now under formal scrutiny. Expect disclosure requirements to tighten. The message is direct: your ability to articulate a novel research question and defend a final evidentiary judgment is the part of the work that cannot be delegated.
Your membership also unlocks:
