What Is Your Laboratory’s Archetype? How It Shapes Your AI Strategy

Artificial intelligence is becoming increasingly common in scientific research, and journals, universities and research institutions are developing their own rules for using large language models. For individual laboratories, however, the challenge is deciding how AI can support research while reflecting the team’s values and priorities. A new white paper suggests that the best approach depends on the kind of research group a laboratory wants to build.

The paper was published on the arXiv preprint server in July by a team of space-science researchers.¹ Because it has not yet been peer reviewed, it does not present a universal policy for adopting AI in research laboratories. Instead, the authors describe four “archetypes” of laboratory groups, each with different priorities and potential uses for AI at various stages of the research process.

The idea for the paper emerged after members of the research team observed AI policies being introduced in astronomy and government agencies. They “struggled with how to pitch the idea to students and postdocs,” says co-author Sarah Burke-Spolaor, an astronomer at West Virginia University in Morgantown. “You can’t say, ‘Don’t use AI,’ because people are already using AI.”

AI tools can be used for many different purposes, including generating computer code, analyzing data and helping researchers who are not fluent in English improve their writing. The team therefore concluded that laboratories should begin by discussing their shared research values rather than focusing only on AI. “We needed to talk about what was important to us as a group, not about AI,” Burke-Spolaor says. “And as a principal investigator, what kind of group would you like to form?”

Which type of AI research laboratory is yours?

To help researchers develop an AI policy for their laboratory, the paper proposes four archetypes, each defined by its priorities and preferred uses of artificial intelligence. The characteristics of each laboratory are displayed in a radar diagram inspired by charts used to measure the difficulty of songs in the video game Dance Dance Revolution, says Michelle Ntampaka, an astronomer at the Space Telescope Science Institute in Baltimore, Maryland, and a co-author of the white paper. The authors also include a worksheet that researchers can use to identify their own laboratory’s priorities.

Collection: Impact of ChatGPT on your scientific career

A “high-leverage” laboratory prioritizes maximizing scientific impact with limited resources. Such a group might use large language models for code generation, data analysis and brainstorming. A “craftsmanship” laboratory focuses on developing expertise through practice, experimentation and failure rather than taking shortcuts. Its researchers might use AI for writing support and meeting agendas.

A “trust” laboratory places the greatest emphasis on reproducibility and transparency. It might use AI for debugging code and handling administrative tasks. A “data-management” laboratory prioritizes responsible data stewardship and may prefer self-hosted AI models to protect confidential or sensitive information.

The authors emphasize that these archetypes are philosophical caricatures rather than rigid categories. They are not mutually exclusive, and most laboratories will probably combine elements of several types. Laboratory priorities can also change according to the research project or situation.

Tali Tan, who oversees AI initiatives in graduate education at Harvard Medical School in Boston, Massachusetts, says that laboratories may sometimes prioritize preserving human judgment and cognitive skills. In such cases, researchers may avoid assigning difficult reasoning tasks to AI simply for convenience or speed. By contrast, teams working on large, fast-moving projects may place greater emphasis on maximizing resources and accelerating research.

Establishing a baseline for AI training