How are national lab staff actually using generative AI?

The paper's central finding is that generative AI use at Argonne National Lab is small, growing and mostly exploratory. The authors surveyed 66 employees, interviewed 22 of them, and analyzed usage of Argo, an internal assistant built on a private instance of OpenAI's GPT-3.5 Turbo, over the first eight months of deployment. They report "small but increasing use" by Science and Operations employees, yet "few had made it a consistent part of their work," and they conclude that "use is largely experimental at this time." The use cases sort into two modalities. A copilot is conversational, where "the user gets real-time responses to questions posed to the AI." A workflow agent "navigates a complex task autonomously or semi-autonomously and returns the output to the user."

The modalities are the authors' framing, not the staff's vocabulary: "participants themselves rarely used the term copilot, rather we impose it for conceptual organization of the findings." Current use is mostly structured writing that people "can easily verify is correct," such as emails and reports. The envisioned use, extracting insights from unstructured text such as scientific literature, was held back by "fears of hallucinations and reliability." The authors' survey figures show the gap. 60 percent of respondents had at least tried summarizing literature, while workflow-agent tasks were less common: analyzing data at 43 percent, merging and cleaning data at 35 percent, and figure generation at 27 percent. The authors call workflow agents "in early stages of development and use," with agents emerging in both Science and Operations workflows.

Against the nearest notes, the central evidence here is self-report. The survey asks how often people use LLMs and what tasks they have tried, which is the kind of perception and reported-practice measure that Can self-ratings replace objective performance scores for AI competence? says cannot stand in for performance. The Argo log is behavioral, but it covers one tool, and the authors say it "under-counts total LLM usage." The contrast with Does generative AI shift knowledge workers away from communication? is instructive. That study uses Microsoft 365 trace data and finds a shift in the mix of activity among heavy users, defined as more than 100 uses. This paper has no trace data of that kind, so it cannot show a comparable shift. On the agent side, How should AI agents and humans divide research tasks? reports a research workflow in which agents already implement revisions, while the Argonne staff describe workflow agents as still early.

The excerpt does not establish productivity, output quality, or whether any of these uses changed how the work was done. Nothing in it measures outcomes; the survey and interviews record what staff say they do and expect. The sample is one organization, the usage log excludes commercial LLMs, and the concerns the paper lists (sensitive data security, academic publishing, job impacts) are reported as concerns, not incidents. The conclusion promises recommendations, but the excerpt does not include them. At the strength the evidence allows, one national lab's early experience supports a working description of how knowledge workers in a science organization are beginning to use an assistant. It does not forecast how science organizations as a whole will adopt generative AI.