UW Astronomy chair Jess Werk proposes "cognitive sanctuaries," arguing department-level reforms are needed to protect PhD training in the AI era

The author proposes the "cognitive sanctuary," a department-level institutional model for protecting the thought process in PhD training, above all for PhD students, in the era of generative AI. Prior discussion of AI and academia has focused largely on assessment or tool-use norms (for example, the Harvard Summit on PhD Math Education in the Age of AI, held the same day, reached similar conclusions about assessment and AI use); the author shifts emphasis to faculty incentives, which she argues is where much of a department's leverage lies. This is a position piece; its argument rests on the author's observations as department chair, consensus at her department's faculty retreat, and cited literature (e.g., Kra 2026, Cohn 2026, Polanyi 1966, Gopen and Swan 1990), not on a controlled experiment or quantitative study.

The article argues that current productivity-weighted incentives shift the "technical debt" onto the most vulnerable members: PhD students competing for postdocs on publication counts, postdocs competing for faculty jobs on h-indices, and early-career faculty judged on grant dollar value and quantity of output—metrics that do not measure or reward scientific merit. The author combines Merton 1968 on the Matthew Effect and Kelly and Jennions 2006 and Caplar et al. 2017 on the h-index inheriting and magnifying bias with the new variable that generative AI makes productivity easy to manufacture, arguing the two together undermine the perceived value of doctoral-level study. Cited literature supports the existing findings on metric bias; the claims that AI makes productivity easy to manufacture and that technical debt is borne by the most vulnerable are the author's argument, with no new quantitative evidence presented.

The article offers three sets of concrete, immediately implementable department-level suggestions: PhD processes (detailed rubrics distinguishing pass, conditional pass, and failure; live, unassisted evaluation checkpoints; removing publication requirements for advancing to candidacy; clear AI-use boundaries), faculty reward systems (counting PhD mentoring in promotion and tenure criteria, publicly posting service assignments, trading advising work for teaching relief), and department community (seminars with discussion time, department-wide AI-use standards). These suggestions do not depend on field-wide reform but are framed as simple, department-level fortifications that can be made now, with AI-use boundaries explicitly binding advisors and external collaborators as well as students. This is an operational suggestion list based on the author's departmental experience; the text reports no outcome data on these measures after implementation.

The article uses a hypothetical scenario discussed at the department's faculty retreat to show the plausibility of the risk: OpenAI posts a preprint, press release, and public decision log of 40,000 steps reporting a five-sigma detection of an evolving dark energy equation of state, inconsistent with a cosmological constant, with error bars a factor of 2.5 tighter than the community could achieve from the same public data, drawn from a future survey in which the department has already invested heavily; no faculty were especially fazed and none thought the scenario implausible. The scenario transposes what is happening in mathematics into astronomy, illustrating the concrete form the gap between discovery and understanding might take. This is a hypothetical retreat scenario, not a real observational result, and the author states it as such; its value lies in reflecting the faculty's judgment rather than providing scientific evidence.

The article addresses department-level decision-makers and faculty, especially in STEM departments like the author's own astronomy department that already use AI heavily; its suggestions are explicitly framed as simple, department-level fortifications that can be made now, and the author states they do not constitute a case against generative AI, instead encouraging its open use "especially where nothing the student is supposed to be learning is at stake." The article also states these suggestions cannot address structural problems such as inadequate funding for basic science and grade-school education's growing reliance on AI.

This is a commentary article, and the loaded text is an incomplete version, lacking any figures or supplementary material the original may contain, so it is not possible to judge whether unshown quantitative analysis exists. Several figures cited in the text (such as astronomy arXiv submissions rising 14% from 2024 to 2025, more than half of papers posted in 2025 written with language-model assistance, and only about one paper in 66 using these AI tools declaring it) come from other works and are not reanalyzed here. The effects of the proposed cognitive-sanctuary measures after implementation, and the judgment that physical intuition cannot be developed through closed-loop agentic AI conversations, remain open to further evidence.