Researchers Are Resisting AI Experimentation Mandates With Evidence

Someone on Bluesky described their organization's mandatory "AI experimentation period" this week — everyone required to try the tools and report back — and announced they were refusing.[¹] Instead, they'd spent the time reading four books and compiling an evidence document. The post got ten likes, which is modest, but the specificity of it captured something the aggregate conversation keeps dancing around: the resistance to AI in research contexts is no longer just instinct. It's becoming methodology.

That dynamic — institutional enthusiasm running ahead of researcher buy-in — is the sharpest tension on this beat right now. Governments are signing headline AI partnerships while the working scientists those partnerships are supposed to benefit remain skeptical, unconvinced, or actively building the counterargument. Grant reviewers are already receiving LLM-generated applications they don't know how to fairly evaluate. A paper circulating in academic circles is asking whether preprints even function the same way in a world where AI can execute research from a public abstract.[²] The infrastructure of scientific communication is changing faster than the norms governing it.

What makes this moment different from earlier rounds of AI-skepticism-in-academia is the texture of the pushback. One Bluesky commenter noted that industry-aligned voices are actively trying to discredit researchers pointing at problems where "the science and data just haven't caught up yet"[³] — framing the skeptics as obstructionists rather than practitioners doing appropriate due diligence. That framing war matters. When you label caution as bad faith, you don't resolve the evidentiary gap, you just make it harder to discuss. The researchers building evidence documents are responding, in part, to that pressure.

There are genuine enthusiasts in this conversation, and they're not naive. A framework being presented for automated scientific discovery in cognitive science — AI systems that support the generation and testing of theories of mind — treats the technology as a collaborator in theory-building, not a replacement for it.[⁴] Separately, work on AI-assisted Earth science teaching is circulating, arguing that grounding AI in set sources and auditing its claims actually sharpens student judgment rather than dulling it.[⁵] These aren't booster takes. They're conditional arguments, with constraints built in. The enthusiasm that's getting traction in research communities is the enthusiasm that comes with a methodology attached.

The infrastructure question is lurking beneath all of this. The University of Utah is preparing to run a TRIGA research reactor to power a small AI data center — a proof-of-concept for powering full-scale compute with microreactors.[⁶] It's a detail that sits oddly beside the evidence-document compilers and the grant-fraud worriers, but it belongs in the same story: science is being asked to both adopt AI and provide the physical substrate for it, simultaneously, without having resolved whether the adoption makes sense. The people being asked to use the tools are also being asked to power them. That's not a contradiction anyone in the conversation has named directly yet. It probably will be soon.