Messeri and Crockett warn AI is narrowing scientific inquiry

A Perspective in Nature by Yale anthropologist Lisa Messeri and Princeton cognitive scientist M. J. Crockett puts a name on something a lot of researchers have been circling. The risk from AI in science, in their reading, is not mainly that the models get things wrong. It is that they make us feel like we understand more than we do while quietly narrowing what science even asks.

The authors sketch four archetypes for how AI is being sold into research. AI as Oracle, digesting and communicating the literature at the study-design stage. AI as Surrogate, standing in for data collection or human participants. AI as Quant, running analysis over data that outstrips human working memory. AI as Arbiter, helping evaluate scientific merit in peer review. Each pitch fixes a human shortcoming, and each, Messeri and Crockett argue, comes with an illusion attached. An illusion of explanatory depth, when a researcher mistakes the Oracle's summary for their own understanding. An illusion of exploratory breadth, when the space of hypotheses collapses to those testable with AI tools. An illusion of objectivity, when a model with a specific standpoint gets treated as having none.

The stakes, in their framing, sit at the level of what they call scientific monocultures. Monocultures of knowing, where AI-suited methods crowd out other modes of inquiry. Monocultures of knowers, where a shrinking, more homogeneous set of perspectives shapes what gets published. As Messeri put it in the Yale write-up, 'there is a risk that scientists will use AI to produce more while understanding less.' Crockett's line is that replacing diverse standpoints with AI tools 'will set back the clock on the progress we've made toward including more perspectives.'

The honest caveat is that this is a Perspective essay, not an empirical audit. Messeri and Crockett are not measuring how much of today's published work already runs through Oracle or Quant workflows, they are not naming the fields most exposed, and they are not spelling out the specific journal-policy or funding-rule changes that would counter a monoculture. What they are doing is shifting the conversation from 'is the AI right' toward 'is the AI narrowing what right even means,' which is the harder question for anyone running a lab, editing a journal, or writing a grant program.

The forward-looking piece, and it is the useful bit, is that whoever takes the epistemic dimension seriously first, the universities that teach AI literacy as a social question, the journals that require disclosure at each archetype stage, the groups that keep AI-heavy and AI-light methods running alongside each other, gets to be the credible check on everyone else.