X user warns epistemic homogenization risk

Nikolai Slavov, a Northeastern University bioengineering professor, wrote on X on May 22 that AI is already useful for literature synthesis, hypothesis generation and recursive critique in research workflows. His thread argued that those gains are strongest in areas that can be framed as search and optimization problems, including parts of computational biology and chemistry. He also said current systems remain weak at identifying wrong assumptions, spotting anomalous findings and reframing a problem when the original setup is flawed. The post landed as AI-for-science tools are drawing new attention from researchers and companies. Nature published a paper on May 19 describing Google’s Empirical Research Assistance, or ERA, as a system designed to help scientists write expert-level empirical software. Google said the tool can search literature, write code, test alternatives and optimize against a stated goal. ### What was Slavov warning about? Slavov said the risk is not only that AI systems make mistakes. He said widespread use of similar tools by many researchers could create “epistemic homogenization,” a condition in which scientists approach questions through the same priors, the same search procedures and similar framings. That, in his account, would make it harder to detect anomalies or generate alternative explanations. Nikolai Slavov’s academic work centers on systems biology and single-cell proteomics, according to Northeastern University and his lab website. (coe.northeastern.edu) Those fields rely heavily on computational workflows, making them a natural place for debates over where AI can accelerate routine analysis and where it may inherit or reinforce faulty assumptions. ### Why does this issue come up now? Nature’s May 19 paper on ERA said the system produced 40 novel methods for single-cell data analysis that outperformed top human-developed methods on a public leaderboard. Google said ERA can evaluate thousands of options and is aimed at one of the most time-consuming parts of research: iterative computational testing. Those claims have helped sharpen discussion over which parts of science are becoming more automatable. (slavovlab.net) A separate Cell commentary on AI agents in biomedicine said researchers envision “AI scientists” as systems that support collaborative discovery by integrating models, tools and experimental platforms. That framing aligns with the broader push to use AI for drafting, coding, benchmarking and experiment design, while leaving validation and skeptical judgment to human researchers. ### Where do the limits show up? Slavov’s thread focused on failure modes that are harder to benchmark than coding speed or leaderboard performance. (nature.com) His point was that a system can be strong at synthesis and still fail when the starting assumptions are wrong. In research, those failures can matter most when a surprising result does not fit existing categories or when the right move is to question the framing itself rather than optimize within it. Google’s own description of ERA reflects that structure. The company said the system works by taking “a scientific problem and a measure of success,” then searching and optimizing against that goal. (cell.com) Slavov’s warning addresses what happens when the problem definition or success metric is itself incomplete. ### Why would many researchers using the same tools matter? The concern is cumulative. If many labs rely on similar models, similar training data and similar optimization routines, the outputs may converge even when the field would benefit from disagreement or reframing. That does not mean the tools are unusable; it means their strengths may be greatest in incremental or well-specified tasks, while anomaly detection and conceptual shifts still depend more heavily on human judgment. Recent public discussion around AI in science has moved in that direction. (research.google) Nature and Google have highlighted benchmark wins and new discovery tools, while researchers such as Slavov are pressing on the conditions under which those systems can miss the very signals that lead to new questions. ### What comes next for this debate? Google said on May 19 that ERA is part of its broader Gemini for Science rollout and that the Computational Discovery prototype is being made available through a trusted tester program in Google Labs. Researchers evaluating those systems now have a public set of examples in genomics, public health and single-cell analysis, alongside criticism from scientists such as Slavov about wrong priors, anomalies and shared model assumptions. (research.google) (nature.com)