David Landsman

Breaking: UC Berkeley Haas just published research in Science showing AI is making researchers more productive. And making science worse. Researchers using AI publish 67.37% more papers. Receive 3.16 times more citations. Become team leaders 4 years earlier than non-adopters. But the field they study shrinks. The diversity of topics collapses. AI is not catalyzing exploration of new ideas. It is accelerating work in domains that were already saturated with data. The team analyzed more than 2 million papers across mathematics, physics, biology, social sciences, and humanities, uploaded between January 2018 and June 2024. Then they used detection algorithms to identify which scientists were likely using AI to write papers and compared output before and after adoption. The individual gains are real. The collective cost is hidden in the structure of what is being published. AI does not surface unexplored corners of science. It rewards researchers for going faster down paths that were already well-paved. The entire scientific enterprise is being optimized for output that the machine already knows how to support, which means the topics where the machine cannot help are being quietly abandoned. This pattern showed up in a separate study published in Nature on December 2024 by MIT doctoral student Aidan Toner-Rodgers. He tracked 1,018 scientists in a major US firm's R&D lab when AI was randomly introduced. AI-assisted researchers discovered 44% more materials. Filed 39% more patents. Generated 17% more downstream product innovation. The output of top-decile scientists rose 81%. The bottom third saw almost no benefit. The paper was hailed as the strongest evidence yet that AI accelerates science. Then in May 2025 MIT formally disavowed it. The university issued a statement saying the paper should be "withdrawn from public discourse" due to concerns about the integrity of the underlying data. The author is no longer at MIT. What remains is the Berkeley Haas finding, which has not been retracted and which describes a quieter problem. AI is not just making science faster. It is making science narrower. The deeper concern in the Haas paper is what it implies about peer review. Reviewers, funders, and policymakers now have to separate worthy research from a flood of polished but superficial work. AI is making it harder for evaluators to tell the difference. The bottleneck of the entire scientific apparatus is shifting from production to verification at exactly the moment verification has become hardest. For two years the AI productivity narrative has been simple. AI makes individual workers faster. The collective benefits will follow. The Berkeley Haas data shows the first half is true and the second half is not. Individual scientists are accelerating. The field they belong to is contracting. The papers are getting written. The discoveries are getting narrower. If the people building human knowledge are running into this trade-off, it is worth asking what version of the same trade-off is happening inside every company that adopted AI in the last 18 months. The output goes up. The diversity of what gets produced goes down. And the cost of that compression does not show up on any quarterly report. Source: UC Berkeley Haas and Cornell University, published in Science journal, January 2026 PDF: arxiv.org/abs/2412.07727