A study built on 15 million AI interactions finds scientists saving an average of 7 hours a week, with 74 percent reporting net time gains. A separate study of 207,000 astronomy papers finds 54 percent of 2025 papers carry detectable language-model traces while 0.81 percent disclose it. Together they describe a practice that is already common and a published record that does not document it, which means the literature alone does not tell you how a given result was produced. Two initiatives appeared within a day of each other on the same blog: a proposal for publicly funded AI research infrastructure modeled on CERN, and an announcement of open-weight mathematical models with explicit community data ownership. Both come from researchers who would rather build an alternative to lab-controlled access than wait for better terms.
AI saves scientists 7 hours a week, but 41 percent have a growing backlog of untested hypotheses
Mihai Codreanu, Arthur Turrell et al., Google and Google DeepMind, September 2026
Drawing on 15 million Gemini interactions, 2,600 specialized AI models, and a survey of 600 scientists, the largest quantitative study of AI use in research finds LLMs and domain-specific models act as complements rather than substitutes: 89 percent of time-savers spend over a tenth of that time checking AI outputs, and the bottleneck has shifted to verification and the growing pile of results awaiting testing.
More than half of recent astronomy papers are written with language-model assistance
Serat M. Saad and Yuan-Sen Ting, arXiv, September 9 2026
Analysis of 207,111 astronomy papers from 2015 to mid-2026 finds 54 percent of 2025 papers carry detectable language-model traces against a disclosure rate of 0.81 percent, roughly one disclosure for every 66 papers showing a trace; the authors also find detectability declining sharply between 2023 and 2026, which they read as authors adapting their language away from detectable patterns, giving any detection-based policy a shortening shelf life.
Editors of higher education journal coauthor editorial with fake references
Retraction Watch, September 8 2026
The editors of a higher education journal published an editorial in which at least nine of 31 references are fabricated or misattributed, with named scholars objecting to papers attributed to them that they never wrote; the failure is notable because the people who did not check the citations are the editors, which is the role the system relies on to catch exactly this.
Reimagining research papers as interactive and reliable AI agents
Jiacheng Miao, Joe R. Davis, Yaohui Zhang, Jonathan K. Pritchard and James Zou, Nature, September 16 2026
Paper2Agent converts a research paper and its codebase into an MCP server, so a reader asks questions in natural language and the agent invokes the paper's own tools and workflows; the authors call it a "virtual corresponding author," and their case studies include multiple paper agents collaborating to prioritize a causal gene for psoriasis.
Rehaan Ahmad and Raj Palleti, alphaXiv, September 2026
A discovery and discussion layer built over arXiv, with paragraph-level comments, engagement metrics, researcher profiles, literature review tooling, and an MCP server exposing the corpus as agent-callable; with $7 million in seed funding and Paper2Agent arriving the same week, the reading interface is migrating off the archive while the paper stays on it.
Anthropic launches a Life Sciences Verification Program
Nathan Frey, Anthropic, September 17 2026
Verified researchers and institutions at academic labs, biotechs, startups, and nonprofits can apply for access to Anthropic's most capable models for biology and drug development work; comments on the announcement raise the access question directly, asking about independent researchers without institutional affiliation and about which jurisdictions are covered.
Julian Jacobs announcing, Google DeepMind, September 16 2026
Google DeepMind has launched a publishing venue for evidence about the deployment of advanced AI systems, with chief AGI scientist Shane Legg as managing editor; the structural question it raises is what editorial independence means when the editor, the funder, and the subject are the same organization.
A CERN for AI-assisted science?
Dimitris Koukoulopoulos (University of Montreal), Terence Tao's blog, September 17 2026
Koukoulopoulos proposes publicly funded frontier AI infrastructure for research, citing the case of a researcher whose two-year result on prime gaps was overtaken within days by AI-assisted work as evidence that unequal access compounds rather than just accelerates existing inequalities; the argument is that company incentives favor rapid demonstrations while science also needs verification and attribution.
SAIR's Open Math Model Initiative
Terence Tao, personal blog, September 18 2026
Tao announces the first program from the Foundation for Science and AI Research, his non-profit for AI in mathematics: open-weight models with explicit data ownership provisions ("the mathematical community owns the data and decides how it is used"), explicit consent required for training, and open licences; the initiative targets everyday mathematical work rather than benchmark problems, with XTX Markets funding in hand.
Why I didn't sign the Fields medallists' letter
Timothy Gowers, Gowers's Weblog, September 17 2026
Gowers (Fields Medal 1998) agrees that mathematics faces a crisis but locates the primary risk not in undigested results but in the people: "a lot of people who would have done a PhD in mathematics and gone on to become custodians of the mathematical tradition will no longer wish to do so," and argues for adaptation rather than restriction.