I’ve worked in open science for a long time, as a tool builder, facilitator, funder, and advisor. Some might say that makes me part of the “Open Science Industrial Complex,” a phrase I’m borrowing from Casey Greene, though it makes me sound more entrenched than I hope I am. In a recent interview with Springer Nature, Greene described “an approach to open science which is open first, science second. That can lead us astray.” His alternative: “It’s science first, rigorous is the goal and open is how we get there.” I agree with him. I have historically been a bit fanatical in my advocacy, but as I’ve aged along with the idea of open science, I’ve worked to revise my thinking based on what has worked, what I’m seeing now, and how science itself has changed. Openness is about making research more rigorous and more trustworthy.
Science is changing at a rapid clip, largely because of AI and LLMs in particular. This post isn’t meant to be an authoritative examination of AI and open science. Instead, I’m raising four points that I’ve been thinking about and others have been writing about, mostly as a way to add to conversations already happening in research circles. The broader world of AI x Science is big and growing fast, and I may nibble at its edges in future posts. Here I’ll stick to the corner I’ve worked in for years: open science.
The most visible discussion about AI and openness concerns open-weight models. Popular news coverage of China’s open-weight models and their capabilities tends to focus on safety and security, but there’s another angle that’s more interesting for science. In a recent Nature comment, “Why scientists should lead the shift away from AI mega data centres,” the authors argue that publicly available models running on local infrastructure can shrink AI’s environmental footprint and give researchers more control over their tools. Part of that control is that an open-weight model can run on your own machine, without any of your data going to the frontier labs. That can reduce costs, and it can also ease scientists’ fear of getting scooped (more on that below). In a companion interview with Berkeley News, co-author Fernando Pérez is blunter. The big frontier models are “complete overkill for many, many real-world tasks,” he says, and “I’m not particularly interested in a future where the future of science is paying a subscription to any one company, no matter how good their tools are.” Building very large centralized models, he adds, “also allows them to collect everyone’s data.”
I heard a version of this argument in person at a recent OSPO meetup on open source AI models at UC Berkeley, hosted by BIDS and the D-Lab. Fernando and Cathryn Carson discussed the nuances of the word “open” in a research context and asked, “How might we (scientists) find ways to have confidence in our data staying secure?” Fernando’s argument was that open source software, and therefore open models, give researchers control. If you run an open model yourself, you don’t have to rely on promises from companies or regulatory agencies about your data’s security. Open-weight models and open source software stacks make “privacy by design” possible, meaning data privacy is considered from the start of a project.
Another phrase I keep hearing is “AI-ready data,” which pairs nicely with concepts around data sharing. I’ve been talking with a lot of people about neuroscience data for one of my consulting projects, and it has come up in many of those conversations, particularly with people running data repositories and services. What the phrase means is still evolving, but I particularly appreciated a recent Science editorial by Julia Lane and Natasha Noy, “Data everywhere, but are they fit for purpose?” Their argument is that the problem is no longer too little data but too much. Google Dataset Search finds roughly two million new datasets a month. What’s missing is infrastructure that signals which data are fit for which purpose: rubrics that score quality and fitness, usage descriptors published with the metadata, and communities that certify trustworthiness. Lane, Noy, and colleagues have tested the idea with food security data on the National Data Platform, publishing data-usage descriptors as search metadata. A recent preprint by Atkins and Parashar sums up the shift: “the challenge of finding data has been overcome by the challenge of finding data that you can trust.”
It’s not enough for data to be accessible to models, whether open-weight or frontier. The data need to be used in ways that let us trust the outcomes. That means good documentation, well-structured data in standard formats, and careful consideration of a dataset’s limitations.
“AI-ready” is now federal funding language, too. NSF’s AI Datasets solicitation cites America’s AI Action Plan and asks for data “fit for their scientific purpose and curated with expertise.” In July, the National Data Platform received $24.5 million under NSF’s new Integrated Data Systems and Services program, part of an $83 million investment in national data infrastructure.
Not everyone thinks “AI-ready” is a new idea. In the abstract for an upcoming NIH seminar, Kesselman and Schuler argue that “FAIR data largely subsumes ‘AI-ready’ data” (FAIR being the long-standing principle that data should be findable, accessible, interoperable, and reusable). The next question, they suggest, is “what makes a repository agent-ready?” Whatever we call it, the curation burden falls on repositories that were underfunded long before AI arrived.
We all know that frontier labs would like to hoover up all the data and information we type into our computers. We can set permissions so they don’t train on our information, but who has read the fine print and really understands what happens when we upload a document and ask Claude a question about it? Apparently the companies aren’t always sure either. When NYU mathematician Tristan Buckmaster wondered whether his Codex sessions had fed into OpenAI’s claimed solution to a major math problem, the company’s first response was that it “cannot rule out that de-identified data derived from their usage of our products helped improve our models,” according to the New York Times, though it later said that kind of influence was impossible in his case.
A real tension for anyone who cares about open science runs the other way. A recent news piece in Nature reports that bots are scraping open data, and some scientists worry about losing control of the data they worked hard to share. The version that worries me is open data being scraped and recombined by AI faster than the people who created it can publish their own analyses.
Perhaps the cleanest way to avoid leakage is to download a model and use it locally, rather than through a connection to the web. That requires technical know-how, the right hardware, and a willingness to use a model that isn’t necessarily the latest and greatest. Anyone who’s been using Claude Opus 5.5 knows this might be a tough sell.
The math is changing, though. Carl Boettiger, another of the Nature co-authors, estimates in the same interview that open models run about six months behind the frontier. Models designed for local hardware are arriving, like Meta’s Muse Spark Open (which runs on two Mac Studios). And the single best model isn’t always the best tool: routed combinations of models now beat individual frontier models on a deep research benchmark. One caution: local tooling is not the same as a local model. Anthropic’s Claude Science can run on a lab’s own infrastructure “rather than sending data off to Anthropic’s servers,” but it runs the same Claude models everyone else uses, so it’s worth reading closely what leaves the lab and what doesn’t.
For those who can’t or won’t run a model locally, there are middle-ground options: enterprise and zero-retention tiers, institutional deployments, and privacy-preserving or federated approaches. The strongest example I know of is OpenMined’s double-blind evaluation of a frontier model using PySyft. Evaluators at AVERI and Singapore’s AI Safety Institute tested a Google DeepMind Gemini model without seeing its weights, and DeepMind never saw their prompts or responses.
This is the angle of AI in science I’m most excited about. I’ve been thinking for a long time about how to make data curation easy enough that sharing gets better. Back in 2013, I worked on a project at the California Digital Library called DataUp. It was ahead of its time by a long shot. We partnered with Microsoft Research, with funding from the Moore Foundation, to develop a tool that could scan a researcher’s Excel spreadsheet and suggest improvements. It checked for basic structural issues like multiple tables in a sheet, helped generate metadata, and could deposit the data into a repository. It was not particularly useful, primarily because the technology wasn’t yet capable of nuanced evaluation and help.
I think about that project sometimes when I’m working on my data with AI’s help. AI’s capabilities for cleaning, annotating, and systematically improving data are extraordinary. This hasn’t gone unnoticed in the research community. I’ve talked to many scientists (most of whom would not consider themselves coders) who have vibe-coded a pipeline for cleaning the data they work with regularly. These pipelines cut down on the tedious work of preparing data for use, and potentially for sharing in a form others can use.
A library of open skills, where researchers share pipelines like these, is starting to appear. The Department of Energy released Genesis Agent Skills on October 15, a free package of workflows built by national lab experts to help researchers use AI coding assistants. The developers of REDI, an open framework for automated data readiness, have packaged it as a skill an AI agent can call. What I haven’t seen yet are shared skills aimed at curation and deposit, the steps that get data documented and into a repository.
I’m also watching the tool space for a more general AI-enabled tool for openness. I envision it meeting researchers at different points in their workflow and making it much easier to add persistent identifiers to their work (RRIDs for materials, DOIs for the datasets they use), identify repositories for their data and other outputs, and prepare those outputs for deposit. It could also generate the kind of usage descriptors Lane and Noy call for.
Pieces of this already exist. Mark Hahnel has written about which parts of the research pipeline AI could automate and about how much of peer review it could take on. OpenAlex is linking research to the code it runs on and has connected more than 700,000 datasets to their funders using DataCite records. Paper2Agent turns published papers into interactive AI agents. Others are testing whether AI agents can reproduce a study’s computational results from its code and data (CORE-Bench).
I haven’t touched on the many groups working at the interface of science and AI, and they vary a lot in how open they are. On the nonprofit side there’s FutureHouse, Ai2, with its fully open Olmo models and Asta tools; AI BioDesign, a $94.6 million open science initiative from the Allen Institute, UW, and Fred Hutch; and Astera, with its new life sciences division, Radial, and a partnership with NSF on automated cloud labs. On the commercial side, there’s FutureHouse’s spinout Edison Scientific; Anthropic, with Claude Science and a wet biology lab; Google DeepMind’s Co-Scientist and AlphaGenome Atlas; Lila Sciences; and OpenAI.
There’s also a big push at the federal level to explore AI in science. The Department of Energy’s Genesis Mission, launched by executive order in November 2025, got more than $5 billion in federal commitments in July, and on October 8 DOE announced $159 million for 12 Phase II projects and up to $100 million in graduate fellowships. Not everyone is convinced. Stuart Buck calls Genesis “likely far too ambitious” and asks why it sits at DOE in the first place.
I’m by no means an expert in AI and science, but I am genuinely excited (and maybe a bit nervous) about what’s becoming possible in research, particularly for open science. Greene’s formulation works as a test for all of it: science first, rigor as the goal, openness as the way to get there. By that measure, open models earn their place because they give researchers control over their tools and their data. AI-ready data is worth the effort only if people can tell which data to trust. And AI tools may finally make documentation and deposit easy enough that sharing stops feeling like a massive burden.
All of this points somewhere specific for funders. Federal money is flowing toward AI for science, and commercial labs are building their own platforms, but the shared infrastructure underneath still runs on thin budgets. Repositories are closing, and AI-readiness is still judged by intuition. Repositories, curation standards, and open models are public goods: fund them sustainably and labs across the research community benefit.
*Disclosure note: I am currently consulting with Astera and OpenMined.
