Ai2’s AutoDiscovery Uses AI to Find Hidden Hypotheses in Big Data
The nonprofit lab’s latest tool uses surprise-driven exploration to decide what researchers should investigate next
Scientific progress has long followed a familiar arc: form a hypothesis, design an experiment, test, and repeat. But when dealing with sprawling datasets, the bottleneck isn’t in collecting information. It’s deciding where to look. Ai2 is betting that artificial intelligence can automate that decision.
The nonprofit AI lab has introduced AutoDiscovery, an open-source experimental tool that turns massive scientific datasets into structured starting points for new lines of inquiry. It stems from a 2025 research paper outlining a method for open-ended, autonomous scientific discovery driven by “Bayesian surprise”—a statistical signal identifying results that meaningfully diverge from prior expectations.
The question, according to Ai2’s senior research scientist Bodhisattwa Majumder, is whether an AI system could take a dataset and independently surface meaningful results, insights, and hypotheses without the constraints of a scientist’s time, research goals, incentives, or limited visibility beyond their expertise.
“You might have some surface-level ideas about what to do with the data, but then you get stuck in this ideation bottleneck—you just don’t know what to do after that,” he tells The AI Economy. AutoDiscovery becomes the antidote to this decision paralysis, analyzing enormous datasets to potentially uncover a gold mine of results that could be impactful for other groups.
From Goal-Driven to Open-Ended Discovery
If Ai2’s AutoDiscovery sounds a lot like its Asta DataVoyager tool, that’s because the two are closely related. Both are extensions of Asta—a framework introduced in August 2025 to bring AI into scientific research. “We’re not just building an assistant, but an ecosystem built on transparency, reproducibility, and scientific rigor,” Ai2 Chief Executive Ali Farhadi explained at the time. AutoDiscovery is also built on top of DataVoyager, relying on its core technology for data ingestion and analysis, as well as experiment and hypothesis execution.
Still, a key difference between the two lies in their approach.
Released in October 2025, DataVoyager offers a standard chat interface that requires scientists to begin with a hypothesis or research question. It supports multiple data formats—including CSV, Excel, JSON, HDF5, TSV, and Parquet—and returns not just clear scientific answers but also the code needed to reproduce the analysis, along with visualizations and a detailed methodology.
Majumder once said, “Asta DataVoyager lets [researchers] ask questions about their data in their own words and receive answers they can trust…Our goal is to shorten the distance between a researcher’s idea and a reproducible scientific result.”
In contrast to AutoDiscovery, however, DataVoyager is built around a defined objective. “It needs a goal. We call that setup…goal-driven discovery.”
AutoDiscovery takes a different approach. Instead of waiting for a prompt, “we are taking away the goal. We’re saying that it is open-ended. It doesn’t wait for the user to give a goal. It looks at the data and starts figuring it out, [coming] up with its own goal and exploring autonomously.”
Majumder stresses that the breakthrough isn’t a new analysis engine, but a system capable of steering the search itself—navigating multiple possible questions over long, autonomous runs. This tool decides what DataVoyager should do next. “This is really an extension towards the long-horizon discovery with the Asta tools that we already have,” he proclaims.
Using AI to Find Surprises in the Data
To begin using AutoDiscovery, scientists first upload their datasets to the system. It can accept any data type and is limited to 20 GB per session—confidential data is prohibited. Next, they provide context for the data, such as whether it consists of GPS coordinates, social science surveys, single-cell analyses, sequence data, or other types of data. Another way to think about it, Majumder says, is like explaining the work to a mentee—that’s the kind of context you provide to the system.
The final step is setting a computational budget, which controls how extensively the system can explore. Scientists specify the maximum number of experiments/hypotheses—with one credit per experiment—and AutoDiscovery can autonomously run up to 500 experiments per session. In honor of this launch, Ai2 is giving users a one-time grant of 1,000 Hypothesis Credits through Feb. 28, 2026, to get started.
Once everything is set up, the AI will start analyzing the data and generating hypotheses on its own. Majumder emphasizes that this process is not done independently: “It’s sort of like a snowball effect—it generates one, verifies it, and then, with inspiration of that, generates the next one. Then, with the inspiration of the regenerate, the other one.”
AutoDiscovery runs are compute-intensive and can take hours to complete—Majumder suggests running them overnight. Users will receive an email when the job has been completed. The sourced data will remain on the platform for seven days after the analysis has been completed before being deleted, though the hypotheses, plans, code, and results will be retained.
He claims Ai2 is doing something “really fundamental,” that being the use of surprise-guided tree exploration. AutoDiscovery treats the whole process as a tree of hypotheses and uses a “surprise” signal—how much a result contradicts the model’s prior belief—to decide which branches of that tree to explore next, rather than wandering the data at random. “It basically encourages surprising findings through the exploration,” Majumder states. “The exploration is not just random. It’s not just by chance that you are getting interesting hypotheses. It drives towards finding interesting hypotheses.”
In fact, he reveals that roughly 60-70 percent of the hypotheses generated by AutoDiscovery were also judged surprising by human scientists. That said, this doesn’t mean that the AI made a discovery. “Surprising means it’s worth looking at,” Majumder clarifies. That means the 60-70 percent is worth examining further, “and out of these, there might be one or two” worth the scientist’s time to write a paper about.
Proving AutoDiscovery Can Work
To demonstrate how AutoDiscovery works in practice, Ai2 spotlights how oncologists at the Paul G. Allen Research Center at the Swedish Cancer Center used it to analyze a dataset of breast cancer mutations. The tool first scanned for patterns in how different genetic mutations appear together in patients. As it narrowed its search, it identified an unexpected signal: among patients with a PIK3CA gene mutation, mutations in another gene, TP53, occurred less often than chance would predict.
The significance: If two mutations rarely occur together, it could mean they serve the same biological function. Either that or cancer cells carrying both can’t survive.
Ai2 states that in the beginning, AutoDiscovery had no strong assumption about whether such a relationship existed. After analyzing the data, the analysis identified the pattern as statistically surprising and flagged it for human scientists to further explore. The oncologists found that signal compelling—AutoDiscovery had identified a potentially meaningful relationship hiding in a massive dataset.
Majumder references another example involving an oceanographer who has spent the past 27 years observing reef ecology in the Gulf of California. “It’s a humungous dataset, five gigs of data,” he shares. “They have been diving into the water, collecting samples, and then digitizing them and creating these datasets. Takes up all of their time. They don’t know where to start—there are so many interesting insights.”
And in using AutoDiscovery, Majumder tells me that the oceanographer identified a “fantastic hypothesis regarding global warming and policies, and how does that affect the species densities in different parts of the sea.”
Unburdening Scientists to Help Advance Science
When the team at Ai2 submitted their paper to the Conference on Neural Information Processing Systems, or NeurIPS, last year, they admittedly never imagined that a system like AutoDiscovery “could actually power some sort of true discovery” like the cancer use case previously mentioned.
That said, Majumder believes this new AI tool can be invaluable to the science community. If anything, it’s enabling researchers to delegate data analysis to a virtual assistant rather than to students or other scientists, freeing humans up to pursue potential discoveries.
He doesn’t believe AI will create a dearth of student assistants, though he thinks it would be foolish to ignore the technology. “Ignoring AI to advance science is impossible. Someone is going to use it. It has to happen. It will happen,” Majumder shares. The important thing, however, is to ensure that AI is used carefully and honestly.
AutoDiscovery is only the latest effort made by Ai2 to bring AI to science. Along with its work on Asta, the nonprofit lab has invested in AI-driven cancer research, launched a tool to help researchers streamline literature reviews, and pursued open-source AI model development to drive scientific breakthroughs.
As is typical for the organization, everything about AutoDiscovery is made open-sourced. The team is releasing not only the tool’s code, but also additional science reports.