Introducing AutoDiscovery—AI that explores your data and generates its own hypotheses

Introducing AutoDiscovery—AI that explores your data and generates its own hypotheses ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏

View in browser

Every dataset holds findings no one has looked for yet. The patterns are there – waiting in the rows and columns – but surfacing them requires asking the right questions. Most AI tools for science don't help with this part. They wait for you to provide a research question before they get to work.

Today we're releasing AutoDiscovery in AstaLabs, an experimental feature that flips that dynamic. Instead of starting with a question, AutoDiscovery starts with your data and asks its own questions—generating hypotheses in natural language, writing and running Python code to test them, interpreting statistical results, and using what it learns to keep exploring. Give it a structured dataset, set an experiment budget, and let it run. You'll get back a complete list of novel research directions, each one fully transparent and reproducible.

How does it decide what to pursue? AutoDiscovery uses Bayesian surprise—a measure of how much the system's beliefs change after seeing evidence. Results that meaningfully shift expectations get prioritized over those that simply confirm what's already assumed. By chasing surprise, AutoDiscovery naturally gravitates toward the unexpected: findings most likely to represent genuine discoveries rather than obvious patterns.

Researchers across disciplines are already using it. Oncologists at the Swedish Cancer Institute used AutoDiscovery to surface mutual-exclusivity patterns in breast cancer mutations that could inform treatment decisions. At the Scripps Institution of Oceanography, a marine ecologist explored 20+ years of rocky reef monitoring data and uncovered trophic-level relationships that would have required extensive manual iterations to discover. And in social science, AutoDiscovery helped an economist surface an unexpected finding about how expertise shapes engagement with AI-generated text—results that were independently verified and later published in a peer-reviewed paper.

Read more about these researcher case studies here.

Everything AutoDiscovery generates is available for you to inspect: every hypothesis, statistical analysis, and line of Python code. We're giving early users 1,000 free Hypothesis Credits to get started, available through February 28. Try it today—you may be surprised by what it discovers.

Need support or more information about AutoDiscovery? Please reach out.

—The Ai2 team

Join our Discord | We're now on Reddit!

Work with us | Ai2 newsletter archive