See Selected Projects

Science & discoveryHealth & careUnited States

Find and summarise the research on a question, with every claim cited

Ai2, a nonprofit research institute in Seattle, offers Asta: open AI agents that find papers, write summaries of the research with every claim cited, and analyse a team's own data, leaving the judgement to the scientist.

Try it withClaudeChatGPT / OpenAI modelsGeminiGrok

Deployed; Asta's paper-finding and literature-summary agents public since 26 August 2025; Asta DataVoyager released 1 October 2025 and in pilots at cancer centres; AutoDiscovery experimental since February 2026, Ai2 (Allen Institute for AI), with pilots at the Cancer AI Alliance and Swedish Cancer Institute

Why it matters

Before anyone can test a new treatment, set a policy or fund a study, someone has to find out what is already known, and that usually means days of searching and reading. Ai2 (the Allen Institute for AI), a nonprofit, built Asta so that researchers can hand that work to AI agents and still check every step: the institute says every output is cited and traceable to its sources, and that the agents are there to assist human researchers. Asta searches a 100 million-paper abstract index and a 12.4 million-paper full-text index. Its data agent, Asta DataVoyager, is being piloted by the Cancer AI Alliance, which unites four cancer centres, and by clinicians at the Paul G. Allen Research Center at Swedish Cancer Institute. The agents and the framework Ai2 uses to test them are open source, so universities, hospitals and public bodies can study them, run them privately or build their own.

How it works

Asta launched in August 2025 with three tools. Find papers reformulates a question, follows citations and explains why each paper is relevant. Summarize literature turns a research question into a structured summary, with each claim tied to the papers behind it. Analyze data, first offered in beta to selected partners and released as Asta DataVoyager in October 2025, takes a dataset (CSV, Excel, JSON, HDF5, TSV or Parquet) and a question in plain language, and returns an answer, code that reproduces the analysis, charts and a methods section that records its assumptions and statistical tests. Teams can use a hosted version, run it on their own servers or in a private cloud, and delete their datasets at any time. In February 2026 Ai2 added AutoDiscovery as an experimental feature: it starts from a dataset, proposes its own hypotheses, and writes and runs Python code to test them. Ai2 also publishes AstaBench, over 2,400 problems across 11 benchmarks, to compare research agents in the open.

Who it's for

Research teams, hospital and public-health analysts, policy units, charities and journalists who need to know what the evidence says, and anyone building a research agent who wants to test it in the open.

The prompt

You are helping me review the evidence on one research question with an AI research agent, so that every claim can be checked by a person. Do not ask me for patient data, personal details or unpublished results.

My question: [for example, does text-message follow-up after hospital discharge reduce readmissions?]
Who will use the answer: [for example, a hospital quality team, a ministry policy unit, a charity's programme team]
Scope: [years, countries, study types to include or exclude]

1. Rewrite my question as three to five precise search questions, and list the terms and synonyms an agent should search for.
2. Propose a summary template: what is established, what is disputed, what is unknown, with every statement tied to a named paper and the sentence it relies on.
3. Write rules for the agent: cite only papers it has actually retrieved, mark preprints and retracted papers, say plainly when evidence is thin, and never fill a gap with its own guess.
4. Give me a checklist to spot-check the output: pick ten citations at random, open each paper and confirm it says what the summary claims.
5. If I also have a dataset, list the questions to ask before any analysis (where it came from, consent, missing data) and what the agent must hand back: the code, the charts and a methods note.
6. List the decisions this review must not be used for on its own.

Leave placeholders rather than invent studies, figures or rules, and mark every step a qualified person must sign off.

Open inClaude (opens in a new tab)ChatGPT (opens in a new tab)Gemini (opens in a new tab)Grok (opens in a new tab)

Sources

The blog post, the agents page and the DataVoyager release are Ai2's own. SiliconANGLE and GeekWire are independent of Ai2 but report its announcements. No independent evaluation of Asta's accuracy was found.

  1. 1. Ai2 (Allen Institute for AI): "every output is cited and traceable to its sources". Asta: Accelerating science through trustworthy agentic AI (opens in a new tab)

  2. 2. Ai2 (Allen Institute for AI): "Quickly and precisely surfaces hard-to-find scientific research papers". Asta agents (opens in a new tab)

  3. 3. Ai2 (Allen Institute for AI), via Business Wire: "a powerful data analysis AI agent within the Asta ecosystem". Ai2 Launches Asta DataVoyager: Data-driven Discovery and Analysis for Science (opens in a new tab)

  4. 4. SiliconANGLE: "AutoDiscovery is available today as an experimental feature in Asta". Ai2 introduces AutoDiscovery, an automated scientific discovery AI system (opens in a new tab)

  5. 5. GeekWire: "find papers, analyze data, summarize literature, generate hypotheses, and cite their sources". An open-source AI sidekick for scientists: Ai2's new Asta agent platform aims to spark big discoveries (opens in a new tab)

Related projects