Research AI Landscape

DR-05-hypothesis-generation

3 min read

DR-05 — AI Tools for Hypothesis Generation

Purpose: Assess the current landscape of AI tools for scientific hypothesis generation and experimental design. This capability area has 5 tools catalogued, all at trial status with no adopt recommendation — suggesting the area needs a clearer picture of maturity and viable use cases.

Use with: Claude Research mode (or equivalent deep research tool)

Prompt

You are helping a UK research university (University of Sheffield) evaluate AI tools that assist researchers in generating, refining, and evaluating scientific hypotheses. This is an emerging area where general-purpose AI tools blur into specialised platforms.

For each tool, research and return:

  1. Primary mechanism — how does it generate or support hypothesis generation? (e.g. literature synthesis, knowledge graph traversal, generative reasoning, experimental design templates)
  2. Evidence base — are there peer-reviewed evaluations of this tool’s hypothesis generation capability? Cite any key papers.
  3. Disciplines served — life sciences, physical sciences, social sciences, etc.
  4. Maturity — is it in active production use at universities, or still experimental/research prototype?
  5. Human oversight model — how does the tool support researcher verification of generated hypotheses?
  6. Licensing and cost
  7. Data training policy and GDPR
  8. Catalogue status recommendation — adopt | trial | assess | hold

Tools to profile

Specialised hypothesis generation tools:

  • Google Co-Scientist (Google DeepMind) — AI system for scientific hypothesis generation; announced early 2025
  • Elicit (elicit.com) — AI research assistant; hypothesis elicitation as a workflow feature
  • Asta — AI research planning and hypothesis tool
  • Consensus — does Consensus’s “Consensus Meter” support hypothesis evaluation?

Emerging / experimental:

  • Kosmos — AI research platform with hypothesis generation claims
  • FutureHouse AI — AI agent systems for scientific research (PaperQA2, etc.)
  • Sakana AI — “AI Scientist” automated research system (preprint 2024)
  • Galactica (Meta AI) — large language model trained on scientific literature (note: retracted from public after concerns — research status only)

General-purpose LLMs used for hypothesis generation:

  • How are researchers using Claude, ChatGPT, and Gemini for hypothesis generation workflows? What prompt frameworks or structured approaches have been published?

Additional research questions

  1. Evidence quality: What does the peer-reviewed literature say about the reliability of AI-generated hypotheses? Are there studies showing AI-generated hypotheses that were validated experimentally? What failure modes (hallucinated references, implausible mechanisms) are documented?

  2. Responsible use frameworks: Have any universities or research funders (UKRI, Wellcome, NIH) published guidance on using AI for hypothesis generation? What are the academic integrity implications?

  3. Discipline differences: Is hypothesis generation AI more mature/reliable in some disciplines (e.g. drug discovery, materials science) than others (social sciences, humanities)? Profile the most mature domain-specific implementations.

  4. Human-in-the-loop: What best practices exist for keeping researchers in control of the hypothesis generation process when using AI? Are there structured workflows that journals or funders accept?

Output format

Return a markdown table:

ToolMechanismEvidence baseDisciplinesMaturityHuman oversightCostData trainingGDPRRecommended status

Follow with:

  • A maturity assessment: is the hypothesis generation field ready for institutional adopt or trial recommendations, or should all tools remain at assess?
  • A responsible use note: what caveats should accompany any institutional recommendation for hypothesis generation AI?

Graph View