# Generate, debate and rank research hypotheses with AI

Google's AI co-scientist independently proposed a gene-transfer mechanism that Imperial College researchers had confirmed in experiments but not yet published.

Organisation: Google Research; Imperial College London
Status: Research
Place: United Kingdom, United States
Categories: Science & discovery, Health & care
Built with: Gemini
Try it with: Claude, ChatGPT / OpenAI models, Gemini, Grok
Time to try: 1 hour

## Why it matters

Antimicrobial resistance is a major health threat. A system that surfaces promising hypotheses in days helps scientists decide what to test next.

## How it works

A Gemini 2.0 multi-agent system generates hypotheses, debates them, ranks them in tournaments and evolves the best, with a scientist guiding it. Asked how capsid-forming phage-inducible chromosomal islands spread across bacterial species, its top-ranked hypothesis matched the Imperial team's unpublished finding (Google Research; Cell paper by Imperial and Google authors).

## Who it's for

Researchers and R&D leads.

## The prompt

```
Research goal: [state the open question in one sentence, plus constraints such as organism, lab methods and budget].

Work in rounds:
1. Generate 8 distinct hypotheses, each with the mechanism, the supporting evidence and a key test. Cite only papers you have actually retrieved with search; otherwise mark the claim [CITATION NEEDED].
2. Debate: run a knockout tournament (4 first-round matches, 2 semi-finals, a final). For each match, argue which hypothesis is better on novelty, plausibility and testability, and name the winner.
3. Show the bracket and a final ranking of all 8.
4. Improve the top 3 by fixing their weakest points.

Output the final top 3, the fastest experiment that would prove each one wrong, and the assumptions you are least sure about.
```

## Caveats

The hypothesis matched results the team had found but not yet published; the system did not run experiments.

## Sources

1. Google Research: Gemini 2.0 multi-agent; debate, tournaments, evolution; cf-PICI result. Accelerating scientific breakthroughs with an AI co-scientist (https://research.google/blog/accelerating-scientific-breakthroughs-with-an-ai-co-scientist/)
2. Cell (Imperial College London and Google Research authors): Top-ranked hypothesis matched experimentally confirmed mechanism; question took years to resolve. Penadés et al., 'AI mirrors experimental science...' (Cell), Imperial repository copy (https://spiral.imperial.ac.uk/server/api/core/bitstreams/03a8f6ae-2dd3-4b61-a828-8f40681087ec/content)
3. BBC News: Independent news report of the co-scientist result. AI cracks superbug problem in two days that took scientists years (https://www.bbc.co.uk/news/articles/clyz6e9edy3o)

Fact-checked: 16 Oct 2026

Page: https://escapevelocitysociety.com/picks/google-ai-co-scientist

See Selected Projects: https://escapevelocitysociety.com/picks.md
