Gemini for Science launches in Google Labs with Co-Scientist, AlphaEvolve and Empirical Research Assistance

Gemini for Science is Google's most explicit attempt yet to plant a flag in scientific research as a vertical for AI applications. The launch bundles several Labs-stage tools: Co-Scientist (collaborative hypothesis generation), AlphaEvolve (DeepMind's evolutionary code/algorithm search system, now pointed at scientific computing problems), Empirical Research Assistance (literature triage and synthesis), and a tighter NotebookLM integration. New Science Skills also land in Google Antigravity, Google's broader agentic platform.

Google DeepMind framed the launch around accelerating literature review, hypothesis generation, and computational modeling — the three places research teams complain most about wall-clock time. The pitch is that scientists can stay inside a single Google-hosted workflow rather than stitching together a PubMed search, a Python notebook, and a chat model. Co-Scientist in particular is positioned as a peer-style collaborator, not a search tool: it proposes hypotheses, critiques them, and iterates.

The competitive picture is interesting because OpenAI, Anthropic and Meta have all positioned 'AI for science' as part of their public-good narrative but none has shipped a coherent product suite. Google's advantage is genuine asset depth — AlphaFold lineage, AlphaEvolve, NotebookLM's grounded-document UX, and Google Scholar/PubMed integration potential — and Gemini for Science is the first time those assets have been packaged for outside researchers rather than buried inside DeepMind. Demis Hassabis's own X post on May 27 framed it as helping scientists 'unlock their next breakthrough.'

What to watch: how Google handles peer-review pressure on AI-assisted hypothesis generation (journals are increasingly wary), how Empirical Research Assistance handles paywalled literature, and whether AlphaEvolve in particular produces a publishable scientific result — that would convert this from an announcement into a category-definer. The gradual Labs rollout starting May 2026 means real third-party evaluations are still weeks away.