State of AI in Science & Research 2026

AI is now embedded in scientific discovery, headlined by AlphaFold's 200-million protein structures and reinforced by two 2024 Nobel Prizes, while investment and regulatory approvals climb sharply.

Artificial intelligence has shifted from a research curiosity to core scientific infrastructure: AlphaFold has predicted over 200 million protein structures used by more than two million researchers, and AI breakthroughs swept the 2024 Nobel Prizes. Surging private investment in generative AI and a steep rise in FDA-cleared AI medical devices show the same momentum moving from the lab into regulated, real-world use.

200M+
Protein structures predicted by AlphaFold
Google DeepMind
2M+
Researchers across 190+ countries using the AlphaFold database
Google DeepMind
$33.9B
Global private investment in generative AI in 2024 (+18.7% YoY)
Stanford HAI AI Index 2025
223
FDA-authorized AI-enabled medical devices by 2023 (6 in 2015)
Stanford HAI AI Index 2025
2.2M
New crystals proposed by DeepMind's GNoME (~380,000 predicted stable)
DeepMind GNoME
6 to 223
FDA-cleared AI medical devices, 2015 to 2023
Stanford HAI AI Index 2025
Protein database growth since 2021 (%)
AlphaFold DB: 585%585%AlphaFold DBUniProt: 31%31%UniProtPDB: 23%23%PDB

Source: Stanford HAI AI Index 2025

Workflow stageRepresentative AI toolWhat it doesFree tier
Literature searchElicitFinds and summarises papers for a research questionYes
Citation checkingSciteShows whether citations support or contrast a claimNo (trial)
Mapping a fieldResearch RabbitBuilds visual citation graphs from seed papersYes (free)
Summarising PDFsScholarcyTurns papers into structured summary flashcardsYes
Academic writingPaperpalLanguage and submission-readiness checks for manuscriptsYes
Quick Q&A with sourcesPerplexityCited answers over the live web and uploaded papersYes

AlphaFold redefined the scale of discovery

DeepMind's AlphaFold predicted the structures of over 200 million proteins, effectively covering nearly all catalogued proteins known to science. More telling than the count is the reach: the AlphaFold database has been used by more than two million researchers across over 190 countries. This is what mature scientific AI looks like, not a demo but shared infrastructure that other labs build upon daily. The breadth of adoption, including in low- and middle-income countries, helps democratize structural biology that once required costly experimental methods.

AI now wins science's highest honors

In 2024, AI-driven research received top recognition when Demis Hassabis and John Jumper shared the Nobel Prize in Chemistry for protein-structure prediction, and deep-learning pioneers were honored in physics. This is a turning point: prizes traditionally reward decades-old foundational work, so honoring AI methods signals the scientific establishment now treats them as legitimate engines of discovery. We read it as validation that AI is not merely accelerating existing science but enabling results that were previously out of reach. That endorsement tends to pull funding and talent toward AI-native research programs.

Investment and databases expand together

Generative AI attracted $33.9 billion in private investment worldwide in 2024, up 18.7% on the prior year, according to Stanford's AI Index. That capital is visible in the data layer of science: since 2021, entries in major protein databases grew sharply, with the AlphaFold database expanding 585% and UniProt up 31%. The pattern is reinforcing, as better models generate more structures, which seed more research, which justifies more funding. For research-tool builders, the signal is durable demand for AI that produces and curates scientific data, not just chats about it.

Regulators are catching up to the lab

The translation from research to practice is clearest in medicine, where the FDA had authorized 223 AI-enabled medical devices by 2023, up from only six as recently as 2015. In parallel, 2024 saw a wave of large medical foundation models such as Med-Gemini alongside specialist systems for radiology and cardiology. Regulatory throughput is becoming a real constraint and enabler rather than a footnote. We expect approval volumes to keep rising as evaluation frameworks for clinical AI mature, pulling more research-stage models into deployment.

What changed in May-June 2026

Three developments moved AI-for-science from the lab into policy and practice this spring. On 5 May 2026 the US Center for AI Standards and Innovation (CAISI) announced agreements with Google DeepMind, Microsoft and xAI allowing government evaluation of frontier models before public release, extending the pre-deployment testing it already runs with OpenAI and Anthropic (CNBC, 5 May 2026). Google DeepMind detailed AlphaEvolve, a Gemini-powered coding agent paired with evolutionary search that has discovered new mathematical constructions and, deployed across Google's own infrastructure, continuously recovers about 0.7% of worldwide compute and sped up a key Gemini training kernel by 23%. And a Harvard Medical School / Beth Israel Deaconess study published in Science reported that an OpenAI reasoning model outperformed experienced physicians at diagnosing and managing emergency-department patients using only their electronic health records - a milestone for AI as a clinical-research collaborator.

Where AI is winning, by discipline

AlphaFold made the headlines, but AI is reshaping discovery across several fields at once. In structural biology it predicted 200M+ structures. In materials science, DeepMind's GNoME proposed about 2.2 million new inorganic crystals (around 380,000 predicted stable) - candidates for batteries, chips and superconductors. In weather and climate, GraphCast produces 10-day global forecasts in under a minute and beats traditional numerical models on many measures. In drug discovery and protein design, David Baker shared the 2024 Chemistry Nobel for computational protein design. In medicine and genomics, 223 FDA-cleared AI devices and medical foundation models such as Med-Gemini push AI into regulated, real-world care.

Methodology & sources

Figures are compiled from publicly reported data - Google DeepMind (AlphaFold, GNoME, GraphCast), the Stanford HAI AI Index 2025 (investment and FDA device counts) and the Nobel Foundation. Every statistic links to its primary source; we do not estimate beyond what those sources report.

Outlook: 2026-2027

We expect three shifts. First, agentic self-driving labs that plan, run and analyse experiments with minimal human steps. Second, domain foundation models beyond biology - chemistry, materials and climate. Third, rising clinical-AI approvals as evaluation frameworks mature. The bottleneck moves from raw model capability to data quality, validation and regulatory throughput.

Tools mentioned

Related guides

FAQ

How widely is AlphaFold actually used in research?

Very widely. DeepMind reports the AlphaFold database has been used by more than two million researchers across over 190 countries, with predictions for more than 200 million protein structures. It functions as shared scientific infrastructure rather than a single lab's tool.

Is AI moving from research papers into real-world science?

Yes. The clearest evidence is in medicine, where FDA-authorized AI-enabled medical devices rose from six in 2015 to 223 by 2023, and 2024's Nobel Prizes recognized AI-driven discovery, signaling AI has become a mainstream scientific method rather than an experiment.

How accurate is AlphaFold?

AlphaFold is highly accurate for many single-chain proteins, often approaching experimental quality, and reports a per-residue confidence score (pLDDT). It is less reliable for intrinsically disordered regions, large multi-protein complexes and the effect of single mutations, so experimental validation is still used for critical results.

Is AlphaFold free to use?

Yes. The AlphaFold Protein Structure Database is openly accessible and the model code has been released, a major reason adoption spread to more than 190 countries.

Which scientific fields benefit most from AI today?

Structural biology (AlphaFold), drug discovery and protein design, materials science (GNoME), weather and climate (GraphCast), and medicine and genomics.

What are the main limitations of AI in science?

Data quality and bias, limited reproducibility, and models that can produce confident but wrong outputs. Predictions still need experimental confirmation; AI accelerates and broadens science rather than replacing the experiment.

What is an AI co-scientist?

An AI system that participates in the research loop itself - proposing hypotheses, designing or running experiments and writing up results - rather than only retrieving information. DeepMind's AlphaEvolve, which searches for new mathematical constructions and code optimisations, is a 2026 example of the pattern.

Do governments test AI models used in science?

Increasingly yes. As of May 2026 the US CAISI has pre-deployment evaluation agreements covering OpenAI, Anthropic, Google DeepMind, Microsoft and xAI models, assessing frontier capabilities including scientific and biosecurity-relevant ones before public release.

Which AI tools do researchers actually use day to day?

The everyday stack is more mundane than AlphaFold: literature tools like Elicit and Research Rabbit, citation checkers like Scite, summarisers like Scholarcy, writing assistants like Paperpal, and general assistants like Perplexity or ChatGPT for first drafts and code.

More reports

Compiled by ToolGlance from publicly reported data; figures link to their sources. Updated 2026-06-11.

How we rate: ToolGlance scores combine pricing, core features, user-review signals and update frequency, compiled from public sources and vendor documentation — see our methodology. Figures are indicative and change often; always verify pricing and features on the vendor site before buying. Last updated 2026-07-14. Compiled by the ToolGlance editorial team.