Stanford HAI AI Index Report 2026: science and research findings
AI in science is real, measurable, and unevenly distributed — but end-to-end autonomous discovery remains far from reliable. The ninth edition of the Stanford HAI AI Index provides the most comprehensive independently sourced picture of AI’s trajectory available. For research institutions, the headline message is that AI is now a routine part of scientific practice across disciplines, yet performance on complete research tasks consistently falls well short of expert-level benchmarks, and the gap between what AI can propose computationally and what scientists can feasibly validate experimentally remains wide.
Source
Stanford HAI AI Index Report 2026, Stanford Institute for Human-Centered Artificial Intelligence (HAI). Independent initiative, ninth edition.
- Online science chapter: https://hai.stanford.edu/ai-index/2026-ai-index-report/science
- Full PDF:
ai_index_report_2026.pdf(378 pages, locally available in this vault) - Chapter 5 (Science) was developed in collaboration with Schmidt Sciences.
AI’s role in science: three maturity tiers
The report frames AI’s role in science across three coexisting tiers that differ significantly in maturity:
| Tier | Description | Maturity (2025) |
|---|---|---|
| 1. ML over scientific data | Machine learning applied to scientific datasets to build predictive and explainable models | Mature — commonplace for several decades |
| 2. Workflow assistance | AI systems that assist scientists with literature synthesis, experiment design, or data analysis | Emerging — expanded considerably in 2025 |
| 3. Autonomous discovery | AI systems generating new scientific discoveries with limited human guidance | Early — gaining traction but unreliable for end-to-end tasks |
The year’s most visible developments were in tiers 2 and 3. Notably, clear breakthroughs cluster in domains with strong existing data infrastructure (structural biology, physics, chemistry, materials science) rather than in fields with the most sophisticated theoretical models.
The report is explicit that these developments do not automatically translate into scientific progress. Experimental validation remains expensive and time-consuming, and the gap between what AI can propose and what scientists can feasibly test is a recurring theme across every domain covered.
Publication and research landscape
Volume
AI-related publications in the natural sciences reached 80,150 in 2025, up from 63,547 in 2024 — a one-year increase of roughly 26%. Physical sciences and life sciences each grew ~27–28% year-on-year; Earth science grew ~23%.
As a share of total scientific output, AI-related work remains a single-digit fraction of each field but is climbing quickly:
| Domain | AI share of publications (2025) | (2010) |
|---|---|---|
| Earth science | 8.8% | <1% |
| Natural sciences (overall) | 6.8% | <1% |
| Life sciences | 6.5% | <1% |
| Physical sciences | 5.8% | <1% |
Who is publishing
Academia produces 68% of AI CS publications globally (2024), with government (12%), industry (12%), and nonprofits (5%) making up the rest. In the US, industry’s share is higher (25%), while China’s government sector is more prominent (25%). Europe has the highest academic share (55%).
Industry contributions to the top 100 most-cited AI publications have fallen sharply — from 17 in 2021 and 19 in 2022 to just 6 in 2024 — even as industry’s share of frontier model releases has grown. This widening split is significant: the most influential research is increasingly academic, while the most capable models are increasingly industrial.
Top organisations in the 2024 top-100 citations: Stanford University and Google (7 papers each), Chinese Academy of Sciences and Microsoft (5 each).
UK and Sheffield context
- UK ranks 5th globally in top-100 citations (9 papers in 2024, up from 6 in 2023)
- UK ranks 6th in top AI authors and inventors per 100,000 population (49.64 per 100k; Switzerland leads at 110.45)
- UK has the highest share of PhD-holding AI authors globally at 51%, suggesting the UK research base is disproportionately postgraduate in character
- 4 in 5 US university students now use AI for schoolwork — UK patterns likely similar
Research workforce
- New AI PhDs in US and Canada increased 22% from 2022 to 2024
- Critically, all of this growth went to academia, reversing a decade-long trend of AI PhDs flowing into industry — this creates potential collaboration opportunity for research universities
- CS undergraduate enrolment at US four-year universities fell 11% between 2024 and 2025
- Gender balance among AI authors and inventors has not improved meaningfully in any country from 2010 to 2025
AI across scientific domains
Physics, astronomy, chemistry, and materials science
AI is replacing expensive first-principles simulations with learned surrogates, and generating novel materials and molecular structures through inverse design. Agent systems began connecting to laboratory equipment to execute experiments autonomously in 2025.
Selected notable 2025 developments:
| Type | Name | Affiliation | Highlights |
|---|---|---|---|
| Dataset | OMol25 | Meta FAIR, Los Alamos, UCD | 100M+ DFT calculations across 83 elements |
| Dataset | Multimodal Universe | Polymathic AI, MIT | 100TB astronomical observations |
| Benchmark | ChemBench | Friedrich Schiller Uni / CSIC | 2,700+ Q&A; best models outperform human chemists but struggle on basic tasks |
| Benchmark | PHYBench | Peking University | 500 physics problems; Gemini 2.5 Pro: 36.9% vs human experts: 61.9% |
| Benchmark | ReplicationBench | Stanford, U Toronto | Full astrophysics paper replication; frontier models score <20% |
| Foundation model | GPhyT | RWTH Aachen, U Virginia | 1.8TB training data; up to 29× better than specialised models; zero-shot generalisation |
| Foundation model | AION-1 | UC Berkeley, NYU, Flatiron | First astronomy FM; 200M+ celestial objects; open release |
| Foundation model | MatterGen | Microsoft Research, SIAT | Diffusion-based materials generation; 2× more novel and stable than prior methods |
| AI agent | Physics Supernova | Princeton, Tsinghua, SJTU | IPhO 2025: 23.5/30, ranked 14th of 406 — gold-medalist level |
| AI agent | ChemAgents | USTC, U Birmingham | Autonomous synthesis and optimisation via robotic platform (Llama-3.1-70B) |
Biological and life sciences
Foundation models trained on genomic and evolutionary data expanded from prediction into generative design in 2025. The gap between genomic sequence data (abundant) and functional perturbation data (scarce) remains a core bottleneck.
| Type | Name | Affiliation | Highlights |
|---|---|---|---|
| Dataset | OpenGenome2 | Arc Institute, Stanford, Nvidia | 9.3 trillion base pairs from all domains of life |
| Benchmark | BixBench | FutureHouse, ScienceMachine | 50+ bioinformatics scenarios; GPT-4o and Claude 3.5 Sonnet: ~17% accuracy |
| Benchmark | BioML-bench | Shift Bioscience, Cambridge | End-to-end biomedical ML; agents underperform human baselines on average |
| Foundation model | Evo 2 | Arc Institute, Nvidia, Stanford | 40B params, 1M token context, 9.3T base pairs; fully open release |
| Foundation model | AlphaGenome | Google DeepMind | Predicts thousands of functional measurements from DNA at single base-pair resolution |
| Foundation model | CellFM | Sun Yat-sen Uni | 800M params, 100M human cells; single-cell analysis, perturbation prediction |
| AI agent | Biomni | Stanford, Genentech, Arc Institute | General-purpose biomedical agent across 25 subfields |
| AI agent | BioAgents | UC Berkeley, Microsoft Research | Multiagent system; expert-level on conceptual genomics tasks |
Earth science
Weather forecasting has advanced furthest due to decades of reanalysis datasets (ERA5) and dense global satellite archives. Climate modelling lags because it requires projections beyond the distribution of any existing training data.
| Type | Name | Affiliation | Highlights |
|---|---|---|---|
| Benchmark | UnivEARTH | Cornell, Columbia | 140 Earth observation questions; LLM agents: 33% accuracy; code fails 58% of the time |
| Benchmark | ExEBench | TU Munich | 7 extreme event categories; tests detection, monitoring, and forecasting |
| Foundation model | FourCastNet 3 | Nvidia, LBNL, UC Berkeley | 60-day global forecast in <4 min on a single GPU; 8–60× faster than prior approaches |
| Foundation model | WeatherNext 2 | Google DeepMind | Hundreds of weather outcomes in <1 min/TPU; 99.9% improvement over predecessor |
| Foundation model | TerraMind | IBM Research, ETH Zurich | First any-to-any generative multimodal Earth observation FM; 9 geospatial modalities |
| AI agent | ClimateAgent | HKUST | 85 climate tasks; 100% completion; quality score 8.32 vs Copilot 6.27, GPT-5 3.26 |
| AI agent | PANGAEA GPT | Alfred Wegener Institute | Multi-agent natural language interface for PANGAEA Earth science database |
Aardvark Weather replaced the full numerical weather prediction pipeline with a single ML system in 2025 — the first time an AI system ran a complete end-to-end weather forecasting pipeline.
Mathematics
AI systems are advancing automated formal proof generation (Goedel-Prover in Lean). Competition-level problem-solving improved sharply — Gemini Deep Think earned a gold medal at IMO 2025 (up from silver in a single year). Rapid gains on FrontierMath and MathArena. Major open problems (long-standing Erdős conjectures) remain well beyond current capabilities.
Benchmark reality check: where AI falls short
Across every scientific domain, the pattern is consistent: AI performs well on isolated subtasks but struggles when required to execute the multistep workflows that real research demands.
| Benchmark | Domain | Best AI performance | Human/expert baseline |
|---|---|---|---|
| BixBench | Bioinformatics | ~17% accuracy | — |
| ReplicationBench | Astrophysics (paper replication) | <20% | — |
| UnivEARTH | Earth observation | 33% accuracy; code fails 58% | — |
| PHYBench | Physics | 36.9% (Gemini 2.5 Pro) | 61.9% (human experts) |
| PaperArena | Cross-domain research Q&A | 38.8% (Gemini 2.5 Pro, multi-agent) | 83.5% (PhD experts) |
| AstaBench | End-to-end agentic science | ~0.53/1.0 (best agent, at ~$3.40/problem) | — |
| ChemBench | Chemistry | Outperforms human chemists on average | Struggles with basic tasks |
On PaperArena, multi-agent configurations consistently outperformed single-agent setups by 2–4 percentage points, but all tested agents lagged substantially behind the PhD expert baseline. The best result (Gemini 2.5 Pro multi-agent) achieved 38.8% accuracy against a 83.5% human baseline — less than half.
Implication for institutional guidance: Claims of AI systems “doing research” should be read against these benchmarks. Current capability is strong for well-defined subtasks and weak for the synthesis, judgment, and multi-step execution that constitute actual research practice.
AI as co-scientist: multi-agent research systems
Several research groups released multi-agent systems designed to approximate the structure of a human research team in 2025. These are the most significant development for the autonomous-laboratories and hypothesis-generation capability areas.
google-ai-co-scientist
Google’s AI Co-Scientist (Gottweis et al., 2025) uses a generate–debate–evolve loop in which agents iteratively produce and refine evidence-grounded hypotheses. It was validated in three biomedical areas including AML drug repurposing and liver fibrosis targets. Top-1 accuracy of 78.4% on GPQA Diamond when selecting its highest-rated hypothesis. Not publicly available at time of writing.
sakana-ai-scientist
Sakana AI Scientist-v2 (Yamada et al., 2025) produced the first fully AI-generated paper accepted at a peer-reviewed workshop (ICLR 2025), using agentic tree search to generate and refine code implementations without human-coded templates.
Kosmos
Kosmos (Mitchener et al., 2025) maintained coherence across runs lasting up to 12 hours, executing an average of 42,000 lines of code and reading 1,500 papers per run. Collaborators reported that a single run approximated six months of human research effort. Not yet publicly available.
Experimentally confirmed results: still rare
Despite these advances, only a handful of multi-agent systems have produced results confirmed through real-world experiments. Published examples include:
- New proteins designed by ProtAgents
- 92 antibody candidates for SARS-CoV-2 from the Virtual Lab (>90% successfully bound their target)
- Two new cancer drug targets (GPR160, ARG2) from OriGene
- Five novel metal-organic frameworks from MOFGen
- A novel chromophore from ChemCrow (chemcrow)
Key roadblocks: workforce training gaps, lack of API and interoperability standards, and funding structures that do not support maintenance and scaling of autonomous research infrastructure.
Responsible AI in research
RAI publication trends
Responsible AI papers at major AI conferences grew sharply, reaching approximately 1,756 papers in 2025 across privacy/data governance, fairness/bias, transparency/explainability, and security/safety topics.
Geographic reversal: In 2025, China led responsible AI research output with 812 accepted RAI papers, more than double the 394 from the United States. This reversal is sharp — in 2024, the US led with 788 to China’s 322. The UK contributed 103 papers, ranking 4th.
AI incidents
Documented AI incidents rose to 362 in 2025, up from 233 in 2024. Improving one responsible AI dimension (e.g. safety) has been found to degrade another (e.g. accuracy).
Transparency declining at the frontier
Industry produced over 90% of notable frontier models in 2025, but the most capable models are the least transparent. Training code, parameter counts, dataset sizes, and training duration are no longer disclosed by OpenAI, Anthropic, or Google for their most resource-intensive systems. This creates a structural tension for research institutions relying on these models.
Implication for AI-assisted research practice
As AI is increasingly used in research workflows — from literature synthesis to hypothesis generation — governance frameworks for AI in research are still underdeveloped. The Sakana AI Scientist case (first AI-generated paper at a peer-reviewed workshop) indicates that research integrity frameworks need to address AI authorship explicitly.
Strategic signals for Sheffield
-
Scientific AI is predominantly academic and government-sourced. Most foundation models for science come from cross-sector collaborations involving universities and national labs, in contrast to the industry-dominated landscape of general-purpose AI. Sheffield should monitor academic AI model releases (Evo 2, AION-1, Walrus, GPhyT) and government data infrastructure (FLUXNET, CAMELS, ERA5) rather than defaulting to commercial API products.
-
Performance gaps demand human oversight. End-to-end benchmark results consistently show AI systems performing at 20–50% of expert-level on complete research tasks. Human oversight is not optional — it is structurally required for research quality.
-
Data infrastructure quality is the primary bottleneck. Domains that have advanced furthest (structural biology, weather forecasting, hydrology) are those with the best training data infrastructure. The clearest institutional contribution Sheffield can make is investing in high-quality, well-structured research data.
-
AI PhDs are returning to academia. All of the 22% growth in new AI PhDs (2022–2024) went to academic roles. This is a structural opportunity for research universities: the talent pipeline that had been flowing to industry is reversing. Sheffield should consider how to attract and retain AI researchers.
-
The co-scientist paradigm requires institutional frameworks. Multi-agent research systems (AI Co-Scientist, AI Scientist, Kosmos) are advancing rapidly. Sheffield needs governance frameworks for: AI contributions to publications, AI authorship policy, and the use of AI in grant preparation and hypothesis generation before these systems become more widely accessible.
What this report does not cover
This synthesis deliberately excludes the report’s Economy, Policy & Governance, and Public Opinion chapters. For those areas, consult the full PDF or the Stanford HAI report homepage. The Medicine chapter (Chapter 6) is only lightly referenced here; a separate synthesis focused on clinical AI would be warranted for Sheffield’s Health Sciences faculty.
Sources
- Stanford HAI AI Index Report 2026: https://hai.stanford.edu/ai-index/2026-ai-index-report/science (accessed 2026-04-15)
- Full report PDF:
ai_index_report_2026.pdf(locally available) - Gottweis et al. (2025): Google AI Co-Scientist paper
- Yamada et al. (2025): Sakana AI Scientist-v2 paper
- Mitchener et al. (2025): Kosmos multi-agent system paper