GCRI · Artificial Intelligence and Autonomous Systems
AI in science
Machine learning now predicts protein structures, emulates weather and climate models, screens materials and helps analyse vast datasets. It can speed discovery, and it can also produce results that are hard to reproduce, overfit or rest on data leakage. This page sets out how AI is used in science, where it fails, how its results are checked, and what evidence serves the researchers, funders and journals involved.
AI in science in seven parts.
- 01The questionAI in science: how the risk develops
- 02Across the domainsAI changes how each domain’s science is done
- 03IndicatorsWhat is measured, and how well
- 04SourcesWhere the evidence comes from
- 05DecisionsWho decides, and what the evidence serves
- 06Core questionsWhat scientists ask about AI in research
- 07MechanismsHow discovery and error both scale
AI in science: how the risk develops.
Discovery and reproducibility. The dossier below sets out how the risk develops, what is measured, where the data come from, how it is modelled and who decides.
Learning from data replaces some simulation
Models trained on simulations or observations can emulate expensive calculations quickly. They are reliable within the range of their training data and uncertain outside it.
Leakage inflates results
When information from the test set reaches training, through duplicates, time overlap or preprocessing, reported accuracy overstates real performance. Leakage is a known cause of irreproducible ML science.
Openness enables checking
Results can be verified only when code, data, models and settings are shared. Proprietary models and data limit independent checks.
Scientists stay accountable
AI can suggest hypotheses and draft text, but researchers remain responsible for methods, interpretation and integrity. Journals and funders are setting rules for AI use.
Model registry.
The screen a team would use to follow this subject: its measures, states and decisions. Names are examples and values are illustrative.
Model registry
AI in science · Model registry · Release gate
Models
| M-489 triage assistant | Public service | High | 20 of 24 suites | Blocked |
| M-530 flood nowcast | Science | Medium | 12 of 12 | Released |
| M-079 document search | Internal | Low | 8 of 8 | Released |
| M-881 agent planner | Agentic | High | 11 of 16 | Evaluating |
Model card
- Intended use
- triage support, human decides
- Out of scope
- eligibility decisions
- Known limits
- lower accuracy on rare cases
- Owner
- Model owner
Incident log
- D-19INCHallucinated citation in summary
- D-8MITRetrieval grounding added
- D-3INCPrompt injection via uploaded file
- D-1MITTool permission narrowed
Evaluation scores
vs release threshold
AI changes how each domain’s science is done.
- WaterCouplingML models now rival process models in some prediction tasks.What to watchBenchmarks against process models
- HealthCouplingAI supports protein and pathogen research, raising dual-use questions.What to watchBiosecurity review of tools
- BiodiversityCouplingAutomated species identification scales monitoring.What to watchIdentification accuracy by region
Studies of AI in science are designed for the Methods and Technical series. Reproducibility requires code, data and model versions, recorded under NS-12 of the proposed Nexus Standards.
- Replications of AI-based results.
- Benchmarks with documented data.
How AI in science is modelled, measured and tested.
The model, its assumptions, its data and its uncertainty, in the form a scientist or engineer would check. Structures follow published methods. The values drawn are illustrative, not findings.
Model, equations and assumptions
The model structure used for AI in science, the assumptions it makes and how each is tested.
Ri = f(Hi, Ei, Vi, Ci)hazard, exposure, vulnerability and coping capacityEAL = ∑k pk · Lkover events k with annual probability pkxt+1 = A xt + B ut + εtA carries the couplings, u the shock, ε the error| Assumption | Why it is made | How it is tested |
|---|---|---|
| Hazard and exposure are independent within an event | Keeps the loss integral tractable | Compare with a copula fit on joint records |
| Couplings are linear near the operating point | Lets shocks propagate through A | Perturb beyond the linear range and compare |
| Vulnerability curves transfer between places | Local curves are rarely available | Validate against observed losses where they exist |
- Inputs
- versioned datasets with provenance
- Outputs
- distributions, never single numbers
- Code
- released with the model card and tests
- Reproduction
- same data, code, environment and seed
What is measured, and how well.
Each indicator carries its unit, its cadence, the kind of source behind it and the evidence label a value would carry. No values are shown: a programme supplies them under its own authority.
| Indicator | Unit | Cadence | Source type | Label |
|---|---|---|---|---|
| Studies sharing code and data | percent of published ML studies | per review | Publication audits | OBSERVED |
| Leakage issues identified in reviews | count by type | per review | Reproducibility studies | REPORTED |
| Out-of-distribution performance | error on held-out regimes | per evaluation | Model evaluation | OBSERVED |
| Emulator error against reference simulation | error metric | per evaluation | Validation runs | OBSERVED |
| Journals with AI use policies | count | annual | Publisher policies | DECLARED |
| Compute used for scientific model training | floating point operations | per model | Model documentation | ESTIMATED |
Where the evidence comes from.
Named public datasets and institutional records, each with its licence class. Sources by agreement are used only under their terms and never republished.
| Source | What it gives | Licence class |
|---|---|---|
| Protein Data Bank and AlphaFold Protein Structure Database | Experimental and predicted protein structures | Open, CC0 and CC BY 4.0 respectively |
| WeatherBench and related benchmarks | Standard evaluation of data-driven weather models | Open, licence per benchmark (to check) |
| Princeton reproducibility studies on ML-based science | Survey of leakage and reproducibility failures | Licence per publication |
| Materials Project | Computed materials properties for screening | Open, CC BY 4.0 |
| Publisher and funder policies on AI use | Rules on authorship, disclosure and review | Public, terms per publisher |
Who decides, and what the evidence serves.
Artificial Intelligence and Autonomous Systems prepares evidence. The decisions below are taken by the authorities named, under their own rules, and never by the platform.
| Who decides | The question | Evidence this page serves |
|---|---|---|
| Research teams | Whether an ML result is robust enough to publish | Leakage checklists; validation protocols |
| Journals and reviewers | What evidence to require for ML-based claims | Reporting guidance; reproducibility evidence |
| Research funders | How to fund open models, data and compute for science | Openness and compute evidence |
Outputs this topic produces
- Leakage and reproducibility checklists for ML studies
- Validation protocols for scientific emulators
- Guidance on reporting AI use in research
- Case studies of AI in specific sciences
What scientists ask about AI in research.
- Does the AI-assisted result hold when tested independently, by experiment or observation?
- Is the model, data and code available so others can reproduce it?
- Has data leakage between training and evaluation inflated reported performance?
- Does the model extrapolate beyond the conditions its training data covered?
How discovery and error both scale.
Machine learning finds patterns in large datasets faster than people can, and surrogate models can emulate expensive simulations, allowing many more scenarios to be explored. In fields with good data and clear validation, such as structural biology, the gains have been substantial.
The risks are equally systematic. Leakage between training and test data has produced inflated results across many disciplines. Models trained on past data may not respect physical laws when extrapolating. Generative systems can fabricate plausible references or data. And when models and data are proprietary, results cannot be checked.
Keeping AI-assisted science reproducible.
- 01
Reporting standards
Checklists covering data splits, leakage checks, baselines and uncertainty. Standards help only if journals and reviewers enforce them.
- 02
Physics-informed and hybrid models
Combining learned components with known physical constraints, using artificial intelligence and machine learning and high-performance, cloud and edge computing. Constraints reduce but do not remove extrapolation risk.
- 03
Uncertainty estimation
Ensembles and Bayesian methods from probability and statistics to flag predictions the model is unsure of. Uncertainty estimates can themselves be miscalibrated.
- 04
Open artefacts
Publishing code, data and model weights where rights allow, so others can reproduce and challenge results.
How it is studied, and what the methods cannot say.
- 01Artificial intelligence and machine learningDistribution shift analysis · Fine-tuning for a specific use · Interpretability methodsApplied here
- 02Probability and statisticsUncertainty on benchmark scores · Calibration assessment · Audit samplingApplied here
- 03High-performance, cloud and edge computingCompute access for public research · Energy measurement of workloadsApplied here
- 04AI evaluation and assurance engineeringAssurance cases for AI systems · Structured red teaming · Post-deployment monitoringApplied here
Where this page’s references lead.
The bodies and documents this page is read against. Status and editions change; each link goes to the issuing body.
Questions and answers.
What is data leakage in machine learning?
Can AI replace scientific simulation?
How should AI use in research be reported?
Can AI replace experiments?
What is data leakage?
Does the platform review scientific papers?
Are AI-generated hypotheses credited?
Contribute evidence from this page.
Reports is the Consortium’s quarterly publication. What you know about this subject can become evidence, a case or a correction in the next edition, credited to you and published under CC BY 4.0 where it is your own work.
- EmailGateway GW-GCRISend it from this pageEvidence, a correction, a case, a method, a review, a translation or profile input. It arrives tagged with this page, so the editor who owns it receives it.
- ZenodoOpen depositDeposit it and submit to the communityPublish your own record on Zenodo with your ORCID iD, then submit it to the Reports community there. Curators review it; the DOI stays yours.
- ReportsGCRISee every way to contributeThe contribution types, how each is reviewed, what credit you receive, the licence and what is never accepted.
Never send personal data, sensitive infrastructure detail or anything you have no right to share. A contribution is reviewed before use; sending it confers no authorship or editorial rights, and sponsorship never buys a finding.
Bring a method, a dataset or a question.
Information and enquiries are open today. An enquiry receives a reference and names the next step; applications, payments, compute allocations and registrations open only when their team, terms and review capacity exist.
Legal documents.
Maps, data and figures on this site are shown for context and general information. They are not advice, forecasts or warnings; the limits of the information are set out in the Disclaimers and Data Notices.