GCRI · Artificial Intelligence and Autonomous Systems
Model evaluation
Model evaluation asks what an AI model can do, how reliably, and where it fails. Benchmarks, red teaming, human studies and field monitoring each measure something different, and none fully predicts behaviour in deployment. This page sets out how evaluations are designed and where they mislead, how results should be reported, and what evidence serves the developers, deployers and regulators who rely on them.
Model evaluation in seven parts.
- 01The questionModel evaluation: how the risk develops
- 02Across the domainsEvaluation must reflect the domain where a model is used
- 03IndicatorsWhat is measured, and how well
- 04SourcesWhere the evidence comes from
- 05DecisionsWho decides, and what the evidence serves
- 06Core questionsWhat evaluators ask
- 07MechanismsWhy scores mislead
Model evaluation: how the risk develops.
Validity, calibration and generalisation. The dossier below sets out how the risk develops, what is measured, where the data come from, how it is modelled and who decides.
Benchmarks saturate and leak
Models quickly reach high scores on fixed benchmarks, and test items can leak into training data. High scores then say more about the benchmark than about general capability.
Test conditions differ from use
Evaluations use set prompts, settings and tasks. Real users phrase things differently, combine tools and face new situations, so measured performance may not transfer.
Adversarial testing finds hidden failures
Red teaming and adversarial prompts reveal unsafe or unreliable behaviour that average-case metrics miss. Coverage depends on the testers’ skill and the time allowed.
Performance drifts after deployment
Model updates, changing inputs and new uses alter behaviour over time. Without monitoring, a model evaluated at release can behave differently months later.
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
Model evaluation · Model registry · Release gate
Models
| M-184 triage assistant | Public service | High | 19 of 24 suites | Blocked |
| M-320 flood nowcast | Science | Medium | 12 of 12 | Released |
| M-157 document search | Internal | Low | 8 of 8 | Released |
| M-563 agent planner | Agentic | High | 12 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-21INCHallucinated citation in summary
- D-9MITRetrieval grounding added
- D-2INCPrompt injection via uploaded file
- D-1MITTool permission narrowed
Evaluation scores
vs release threshold
Evaluation must reflect the domain where a model is used.
- HealthCouplingClinical AI must be evaluated on the populations it serves.What to watchPerformance by subgroup
- WaterCouplingHydrological ML must be tested out of sample in new basins.What to watchPerformance in ungauged basins
- FoodCouplingCrop monitoring models need local ground truth.What to watchAccuracy against field data
Evaluations are designed for the Labs Evidence series. The International AI Safety Report (2025) describes the limits of current evaluations; NIST’s AI RMF and its generative AI profile (2024) set measurement practice. Models would carry cards under NS-12 of the proposed Nexus Standards.
- Held-out test sets for a domain.
- Evaluation protocols with known failure cases.
How model evaluation 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 model evaluation, 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 |
|---|---|---|---|---|
| Benchmark score with uncertainty | score with confidence interval | per evaluation | Published benchmark results | REPORTED |
| Contamination checks reported | percent of evaluations reporting checks | per evaluation | Evaluation reports | DECLARED |
| Red-team findings by severity | count by severity class | per exercise | Red-team reports | REPORTED |
| Performance gap between subgroups | difference in error rates | per evaluation | Disaggregated evaluation | OBSERVED |
| Post-deployment incident reports | count by type | monthly | Incident monitoring | REPORTED |
| Calibration error | expected calibration error | per evaluation | Evaluation on held-out data | OBSERVED |
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 |
|---|---|---|
| Stanford HELM (Holistic Evaluation of Language Models) | Open, multi-metric evaluations of language models | Open source, Apache 2.0 for code; results public |
| MLCommons benchmarks, including AILuminate | Standardised performance and safety benchmarks | Public results, MLCommons terms (to check) |
| International AI Safety Report | Synthesis of evidence on capabilities, risks and evaluation limits | Public, terms per publisher |
| Model cards and system cards from developers | Developer-reported evaluations and limitations | Public documents, terms per developer |
| Epoch AI benchmark and model databases | Compiled benchmark results and model characteristics | Open, CC BY 4.0 |
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 |
|---|---|---|
| Developers | Whether a model is ready for release and what limits to state | Evaluation protocols; red-team findings |
| Deploying organisations | Whether a model is fit for a specific use | Use-specific evaluations; subgroup performance |
| Regulators and AI safety institutes | What evaluation evidence to require | Evaluation standards and reporting templates |
Outputs this topic produces
- Evaluation protocols with uncertainty and contamination checks
- Disaggregated evaluation designs for subgroup performance
- Red-team methods for public sector use cases
- Evaluation reporting templates
What evaluators ask.
- Does the test measure the capability or property that matters for the intended use (construct validity)?
- When the model reports high confidence, is it right as often as that confidence implies (calibration)?
- Does performance hold on data from different places, times or populations (generalisation)?
- Has test data leaked into training, inflating results (contamination)?
Why scores mislead.
Models learn statistical regularities in their training data, including spurious ones. A medical imaging model may learn the hospital’s scanner signature rather than the disease; a language model may recognise benchmark questions it saw during training. When deployment data differ from test data, a problem known as distribution shift, performance can fall sharply and without warning.
Large general-purpose models add new difficulties. Their capabilities are broad and emergent, outputs vary with phrasing, and the line between capability and memorisation is blurred. Aggregate scores can hide failures concentrated in particular groups or tasks.
How evaluation is done, and its limits.
- 01
Held-out and out-of-distribution testing
Testing on data deliberately different from training, analysed with probability and statistics. No test set covers every deployment condition.
- 02
Calibration analysis
Comparing stated confidence with observed accuracy, and checking how it changes under shift.
- 03
Disaggregated evaluation
Reporting performance by subgroup and task rather than only on average. Small subgroups give uncertain estimates.
- 04
Human and expert evaluation
Structured assessment of open-ended outputs, using social and behavioural science methods for rater agreement. Human judgement is costly and can be inconsistent.
How it is studied, and what the methods cannot say.
- 01AI evaluation and assurance engineeringAssurance cases for AI systems · Structured red teaming · Post-deployment monitoringApplied here
- 02Probability and statisticsUncertainty on benchmark scores · Calibration assessment · Audit samplingApplied here
- 03Artificial intelligence and machine learningDistribution shift analysis · Fine-tuning for a specific use · Interpretability methodsApplied here
- 04Social and behavioural scienceHuman oversight studies · Affected community research · Workplace adoption studiesApplied 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.
Where this page leads across the Consortium.
Questions and answers.
Why are benchmark scores not enough?
What is red teaming in AI?
What is data contamination?
Is a high benchmark score evidence a model is safe to deploy?
What is benchmark saturation?
Why evaluate after deployment too?
Will the platform publish evaluation protocols?
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.