Skip to content

GCRI · Inteligencia artificial y sistemas autónomos

Evaluación de modelos

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.

02La pregunta

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.

Mecanismo

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.

Mecanismo

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.

Mecanismo

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.

Mecanismo

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.

03Seguimiento

Registro de modelos.

La pantalla que utilizaría un equipo para seguir esta materia: sus mediciones, estados y decisiones. Los nombres son ejemplos y los valores son ilustrativos.

Registro de modelos

Model evaluation · Model registry · Release gate

Un modelo bloqueado en la puertaGCRI · Inteligencia artificial y sistemas autónomos

Modelos

M-184 triage assistantServicio públicoAlto19 of 24 suitesBloqueado
M-320 flood nowcastCienciaMedio12 de 12Publicado
M-157 document searchInternoBajo8 of 8Publicado
M-563 agent plannerAgénticoAlto12 of 16En evaluación

Ficha del modelo

Uso previsto
apoyo al triaje, decide una persona
Fuera de alcance
decisiones de elegibilidad
Límites conocidos
menor precisión en casos raros
Propietario
Propietario del modelo

Registro de incidentes

  1. D-21INCCita alucinada en el resumen
  2. D-9MITFundamentación de la recuperación añadida
  3. D-2INCInyección de instrucciones mediante un archivo cargado
  4. D-1MITPermiso de herramienta restringido

Evaluation scores

vs release threshold

Task accuracy87 of 85 requiredRobustez80 of 75 requiredFairness gap (lower is better)6 of 5 requiredRefusal on misuse85 of 95 required
Model registry: the screen’s structure, terms and states for Model evaluation.Nombres de ejemplo y valores ilustrativos; un diseño, no un registro ni un sistema en vivo.
04En todos los dominios

Evaluation must reflect the domain where a model is used.

Evaluations are designed for the serie Labs Evidence. 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.
05Métodos y evidencia

How model evaluation is modelled, measured and tested.

El modelo, sus supuestos, sus datos y su incertidumbre, en la forma en que un científico o un ingeniero los comprobaría. Las estructuras siguen métodos publicados. Los valores representados son ilustrativos, no hallazgos.

Modelo, ecuaciones y supuestos

The model structure used for model evaluation, the assumptions it makes and how each is tested.

Riesgo del elemento iRi = f(Hi, Ei, Vi, Ci)amenaza, exposición, vulnerabilidad y capacidad de afrontamiento
Pérdida anual esperadaEAL = ∑k pk · Lksobre los eventos k con probabilidad anual pk
Sistemas acopladosxt+1 = A xt + B ut + εtA carries the couplings, u the shock, ε the error
SupuestoPor qué se haceCómo se prueba
La amenaza y la exposición son independientes dentro de un eventoMantiene manejable la integral de pérdidasComparar con un ajuste de cópula sobre registros conjuntos
Los acoplamientos son lineales cerca del punto de operaciónPermite que las perturbaciones se propaguen a través de APerturbar más allá del rango lineal y comparar
Las curvas de vulnerabilidad se transfieren entre lugaresLas curvas locales rara vez están disponiblesValidar frente a las pérdidas observadas cuando existan
Entradas
conjuntos de datos versionados con procedencia
Resultados
distribuciones, nunca cifras únicas
Código
publicado con la ficha de modelo y las pruebas
Reproducción
mismos datos, código, entorno y semilla
Formulaciones estándar de riesgo y de espacio de estados; los supuestos enumerados son los típicos del método, expuestos para su revisión, no resultados.
Principios rectores FAIR · GO FAIR y FORCE11W3C PROV · World Wide Web ConsortiumOrientación para un tratamiento coherente de las incertidumbres · IPCCCEOS Analysis Ready Data · Comité sobre Satélites de Observación de la Tierra
06Indicadores

Qué se mide y con qué calidad.

Cada indicador lleva su unidad, su periodicidad, el tipo de fuente que lo respalda y la etiqueta de evidencia que llevaría un valor. No se muestran valores: un programa los aporta bajo su propia autoridad.

IndicadorUnidadCadenciaTipo de fuenteEtiqueta
Benchmark score with uncertaintyscore with confidence intervalper evaluationPublished benchmark resultsNOTIFICADO
Contamination checks reportedpercent of evaluations reporting checksper evaluationEvaluation reportsDECLARADO
Red-team findings by severitycount by severity classper exerciseRed-team reportsNOTIFICADO
Performance gap between subgroupsdifference in error ratesper evaluationDisaggregated evaluationOBSERVADO
Post-deployment incident reportscount by typemensualIncident monitoringNOTIFICADO
Error de calibraciónexpected calibration errorper evaluationEvaluation on held-out dataOBSERVADO
07Fuentes

De dónde procede la evidencia.

Conjuntos de datos públicos y registros institucionales identificados, cada uno con su tipo de licencia. Las fuentes obtenidas por acuerdo solo se utilizan conforme a sus condiciones y nunca se republican.

FuenteLo que aportaTipo de licencia
Stanford HELM (Holistic Evaluation of Language Models)Open, multi-metric evaluations of language modelsOpen source, Apache 2.0 for code; results public
MLCommons benchmarks, including AILuminateStandardised performance and safety benchmarksPublic results, MLCommons terms (to check)
Informe internacional sobre seguridad de la IASynthesis of evidence on capabilities, risks and evaluation limitsPublic, terms per publisher
Model cards and system cards from developersDeveloper-reported evaluations and limitationsPublic documents, terms per developer
Epoch AI benchmark and model databasesCompiled benchmark results and model characteristicsAbierta, CC BY 4.0
08Decisiones

Quién decide y a qué sirve la evidencia.

Estudio de Inteligencia artificial y sistemas prepares evidence. The decisions below are taken by the authorities named, under their own rules, and never by the platform.

Quién decideLa preguntaEvidencia a la que sirve esta página
DesarrolladoresWhether a model is ready for release and what limits to stateEvaluation protocols; red-team findings
Deploying organisationsWhether a model is fit for a specific useUse-specific evaluations; subgroup performance
Regulators and AI safety institutesWhat evaluation evidence to requireEvaluation standards and reporting templates

Productos que genera este tema

  • Evaluation protocols with uncertainty and contamination checks
  • Disaggregated evaluation designs for subgroup performance
  • Red-team methods for public sector use cases
  • Evaluation reporting templates
09Cuestiones centrales

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)?
10Mecanismos

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 cambio de distribución, 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.

11Métodos y datos

How evaluation is done, and its limits.

  1. 01

    Held-out and out-of-distribution testing

    Testing on data deliberately different from training, analysed with probabilidad y estadística. No test set covers every deployment condition.

  2. 02

    Calibration analysis

    Comparing stated confidence with observed accuracy, and checking how it changes under shift.

  3. 03

    Disaggregated evaluation

    Reporting performance by subgroup and task rather than only on average. Small subgroups give uncertain estimates.

  4. 04

    Human and expert evaluation

    Structured assessment of open-ended outputs, using ciencias sociales y del comportamiento methods for rater agreement. Human judgement is costly and can be inconsistent.

13Where this page’s references

Where this page’s references lead.

Los organismos y documentos con los que se lee esta página. El estado y las ediciones cambian; cada enlace lleva al organismo emisor.

15Preguntas

Preguntas y respuestas.

Why are benchmark scores not enough?
Because models can be trained on test items, benchmarks saturate, and test conditions differ from real use. Use-specific, adversarial and post-deployment evaluation are needed as well.
What is red teaming in AI?
Structured adversarial testing in which people or automated tools try to make a model produce harmful, insecure or unreliable outputs, to find failures before deployment.
What is data contamination?
The presence of benchmark test items in a model’s training data, which inflates scores and makes results unreliable as measures of general capability.
Is a high benchmark score evidence a model is safe to deploy?
No. It shows performance on that test under its conditions. Fitness for a use needs evaluation on representative data, with failure analysis and monitoring after deployment.
What is benchmark saturation?
When models reach near-perfect scores, the benchmark no longer discriminates between them, and further gains may reflect overfitting to the test rather than real capability.
Why evaluate after deployment too?
Because data, users and the model itself can change. Monitoring detects drift and new failure modes that pre-release testing could not anticipate.
Will the platform publish evaluation protocols?
Protocols are designed to appear on modelos and in publications once reviewed, each with its scope and limits.
serie ReportsContribuir

Aporte evidencia desde esta página.

serie 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.

GCRIInteligencia artificial y sistemas autónomosTecnología exponencialEvidencia de Labs
Inteligencia artificial y sistemas autónomos

Aporte un método, un conjunto de datos o una pregunta.

La información y las consultas están abiertas hoy. Una consulta recibe una referencia e indica el siguiente paso; las solicitudes, pagos, asignaciones de cómputo y registros solo se abren cuando existen su equipo, sus condiciones y la capacidad de revisión.