Skip to main content
v2026.11,986 entries · CC-BY 4.0
CASRAIAsk CASRAI · AI assistant for researchAsk a research question. Answers come with their sources.Funder rules, compliance, methods and publishing. Answers link the sources they rest on.
NIKOLAI elementN5 · Evidence and evaluationsProposednikolai-v0.3

Evaluation

Definition

An AI evaluation is a defined test or study that produces evidence about a model's capabilities, its tendencies to behave in certain ways, or how well its safeguards work, using methods such as benchmarks, red-teaming, uplift studies or simulations.

Also called: Model evaluation, Eval, Capability evaluation

NIKOLAI's proposed operational definition

NIKOLAI proposal: a defined procedure that produces evidence about a model's capabilities, propensities, or safeguard effectiveness, recorded with its purpose, method type, and reporting format as a distinct instance — independent of any single lab's own vocabulary for it.

The operational definition is CASRAI's own proposal, not a definition any named organisation has agreed to. See what NIKOLAI is and is not, or browse the AI safety glossary.

Examples

Examples from published documents

  • The EU Code of Practice lists Q&A sets, task-based evaluations, benchmarks, red-teaming and other adversarial testing, human uplift studies, model organisms, simulations and proxy evaluations for classified materials as example evaluation methods (EU GPAI Code of Practice, Safety and Security Chapter, Measure 3.2)
  • OpenAI pairs automated scalable evaluations, which carry pre-set indicative thresholds, with deep dives such as human expert red-teaming and bio wet lab studies (OpenAI Preparedness Framework v2, section 3.1)
  • Microsoft documents each capability evaluation with its assumptions about the model, the capability evaluated, the method and the results, plus a statement on the method's robustness and any concerns about its validity (Microsoft Frontier Governance Framework, p.6)

Counter-examples

Not to be confused with

  • A threat model describes how harm could happen and guides which evaluations to run, but it produces no evidence about a model by itself.
  • A system card or risk report is a document that reports evaluation results; the evaluation is the procedure that produced them.

Related terms

Related AI safety terms

Terms defined under this element that are not NIKOLAI elements themselves, each as the cited document uses it. Every term is also listed in the AI safety glossary.

Dangerous capability evaluation
A dangerous capability evaluation is a test of whether an AI model can do things that could enable severe harm, such as helping with weapons or cyberattacks, used to judge how close it is to a risk threshold.
Source: Google DeepMind Frontier Safety Framework v3.1
Propensity evaluation
A propensity evaluation measures what an AI model tends to do rather than what it can do, for example its tendency toward deception or other misaligned behaviour, because capability alone does not show whether harm is likely.
Source: Meta Advanced AI Scaling Framework v2
Pre-deployment evaluation
A pre-deployment evaluation is testing of an AI model before release, often including early access for government or independent evaluators, so that the findings can inform the decision to deploy.
Source: GPT-5.6 deployment safety card
Open-ended testing
Open-ended testing is exploratory evaluation without a fixed task list, aimed at finding unexpected behaviours, capability boundaries or emergent properties that predefined evaluations would miss.
Source: EU GPAI Code of Practice, Safety and Security Chapter
Model organism
A model organism, in AI safety, is a model deliberately trained to have a known problem, such as a hidden misaligned behaviour, so that evaluators can check whether their auditing and evaluation methods detect it.
Source: Anthropic Risk Report (August 2026)
Interpretability evidence
Interpretability evidence is information about a model's internal computations, gathered with tools such as natural-language autoencoders or activation steering, used alongside behavioural tests to look for hidden capabilities or misalignment.
Source: Anthropic Risk Report (August 2026)

Source of record

Where this definition comes from

Crosswalk

How named organisations use this concept

Every row below is a shadow mapping. A shadow row is CASRAI's own reading of a published document. No lab, evaluator or regulator named on a shadow row has declared, endorsed, or been consulted on it. That changes only when an organisation files its own Mapping Declaration.
OrganisationTheir term, as publishedMatch & verificationSource
AnthropicShadow mapping
Anthropic Risk Report (August 2026)
“CoBench: "an internal evaluation measuring how well a model, placed at a historical point in Anthropic's infrastructure ..., can diagnose the root causes of issues that Anthropic engineers actually solved" (§3.4.3).”exactEQ
confidence: high
Anthropic Risk Report (August 2026)
AnthropicShadow mapping
Anthropic Advanced AI Framework
“the safety framework must describe "capability evaluations performed" (p.5).”exactEQ
confidence: high
Anthropic Advanced AI Framework
OpenAIShadow mapping
OpenAI Preparedness Framework v2
“"Scalable Evaluations: automated evaluations designed to measure proxies that approximate whether a capability threshold has been crossed." "Deep Dives: designed to provide additional evidence validating the scalable evaluations' findings ..." (§3.1)”exactEQ
confidence: high
OpenAI Preparedness Framework v2
Google DeepMindShadow mapping
Frontier Safety Framework v3.1
“"Early Warning Evaluations: are evaluations which measure the dangerous capabilities of a model. They specifically target the threats and risk scenarios identified through our threat modeling to determine a model's proximity to a CCL or TCL" (glossary).”exactEQ
confidence: high
Google DeepMind Frontier Safety Framework v3.1
xAIShadow mapping
xAI Frontier AI Framework (30 Jun 2026)
“Model evaluations: "state-of-the-art model evaluations relevant to the systemic risk to assess the model's capabilities, propensities, affordances, and/or effects, which may include Q&A sets, task-based evaluations, benchmarks, red-teaming and other methods of adversarial testing, human uplift studies, model organisms, simulations, and/or proxy evaluations for classified materials" (s.2.2(2)).”
This row cites xAI's Frontier AI Framework (30 Jun 2026), whose PDF metadata title reads "Privileged/Confidential DRAFT working FRAMEWORK DOC"; no xAI statement settling draft or final status was found, and the cited URL returned HTTP 404 on 13 Nov 2026. Treat as provisional.
exactEQ
confidence: medium
xAI Frontier AI Framework (30 Jun 2026)
xAIShadow mapping
Grok 4.6 model card
“"safety-threshold evaluations"”exactEQ
confidence: high
Grok 4.6 model card
MetaShadow mapping
Meta Advanced AI Scaling Framework v2
“"Evaluation(s) refers to the assessments we do to understand capabilities and performance. We use this term to describe automated and human evaluations that assess capabilities, as well as evaluations to assess potential for misuse, such as red teaming and uplift studies." (Appendix I, p.43)”exactEQ
confidence: high
Meta Advanced AI Scaling Framework v2
EUShadow mapping
EU GPAI Code of Practice, Safety and Security Chapter
“Measure 3.2: "Signatories will conduct at least state-of-the-art model evaluations in the modalities relevant to the systemic risk to assess the model's capabilities, propensities, affordances, and/or effects", including "open-ended testing ... with a view to identifying unexpected behaviours, capability boundaries, or emergent properties"; Appendix 3.1 requires "internal validity", "external validity" and "reproducibility".”exactEQ
confidence: high
EU GPAI Code of Practice, Safety and Security Chapter
EU / xAI (cross-reference)Shadow mapping
EU CoP Appendix 3 example methods, echoed in xAI FAIF s.2.2(2)
“The example methods in the EU Code of Practice (Measure 3.2 / Appendix 3) are echoed almost verbatim in xAI s.2.2(2): Q&A sets, task-based evaluations, benchmarks, red-teaming, human uplift studies, model organisms, simulations, proxy evaluations for classified materials.”
This document's PDF metadata /Title reads "Privileged/Confidential DRAFT working FRAMEWORK DOC"; no xAI statement was found disambiguating draft vs. final status. Treat as provisional.
exactEQ
confidence: medium
xAI Frontier AI Framework (30 Jun 2026)
California SB 53Shadow mapping
California SB 53
“Large developers summarise "(A) Assessments of catastrophic risks from the frontier model conducted pursuant to the large frontier developer's frontier AI framework" (22757.12(c)(2)).”broadBR
confidence: medium
California SB 53
US Government (NIST CAISI)Shadow mapping
NIST CAISI bulletin
“CAISI "will conduct pre-deployment evaluations and targeted research" (para. 1) — undefined.”noneDU
confidence: medium
NIST CAISI bulletin
Frontier Model ForumShadow mapping
FMF Third-Party Assessments technical report
“"Capability Assessments : the process begins with frontier capability assessments that evaluate whether a model crosses any enabling capability thresholds or outcomes-based thresholds" (1.2).”closeCL
confidence: medium
Frontier Model Forum, Third-Party Assessments
MicrosoftShadow mapping
Microsoft Frontier Governance Framework
“"Capability evaluation: This involves robust evaluation of whether a model possesses tracked capabilities at high or critical levels, including through adversarial testing and systematic measurement using state-of-the-art methods. Evaluations are documented in a consistent fashion setting out the assumptions regarding the model, capability being evaluated, the method used, and evaluation results." (Deeper capability assessment, p.6)”exactEQ
confidence: high
Microsoft Frontier Governance Framework
What do these codes mean?
exact
The source term is equivalent to this element
close
The source term is close but not equivalent to this element
broad
The source term is broader than this element
narrow
The source term is narrower than this element
none
No mapping claim — used for false-friend and declared-but-undefined rows
EQ
Equivalent
CL
Close
BR
Source is broader than the element
NR
Source is narrower than the element
FF
False friend — same or similar label, different meaning
DU
Declared but undefined by the source
UV
Unverified

Related, not mapped

Pointers that are not crosswalk claims

These sources mention this concept but do not define or map it clearly enough to count as a crosswalk row — noted here so the research is visible without overstating it as a mapping.

  • METR

    "Full Capability Elicitation During Evaluations"; "Timing and Frequency of Evaluations" — named as process elements, not an evaluation definition (not a mapping).

    METR

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →