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 elementN4 · Claims and argumentProposednikolai-v0.3

Risk Level

Definition

A risk level is the overall grade, such as low, high or critical, given to the risk a model or AI activity poses after weighing likelihood, severity and safeguards. It differs from a capability level, which grades only what a model can do.

Also called: Risk tier, Risk threshold, Systemic risk tier, Overall risk assessment, Model risk score

NIKOLAI's proposed operational definition

NIKOLAI proposes Risk Level as the overall graded risk assigned to a model, domain, or company activity as a whole -- distinct from a capability threshold or capability level, which describes what a model can do rather than the net risk after mitigations. This is a NIKOLAI editorial proposal made necessary by a genuine label collision in the corpus: several sources use identical words ('High', 'Critical') for capability levels in one place and risk levels in another.

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

  • Anthropic grades misalignment risk "Low (an increase from our previous assessment of "very low," in light of general increased uncertainty around recent incident disclosures related to model behavior in cybersecurity evaluations)" (Anthropic Risk Report, August 2026, Table 1.2.A)
  • Meta assigns models to "Critical", "High" or "Moderate or lower" risk thresholds, mapped to "Develop with Mitigations", "Deploy with mitigations" and "Deploy" (Meta Advanced AI Scaling Framework v2, Table 1)
  • Microsoft assesses whether a capability "poses a low, medium, high, or critical risk to national security or public safety" (Microsoft Frontier Governance Framework, Feb 2026, p.3)

Counter-examples

Not to be confused with

  • OpenAI's "High" and "Critical" are capability thresholds, not risk levels: they grade what a model can do before safeguards, so "High" in OpenAI's framework and "High" in Meta's are different measurements.
  • A severity threshold defines how bad a harm must be to count as catastrophic; a risk level grades the overall risk, combining severity with likelihood and mitigation.

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.

Risk tier
A risk tier is a graded level that concretely describes what an AI system might help someone do, or do on its own, that could meaningfully increase the risk of severe harm. OpenAI's Frontier Governance Framework uses Tiers 1 to 3.
Source: OpenAI Frontier Governance Framework
Systemic risk tier
A systemic risk tier is a measurable level of systemic risk defined mainly by model capabilities. The EU GPAI Code requires providers to define tiers for each identified risk, including at least one tier the model has not reached.
Source: EU GPAI Code of Practice, Safety and Security Chapter
Model risk score
A model risk score is NVIDIA's 1-to-5 grade for a model, set by the highest score across intended use case, capabilities and level of autonomy, each criterion having discrete thresholds between 1 and 5.
Source: NVIDIA Frontier AI Risk Assessment
Qualitative risk score
A qualitative risk score expresses an estimated risk in words, such as "moderate" or "critical", instead of numbers. The EU GPAI Code allows it alongside risk matrices and probability distributions.
Source: EU GPAI Code of Practice, Safety and Security Chapter

Source of record

Where this definition comes from

  • Anthropic Risk Report, August 2026, Table 1.2.A, §4.6.1, §5.4

    “Overall risk assessment: "Low (an increase from our previous assessment of "very low," in light of general increased uncertainty around recent incident disclosures related to model behavior in cybersecurity evaluations)." (Table 1.2.A); "we now also assess the risk posed by our models in February 2026 as low, rather than the "very low" description we gave at the time in our previous risk report" (§4.6.1); "Risks from autonomy-related threats are presently low" and "Risks from chemical and biological weapons are presently low." (§5.4)”

    https://www-cdn.anthropic.com/f61d49fa5596956a5dec75fea0e973bf6a6a8378/Redacted%20Risk%20Report%20August%202026%20.pdf
  • Meta Advanced AI Scaling Framework v2, Table 1

    “"Critical", "High", "Moderate or lower" risk thresholds with measures "Develop with Mitigations", "Deploy with mitigations", "Deploy" (Table 1).”

    https://ai.meta.com/static-resource/Meta_Advanced-AI-Scaling-Framework-v2

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 / Anthropic Advanced AI Framework
“Overall risk assessment for misalignment: "Low (an increase from our previous assessment of "very low," ...)" (Table 1.2.A); for automated R&D: "Low." (Table 1.2.B); for CB-1: "Low, but higher than our previous estimate" and for CB-2: "Low risk, but with substantial uncertainty." (Table 1.2.C). CB-1 reassessment: "we now also assess the risk posed by our models in February 2026 as low, rather than the "very low" description we gave at the time in our previous risk report" (§4.6.1).”exactEQ
confidence: high
Anthropic Risk Report, August 2026
AnthropicShadow mapping
Anthropic Advanced AI Framework
“Risk reports should offer "an overall assessment of each of the Enumerated Risks posed by the developer's models" (p.6); external evaluators address "whether the evaluator disagrees with any of the report's key claims, especially its overall assessment of the level of risk for each Enumerated Risk category" (p.8).”exactEQ
confidence: high
Anthropic Advanced AI Framework
OpenAIShadow mapping
OpenAI Frontier Governance Framework / Preparedness Framework v2
“FGF "Risk tiers" with Tier 1 to Tier 3 for cyber, CBRN and loss of control (§2.4). PF v2 retired "low" and "medium" "because those levels were not operationally involved in the execution of our Preparedness work" (App. A).”closeCL
confidence: high
OpenAI Frontier Governance Framework
xAIShadow mapping
xAI Frontier AI Framework, 30 Jun 2026
“"we review the risk tiers for each systemic risk category" (s.2.3). Not defined anywhere.”
This row cites xAI's Frontier AI Framework (30 Jun 2026), whose PDF metadata /Title reads "Privileged/Confidential DRAFT working FRAMEWORK DOC" with no xAI statement found disambiguating draft vs. final status. Treat as citing a document of unconfirmed draft/final status.
noneDU
confidence: medium
xAI Frontier AI Framework, 30 Jun 2026
MetaShadow mapping
Meta Advanced AI Scaling Framework v2
“"Critical", "High", "Moderate or lower" risk thresholds with measures "Develop with Mitigations", "Deploy with mitigations", "Deploy" (Table 1).”exactEQ
confidence: high
Meta Advanced AI Scaling Framework v2
EUShadow mapping
EU GPAI Code of Practice, Safety and Security Chapter
“Same tiers/estimates machinery as B4/C3 (systemic risk tiers, Measure 4.1; qualitative or quantitative estimate, Measure 3.4); the chapter gives illustrative labels only ("moderate", "critical") rather than a fixed scale.”closeCL
confidence: high
EU GPAI Code of Practice, Safety and Security Chapter
MicrosoftShadow mapping
Microsoft Frontier Governance Framework, Feb 2026
“"low, medium, high, or critical risk to national security or public safety" (p.3).”exactEQ
confidence: high
Microsoft Frontier Governance Framework, Feb 2026
NVIDIAShadow mapping
NVIDIA Frontier AI Risk Assessment
“A numeric "model risk (MR) score between 1 and 5 based on the highest MR score within this criteria" (p.2) across three criteria -- intended use case, capabilities, level of autonomy -- each with "discrete thresholds between 1 and 5"; "correlated to the maximum permissible harm relative to our trustworthy AI principles" (p.2).”exactEQ
confidence: high
NVIDIA Frontier AI Risk Assessment
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.

Divergence

Where sources materially disagree

False-friend warning: OpenAI's 'High' and 'Critical' are capability thresholds, not risk levels, in OpenAI's own other vocabulary (see the capability-threshold element). Meta's and Microsoft's 'High'/'Critical' above ARE risk levels, and Anthropic's 'low' above is a risk level, not a capability level. A naive string match across these sources on the word 'Critical' or 'High' alone would silently conflate a capability threshold with a risk level; NIKOLAI's crosswalk keeps these element types distinct precisely to prevent that collision.

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 →