Source of record
Where this definition comes from
Anthropic Risk Report, August 2026, Table 2.4.A discussion
“the brief records the scale as "'very low', 'low', 'moderate/somewhat', 'high'" and notes it is never numerically anchored. CB-2 baseline per-decade probabilities "as high as 1 in 50 or so, or as low as 1 in 20,000".”
https://www-cdn.anthropic.com/f61d49fa5596956a5dec75fea0e973bf6a6a8378/Redacted%20Risk%20Report%20August%202026%20.pdfEU GPAI Code of Practice, Safety and Security Chapter, Measure 3.4
“Signatories will estimate the probability and severity of harm for the systemic risk. ... Estimates will be expressed as a risk score, risk matrix, probability distribution, or in other adequate formats, and may be quantitative, semi-quantitative, and/or qualitative”
https://ec.europa.eu/newsroom/dae/redirection/document/118119
Crosswalk
How named organisations use this concept
| Organisation | Their term, as published | Match | Source |
|---|---|---|---|
| Anthropic Anthropic Risk Report, August 2026 | “Ordinal terms across "probability of occurrence, expected harm conditional on occurrence, fraction unmitigated, and overall risk"; scale "'very low', 'low', 'moderate/somewhat', 'high'", never numerically anchored. CB-2 baseline per-decade probabilities "as high as 1 in 50 or so, or as low as 1 in 20,000".” | close confidence: high | Anthropic Risk Report, August 2026 |
| OpenAI OpenAI Frontier Governance Framework | “Severity and probability are estimated for CBRN, cyber and loss of control (§2.2); "The probability side of 'severity and probability' is not quantified at all" (brief gap).” | none confidence: medium | OpenAI Frontier Governance Framework |
| Google DeepMind Gemini 3.7 Flash FSF Report | “"reasonable confidence" and "moderate confidence" are used for rule-out determinations "but never defined" (pp.3, 13).” | none confidence: medium | Gemini 3.7 Flash FSF Report |
| xAI xAI Frontier AI Framework, 30 Jun 2026 | “"Systemic risk estimation: estimating 'the probability and severity of harm for the systemic risk using state-of-the-art risk estimation methods'" (s.2.2(4)). No scale given.” This row cites xAI's Frontier AI Framework (30 Jun 2026, {FAIF26}), whose PDF metadata /Title reads "Privileged/Confidential DRAFT working FRAMEWORK DOC" with no xAI statement found disambiguating draft vs. final status (ALIGNMENT-MATRIX.md §6 item 4, unresolved as of the source pass). Treat as citing a document of unconfirmed final/draft status. | none confidence: medium | xAI Frontier AI Framework, 30 Jun 2026 |
| EU EU GPAI Code of Practice, Safety and Security Chapter | “Measure 3.4 "Systemic risk estimation": estimates "will be expressed as a risk score, risk matrix, probability distribution, or in other adequate formats, and may be quantitative, semi-quantitative, and/or qualitative", e.g. "a qualitative systemic risk score (e.g. 'moderate' or 'critical')", "a qualitative systemic risk matrix (e.g. 'probability: unlikely' x 'impact: high')", or "a quantitative systemic risk matrix (e.g. 'X-Y%' x 'X-Y EUR damage')" -- the only source in this corpus that explicitly permits and names a quantitative probability-times-damage format alongside qualitative bands.” | close confidence: high | EU GPAI Code of Practice, Safety and Security Chapter |
| International AI Safety Report 2026 (Chair: Yoshua Bengio) International AI Safety Report 2026 | “Describes risks qualitatively -- "already materialising, with documented harms" versus "more uncertain but could be severe if they materialise" -- and frames the problem as an "evidence dilemma"; states "Our global risk management frameworks are still immature, with limited quantitative benchmarks and significant evidence gaps." No numbered likelihood/severity matrix is adopted on the evidence read.” IMPORTANT provenance caveat carried from ALIGNMENT-MATRIX.md: the "immature ... limited quantitative benchmarks" sentence was confirmed (16 Sep 2026 pass) to be a quote from Minister Ashwini Vaishnaw's framing statement in the report's own PR Newswire launch release, NOT the report's analytical body text -- it must not be cited as "the report says" without that distinction. Separately, the ~300-page PDF body (arXiv:2602.21012) could not be parsed in the source pass, so a structured likelihood/severity table in an unread chapter cannot be ruled out; this remains an open [VERIFY] gap, not a settled absence. | none confidence: medium | International AI Safety Report 2026 |
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.
- California SB 53
RL -- an undefined qualitative standard in statute, a pointer rather than a controlled term mapping.
California SB 53 - Frontier Model Forum
RL -- framing commentary on the state of the field's likelihood quantification, not itself a likelihood-term scale.
Frontier Model Forum: Risk Taxonomy and Thresholds - CISA (U.S. Cybersecurity and Infrastructure Security Agency)
RL -- cited in the source brief via an internal research file, not a public URL in the supplied sources table; recorded as a pointer only, no URL invented.
Gap
No source anchors its likelihood words. NIKOLAI should require a `likelihood.scheme` reference and allow "unanchored" as an explicit value. This is no longer only an absence in the corpus NIKOLAI happened to read -- the International AI Safety Report 2026 independently confirms it is an absence in the field itself (subject to the minister-attribution caveat above), which strengthens rather than weakens the case for NIKOLAI's proposed `likelihood.scheme` field.
