Examples
Examples from published documents
- Anthropic says its CB-1 assessment "also works as a description of the level of risk our systems impose over and above the risks posed by other AI developers' systems" and also "holds for the level of risk that would be imposed industry-wide" (Anthropic Risk Report, August 2026, §4.6.1)
- Google DeepMind measures uplift "relative to a baseline without generative AI" (Google DeepMind Frontier Safety Framework v3.1, footnote 8)
- OpenAI's Tier 2 is "meaningful counterfactual assistance (relative to unlimited access to baseline of tools available in 2021) to novice actors" (OpenAI Frontier Governance Framework, CBRN risk tiers, p.11)
Counter-examples
Not to be confused with
- Uplift measured against tools available in 2021 uses a fixed historical baseline; it is not a marginal basis, even though both compare against something other than zero risk.
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 creep
- Risk creep is the gradual build-up of risk when many models each add only a small marginal increase that crosses no individual threshold, so that total risk grows over time. The Frontier Model Forum names it as a weakness of purely marginal baselines.
- Source: Frontier Model Forum: Risk Taxonomy and Thresholds
- Net new risk
- Net new risk is harm that could not currently be achieved with existing tools and resources without access to frontier or general-purpose AI. OpenAI's and Meta's frameworks use it as a test for which risks to track.
- Source: OpenAI Preparedness Framework v2
- Static historical standard
- A static historical standard fixes a risk baseline at a specific point in time, for example 2023, and never adjusts it, regardless of what other developers later release. The Frontier Model Forum contrasts it with marginal risk assessment.
- Source: Frontier Model Forum: Risk Taxonomy and Thresholds
- Counterfactual assistance
- Counterfactual assistance is help a model provides beyond what a person could already get without it, measured in OpenAI's framework relative to unlimited access to the baseline of tools available in 2021.
- Source: OpenAI Frontier Governance Framework
Source of record
Where this definition comes from
Anthropic Risk Report, August 2026, §1.1
“"Marginal risk": "the level of risk our systems pose over and above the risks posed by other AI developers' systems"; "Absolute risk": "the level of risk that would be posed industry-wide, if all AI developers had models and practices similar to ours" (§1.1).”
https://www-cdn.anthropic.com/f61d49fa5596956a5dec75fea0e973bf6a6a8378/Redacted%20Risk%20Report%20August%202026%20.pdfFrontier Model Forum: Risk Taxonomy and Thresholds, s3.2
“"Marginal Risk Assessments: On the other hand, developers would consider the additional risk their model adds beyond what's already possible in the ecosystem." "they may facilitate gradual risk escalation through "risk creep" – where multiple models, each introducing only small marginal increases which do not cross individual "net new" thresholds, may collectively create significant risk growth over time as these small increases compound." (s3.2)”
https://www.frontiermodelforum.org/technical-reports/risk-taxonomy-and-thresholds/
Crosswalk
How named organisations use this concept
| Organisation | Their term, as published | Match & verification | Source |
|---|---|---|---|
| AnthropicShadow mapping Anthropic Risk Report, August 2026 | “"Marginal risk": "the level of risk our systems pose over and above the risks posed by other AI developers' systems"; "Absolute risk": "the level of risk that would be posed industry-wide, if all AI developers had models and practices similar to ours" (§1.1).” | exactEQ confidence: high | Anthropic Risk Report, August 2026 |
| OpenAIShadow mapping OpenAI Preparedness Framework v2 | “Marginal-risk adjustment if another developer releases a comparable model and "may do so without instituting comparable safeguards to the ones we have committed to" (§4.3). The Safeguards sufficiency determination considers "The baseline risk from other deployments, based on a review of any non-OpenAI deployments of models which have crossed the capability thresholds" (§4.2). The "Net new" criterion compares against "existing tools and resources (e.g., available as of 2021) but without access to frontier AI" (§2.1).” | closeCL confidence: high | OpenAI Preparedness Framework v2 |
| Google DeepMindShadow mapping Google DeepMind Frontier Safety Framework v3.1 | “"(e.g. if other models are similarly capable and have few mitigations, then the marginal risk added by our external deployment is likely low)" (s.1.3.5). Uplift baseline: "we intend this to mean relative to a baseline without generative AI" (fn 8).” | closeCL confidence: high | Google DeepMind Frontier Safety Framework v3.1 |
| MetaShadow mapping Meta Advanced AI Scaling Framework v2 | “"Net new" criterion: "The outcome cannot currently be realized as described (e.g. at that scale / by that threat actor / for that cost) with existing tools and resources but without access to general-purpose AI." (§3.2)” | closeCL confidence: high | Meta Advanced AI Scaling Framework v2 |
| Frontier Model ForumShadow mapping Frontier Model Forum: Risk Taxonomy and Thresholds | “"Marginal Risk Assessments: On the other hand, developers would consider the additional risk their model adds beyond what's already possible in the ecosystem. While dynamic approaches allow responsiveness to changing risk landscapes, they may facilitate gradual risk escalation through "risk creep" – where multiple models, each introducing only small marginal increases which do not cross individual "net new" thresholds, may collectively create significant risk growth over time as these small increases compound." "Static Historical Standards: Under this approach, risk thresholds would be established at a specific point in time (e.g. 2023) and never adjusted" (s3.2).” | exactEQ confidence: high | Frontier Model Forum: Risk Taxonomy and Thresholds |
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.
- EU GPAI Code of Practice
A comparative-baseline exemption mechanism related to, but structurally distinct from, a marginal/absolute risk basis; recorded as a pointer, not a mapping.
EU GPAI Code of Practice, Safety and Security Chapter - California SB 53
A publicly-accessible-information carve-out, conceptually adjacent to a 'baseline of publicly available tools' but stated as a legal exclusion, not a risk-basis definition.
California SB 53
Divergence
Where sources materially disagree
Four different baselines are in active use across the corpus: OpenAI's 2021-tools baseline; Google DeepMind's no-generative-AI baseline; Anthropic's and OpenAI's other-developers'-current-models baseline; and Anthropic's industry-wide hypothetical. Frontier Model Forum names the collective-action failure ('risk creep') that a purely marginal basis permits when every developer measures only against its immediate competitors rather than a fixed historical point. NIKOLAI's crosswalk keeps these four baselines as distinct controlled values rather than treating 'marginal risk' as one interoperable term.







