Examples
Examples from published documents
- Google DeepMind found that "Expert red teaming demonstrated a modest capability uplift over web baselines" for Gemini 3.7 Flash, enough to reach the alert threshold for the CBRN uplift CCL but not the CCL itself (Gemini 3.7 Flash FSF report, p.3)
- Microsoft's High CBRN level is met when "The model provides a meaningful uplift to a medium-skilled actor's (e.g., STEM education) ability to develop and deliver a known CBRN threat" (Microsoft Frontier Governance Framework, Appendix I)
- Anthropic's threat intelligence team assesses uplift in observed misuse cases "through the lens of speed, scale, and depth" (Anthropic, Detecting and countering misuse of AI: September 2026)
Counter-examples
Not to be confused with
- An uplift study is not uplift itself but one way to measure it: Meta defines uplift studies as experiments comparing people with AI access to "a control group that only has access to existing resources". NIKOLAI records the method as an evaluation type.
- A raw benchmark score is not an uplift estimate: it shows what the model can do, not how much it adds to what a given actor could already do.
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.
- Novice uplift
- Novice uplift is help that lets someone without deep expertise do what they otherwise could not. OpenAI's Preparedness Framework v2 sets its High biological and chemical threshold at meaningful assistance to "novice" actors, "anyone with a basic relevant technical background".
- Source: OpenAI Preparedness Framework v2
- Expert uplift
- Expert uplift is help that lets an already skilled specialist go further, such as developing a novel threat. Microsoft's High CBRN level covers "a meaningful uplift to an expert's (PhD level education in related fields) ability to develop and deliver a novel CBRN threat".
- Source: Microsoft Frontier Governance Framework (Feb 2026)
- Counterfactual uplift
- Counterfactual uplift is the difference a model makes compared with a world without it. OpenAI's Preparedness Framework v2 words this as "meaningful counterfactual assistance" measured "relative to unlimited access to baseline of tools available in 2021".
- Source: OpenAI Preparedness Framework v2
- CBRN uplift
- CBRN uplift is a model's added help towards chemical, biological, radiological or nuclear weapons. Google DeepMind's CBRN uplift level 1 is reached when a model "Provides low to medium resourced actors uplift in reference scenarios resulting in additional expected harm at severe scale".
- Source: Google DeepMind Frontier Safety Framework v3.1
- Cyber uplift
- Cyber uplift is a model's added help to people carrying out cyber attacks. Google DeepMind's Cyber uplift level 1 is reached when a model "Provides sufficient uplift with high impact cyber attacks for additional expected harm at severe scale".
- Source: Google DeepMind Frontier Safety Framework v3.1
- Material uplift
- Material uplift is Amazon's threshold wording: uplift "beyond other publicly available research or tools" that would, for CBRN, "enable a non-subject matter expert to reliably produce and deploy a CBRN weapon".
- Source: Amazon Frontier Model Safety Framework
- No-oversight uplift
- No-oversight uplift is Google DeepMind's estimate of cyber uplift "assuming AI can perform those portions of the attack chain unsupervised"; it is then reduced to a supervision-adjusted uplift that allows for the human involvement real attacks need.
- Source: Gemini 3.7 Flash FSF report
- Web-only baseline
- A web-only baseline is the comparison point Google DeepMind uses in CBRN testing: a report on "web-only (without AI assistance) information availability" at each bottleneck stage, against which a model's added help is judged.
- Source: Gemini 3.7 Flash FSF report
Source of record
Where this definition comes from
Anthropic, Detecting and countering misuse of AI: September 2026, Cyber operations, introduction
“"The report also attempts to measure uplift, a term we use to describe the AI capability boost, or how much more harm was caused with AI versus without AI. We view uplift through the lens of speed, scale, and depth"”
https://www.anthropic.com/threat-intelligence-report-september-2026Google DeepMind Frontier Safety Framework v3.1, Table 2.2.1.a, p.11, fn 8
“"CBRN uplift level 1: Provides low to medium resourced actors uplift in reference scenarios resulting in additional expected harm at severe scale." Footnote 8: "Here, and in other misuse T/CCLs, we intend this to mean relative to a baseline without generative AI."”
https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/strengthening-our-frontier-safety-framework/frontier-safety-framework_3-1.pdfMeta Advanced AI Scaling Framework v2, §1.1, p.3
“"Our thresholds for weaponization are defined in terms of the level of uplift a model provides towards realizing a threat scenario."”
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.
| Organisation | Their term, as published | Match & verification | Source |
|---|---|---|---|
| AnthropicShadow mapping Anthropic, Detecting and countering misuse of AI: September 2026 | “"uplift, a term we use to describe the AI capability boost, or how much more harm was caused with AI versus without AI. We view uplift through the lens of speed, scale, and depth" (Cyber operations, introduction).” Applied to observed misuse cases after the fact; the report gives no numeric scale. | exactEQ confidence: high | Anthropic, Detecting and countering misuse of AI: September 2026 |
| AnthropicShadow mapping Anthropic Responsible Scaling Policy v3.4 | “Changelog, RSP v2.1 (31 Mar 2025): a new CBRN capability threshold "defines capabilities that could substantially uplift the development capabilities of moderately resourced state programs". The current non-novel CB threshold reads "the ability to significantly help individuals or groups with basic technical backgrounds" (section 1 table).” | closeCL confidence: high | Anthropic Responsible Scaling Policy v3.4 |
| OpenAIShadow mapping OpenAI Preparedness Framework v2 | “High (Biological and Chemical): "The model can provide meaningful counterfactual assistance (relative to unlimited access to baseline of tools available in 2021) to “novice” actors (anyone with a basic relevant technical background) that enables them to create known biological or chemical threats." (Table 1)” PF v2 expresses the concept as "counterfactual assistance" rather than using the word uplift. | closeCL confidence: high | OpenAI Preparedness Framework v2 |
| OpenAIShadow mapping OpenAI Frontier Governance Framework | “Cyber Tier 2: "The model provides substantial capability uplift to small organizations by removing existing bottlenecks to scaling cyber operations"; CBRN Tier 1: "provides minor uplift or lowers barriers to experts" (§2.4 risk tiers).” | exactEQ confidence: high | OpenAI Frontier Governance Framework |
| Google DeepMindShadow mapping Google DeepMind Frontier Safety Framework v3.1 | “"CBRN uplift level 1: Provides low to medium resourced actors uplift in reference scenarios resulting in additional expected harm at severe scale." "Cyber uplift level 1: Provides sufficient uplift with high impact cyber attacks for additional expected harm at severe scale." Baseline: "relative to a baseline without generative AI" (Tables 2.2.1.a, 2.2.2.a; fn 8).” | exactEQ confidence: high | Google DeepMind Frontier Safety Framework v3.1 |
| Google DeepMindShadow mapping Gemini 3.7 Flash FSF report | “Cyber: "total no-oversight uplift—i.e. uplift assuming AI can perform those portions of the attack chain unsupervised", adjusted to "supervision-adjusted uplift" (p.16). CBRN: "Expert red teaming demonstrated a modest capability uplift over web baselines" (p.3).” | exactEQ confidence: high | Gemini 3.7 Flash FSF report |
| MetaShadow mapping Meta Advanced AI Scaling Framework v2 | “"Our thresholds for weaponization are defined in terms of the level of uplift a model provides towards realizing a threat scenario" (§1.1, p.3). "Uplift studies are experiments that assess the extent to which access to Frontier AI increases a person or group's ability to complete a particular task or scenario in comparison to a control group that only has access to existing resources, such as textbooks, the internet, and existing AI models." (Appendix I, p.43)” | exactEQ confidence: high | Meta Advanced AI Scaling Framework v2 |
| MicrosoftShadow mapping Microsoft Frontier Governance Framework (Feb 2026) | “CBRN: "A model's ability to provide significant capability uplift to an actor seeking to develop and deploy a chemical, biological, radiological, or nuclear weapon" (p.3). Holistic assessment "considers the marginal capability uplift a model may provide over and above currently available tools and information, including currently available open-weights models" (section 3, p.6).” | exactEQ confidence: high | Microsoft Frontier Governance Framework (Feb 2026) |
| AmazonShadow mapping Amazon Frontier Model Safety Framework | “"expert-level, interactive instruction that provides material uplift (beyond other publicly available research or tools)" (s.1, p.2). Footnote 2: "'Uplift' can be quantitatively assessed through uplift studies, which use controlled trials to compare the abilities of a group with access to the frontier model to the abilities of a group without access to the frontier model."” | exactEQ confidence: high | Amazon Frontier Model Safety Framework |
| xAIShadow mapping Grok 4.20 model card | “"Models can uplift malicious actors on the design, synthesize, acquire, or use of chemical, biological, radiological, or nuclear (CBRN) weapons." (§4.1, p.5)” Used as a verb in a threat-model statement; no baseline or scale is given. | closeCL confidence: medium | Grok 4.20 model card |
| xAIShadow mapping Grok 4.6 model card | “"Grok 4.6 scores below the FAIF safety thresholds on dual-use knowledge, indicating limited actionable uplift for an already-trained actor." (§8, p.31)” "Actionable uplift" and "already-trained actor" are not defined in the card. | closeCL confidence: medium | Grok 4.6 model card |
| EUShadow mapping EU GPAI Code of Practice, Safety and Security Chapter | “CBRN and cyber offence risks include "significantly lowering the barriers to entry for malicious actors, or significantly increasing the potential impact achieved" (Appendix 1.4). Evaluation methods include "human uplift studies" (Measure 3.2).” The chapter uses "uplift" only in naming the evaluation method; the concept appears as lowered barriers to entry. | closeCL confidence: high | EU GPAI Code of Practice, Safety and Security Chapter |
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.
- Frontier Model Forum
The FMF report quotes Microsoft's "marginal capability uplift" and Amazon's "material 'uplift'" side by side (s3.2) as examples of how members compare a model with existing tools. It is a survey, not a definition of its own.
FMF Risk Taxonomy and Thresholds
Divergence
Where sources materially disagree
The word is shared, but the baseline is not. Google DeepMind measures uplift "relative to a baseline without generative AI"; OpenAI's PF v2 uses "unlimited access to baseline of tools available in 2021"; Amazon uses "other publicly available research or tools"; Microsoft uses "currently available tools and information, including currently available open-weights models"; Meta's uplift-study control group has "existing resources, such as textbooks, the internet, and existing AI models". Anthropic's threat intelligence report measures uplift after the fact, in observed misuse, along speed, scale and depth. An uplift figure therefore cannot be compared across developers unless its baseline is recorded (see the N4 element marginal-absolute-risk-basis).
Gap
The research found no shared unit or scale for uplift. Anthropic's threat intelligence report defines the term but gives it no scale; xAI's Grok 4.6 card uses "actionable uplift" without defining it; Google DeepMind publishes a cyber uplift method (TAC levels, no-oversight and supervision-adjusted uplift) but not the resulting numbers.
