Appendix C | Risk Framework Comparisons

In the 2024 AI Assurance Technology Market Report, we present an original framework to categorise and describe risks resulting from the development and deployment of AI.

That framework was developed with the purposes of the report in mind, but it was not cherry-picked to only correspond to existing AI Assurance Technology solutions; instead, it seeks to give a comprehensive overview of known and expected AI-related risks.

In order to validate the coverage and sensibility of our framework, we consulted the existing literature and compared risk categorisations from diverse sources with the categories we produced.

We relied on some of these sources in developing the framework and added others to the comparison afterwards. Table 1 gives an overview of the frameworks we encountered in the literature.

Table 2 compares a few selected frameworks with our framework, finding that our categorisation is able to capture the risks identified by other authors. Given the large volume of AI-related risk analyses and discussions published in recent years, we cannot claim that our overview below is completely exhaustive.

However, we would argue that its coverage of a decent sample of prominent AI risk frameworks is a decent basis for validating and cross-checking the comprehensiveness of our framework.

Table 1: Comparison of a few prominent AI Risk Frameworks

Source	Categories	Main dimension for categorisation	Time period	Type of AI	Focus Area(s)
Our framework (2024)	Misuse of AI system - Physical harm - Digital harm - Informational harm Internal technical failure of AI system - Misalignment - Unreliability Vulnerability of AI system to exogenous interference - Natural hazard or accident - Adversarial attack (Socioeconomic disruption) (Drivers of risk and vulnerability)	Pathways to harm or loss	Now and prospective	All / Undefined	Comprehensive, with a bias towards business or economic risk
U.S. NIST (2023)	Valid and reliable Safe Secure and resilient Accountable and transparent Explainable and interpretable Privacy enhanced Fair with their harmful biases managed	AI characteristics to prevent harm	Now and prospective (with adaptations)	All / Undefined	Comprehensive (harms affecting “individuals, groups, organisations, communities, society, the environment, and the planet”)
UNSG (2023)	Individuals Groups Society Economy (Eco)systems Values and norms Technical limitations Inappropriate use Human-machine interaction Catastrophic risks	Primary: Targets of harm Secondary, for illustration: Pathways to harm and Severity of harm (inconsistent)	Now and prospective	All / Undefined	Comprehensive
EU AI Act (2024)	Unacceptable risk High risk (Systemic risk) Moderate risk No risk x	Severity	Now (current systems and areas of application are categorised)	All / Undefined	Comprehensive
Fjeld et al. (2020)	Privacy Accountability Safety and security Transparency and explainability Fairness and non-discrimination Human control of technology Professional responsibility Promotion of human values	AI characteristics to prevent harm	Now and prospective (with adaptations)	All / Undefined	Comprehensive (synthesis of AI principles documents released by governments as well as international, multi-stakeholder, civil society, and corporate actors)
Wirtz et al. (2022)	Technological, data, and analytical Informational and communicational Economic Social Ethical Legal and regulatory	Sector of harm	Now and prospective	All / Undefined	Comprehensive (systematic review and synthesis of academic AI risk frameworks)
Hendrycks (n.d.)	Malicious use AI Race Organizational Risks Rogue AIs	Pathways to harm?	Mainly prospective	Advanced ML-based AI	Catastrophic risks
Maham & Küspert (2023)	Unreliability Risks - Discrimination and Stereotype –Reproduction - Misinformation and Privacy –Violations - Accidents Misuse Risks - Cyber Crime - Biosecurity Threats - Politically Motivated Misuse Systemic Risks - Economic Power Centralisation and Inequality - Ideological Homogenization from Value Embedding - Disruptions from Outpaced Societal Adaptation	Pathways to harm? Relevance of specific hazards (for selection of subcategories)	Now and prospective	General-purpose AI	Comprehensive (?) Focus on societal risks (?)
Shelby et al. (2023)	Representational harms - Stereotyping social groups - Demeaning social groups - Erasing social groups - Alienating social groups - Denying people opportunity to self-identify - Reifying essentialist social categories Allocative harms - Opportunity loss - Economic loss Quality of service harms - Alienation - Increased labor - Service/benefit loss Interpersonal harms - Loss of agency - Tech-facilitated violence - Diminished health and well-being - Privacy violations Social system harms - Information harms - Cultural harms - Political and civic harms - Socio-economic harms - Environmental harms	Type of harm	Now	Algorithmic systems	Sociotechnical harms Focus on harms that arise directly from the interaction of humans with algorithmic systems
Miles et al. 2018	Digital Security Physical Security Political Security	Sector of harm	Now and prospective	ML-based AI with (current or near-future capabilities)	Misuse risks
Katzman et al. 2023	Reifying social groups Demeaning social groups Erasing social groups Stereotyping social groups Denying people the ability to self-identify	Types of harm	Now and prospective (?)	Imagine-tagging systems	Representational harms (bias, discrimination, etc)
Weidinger et al. 2022	Discrimination, Hate speech and Exclusion Information Hazards Misinformation Harms Malicious Uses Human-Computer Interaction Harms Environmental and Socioeconomic harms	Pathways to harm	Now and prospective	Large Language Models	Ethical and social risks
Bommasani 2022	Inequity and fairness Misuse Environment Legality Economics Ethics of scale	Sectors of harm/impact	Now and prospective	Foundation models	Societal impact (comprehensive)
Slaughter et al. 2021	Design flaws - faulty inputs - faulty conclusions - failure to adequately test Systemic effects - by facilitating proxy discrimination - by enabling surveillance capitalism - by inhibiting competition in markets	Causes of harm	Now	Algorithmic decision-making in high-stakes spheres	Economic justice; focused on harms that can be addressed by FTC regulators
Smuha 2021	Individual harm Collective harm Societal harm - Equality - Democracy - Rule of law	Targets of harm	Now and prospective	All / Undefined	Comprehensive, but analysis focuses on societal harms
Shevlane et al. 2023 (pdf)	Misuse Misalignment They also list a number of distinct risky capabilities, and give examples of risks resulting from each of those capabilities	Pathways to harm	Prospective	Frontier AI models	Focus on Extreme risks

Table 2: Mapping existing AI Risk Frameworks to the framework developed for this report

Categories in our framework	Hendrycks (n.d.)	Maham & Küspert (2023)	UNSG (2023)	U.S. NIST (2023)	Shelby et al. (2023)	Wirtz et al. (2022)
Misuse–Physical Harm	Malicious use	Misuse Risks–Biosecurity Threats	Inappropriate use (not explicitly discussed) Catastrophic risks (LAWS)	Secure and resilient Explainable and interpretable (requirement for mitigating this risk (??))	Interpersonal harms–Diminished health and well-being (not mentioned explicitly)	Technological, data, and analytical Ethical (LAWS)
Misuse–Cyber Harm	Malicious use	Misuse Risks–Cyber Crime Misuse Risks–Politically Motivated Misuse	Inappropriate use (not explicitly discussed)	Secure and resilient Explainable and interpretable (requirement for mitigating this risk (??))	Interpersonal harms–Tech-facilitated violence Interpersonal harms–Privacy violations Social system harms–Political and civic harms (surveillance)	Technological, data, and analytical Social (surveillance, privacy) Ethical (unethical use of data)
Misuse–Informational Harm	Malicious use	Misuse Risks–Politically Motivated Misuse	Inappropriate use (explicitly discussed) Catastrophic risks (mass surveillance)	Secure and resilient Accountable and transparent Explainable and interpretable (requirement for mitigating this risk (??))	Interpersonal harms–Loss of agency (?) Interpersonal harms–Tech-facilitated violence Interpersonal harms–Diminished health and well-being Social system harms–Information harms Social system harms–Political and civic harms (propaganda, manipulation)	Technological, data, and analytical Informational and communicational
Internal technical Failure–Unreliability	Organizational Risks Rogue AIs	Unreliability Risks–Discrimination and Stereotype Reproduction Unreliability Risks–Misinformation and Privacy Violations Unreliability Risks–Accidents	Technical limitations	Valid and reliable Safe Accountable and transparent Explainable and interpretable (requirement for mitigating this risk (?)) Privacy enhanced (prevent accidental data leak) Fair with their harmful biases managed (prevent unfair outcomes due to imbalanced or insufficient data)	Representational harms (all subcategories) Allocative harms (all subcategories) Quality of service harms (all subcategories) Interpersonal harms–Diminished health and well-being Interpersonal harms–Privacy violations Social system harms–Information harms (unintended hallucinations) Social system harms–Cultural harms (bias against non-majority cultures) Social system harms–Socio-economic harms (financial crash due to unreliability in AI-powered financial traders)	Technological, data, and analytical Social (discrimination) Ethical (discrimination, harmful accidents)
Internal technical Failure–Misalignment	Rogue AIs	Unreliability Risks–Accidents Systemic Risks–Ideological Homogenization from Value Embedding Systemic Risks–Disruptions from Outpaced Societal Adaptation	Catastrophic risks (uncontrollable AI)	Safe Accountable and transparent Explainable and interpretable (requirement for mitigating this risk (?)) Fair with their harmful biases managed (prevent unfair outcomes due to misspecified reward functions)	Interpersonal harms–Loss of agency Interpersonal harms–Diminished health and well-being Social system harms–Information harms (unintended homogenization of the information environment as a side-result of algorithms pursuing some other goal) Social system harms–Cultural harms (unintended marginalisation of some cultures as a side-result of algorithms pursuing some other goal)	Technological, data, and analytical Informational and communicational (filter bubbles as an unintentioned side-effect of misaligned recommender algorithms) Ethical (omitted values; difficulty of encoding human values)
Vulnerability to exogenous interference–Accident	N/A	N/A	N/A	Secure and resilient	N/A	N/A
Vulnerability to exogenous interference–Attack	N/A	Unreliability Risks–Accidents Unreliability Risks–Misinformation and Privacy Violations	N/A	Secure and resilient Privacy enhanced (prevent data theft)	N/A	Technological, data, and analytical Social (privacy)
(Socioeconomic disruption)	Malicious use AI Race	Systemic Risks–Economic Power Centralisation and Inequality Systemic Risks–Disruptions from Outpaced Societal Adaptation	Human-machine interaction	Privacy enhanced (prevent excessive data collection by AI developers) Fair with their harmful biases managed (prevent unfair disparities in access to AI services)	Interpersonal harms–Loss of agency (??) Interpersonal harms–Privacy violations (excessive data collection by AI developers) Social system harms–Socio-economic harms Social system harms–Environmental harms	Economic (e.g., inequality) Social (e.g., unemployment) Ethical (meaning crisis) Legal and regulatory (copyright issues)
(Drivers of risk)	AI Race Organizational Risks	Systemic Risks–Disruptions from Outpaced Societal Adaptation Mentioned throughout	Human-machine interaction	Accountable and transparent Explainable and interpretable Fair with their harmful biases managed (prevent unfair outcomes due to how humans interact with AI-generated information)	Representational harms (those harms are made worse because of automation bias and system opacity)	Technological, data, and analytical (opacity, oversight) Social (human-machine interaction, opacity and lack of public understanding) Ethical (difficulty of encoding human values) Legal and regulatory (accountability, unpredictability, opacity)
No match	N/A	N/A	N/A	N/A	N/A	N/A
Note			Categories serve mostly illustrative purposes; the authors’ primary approach to categorisation separates hazards by target / vulnerability (individuals, groups, society, economy, (eco)system, norms and values)	Categories describe characteristics to ensure trustworthy AI; we matched them with hazard categories by asking which type of hazard each characteristic is aiming to address.		Categories are broad, encompassing several risks, hence the many-to-many matching.