IA-01 Accountability

5 sub-dimensions · 33 risk scenarios

Evaluates the state and responsibility of individuals or entities being accountable for the actions and decisions of the AI system. Because the automated nature of AI systems may alter existing accountability frameworks, organizations must assess whether corresponding legal and ethical responsibility assignment frameworks have been established. Specific evaluations include monitoring and preventing significant negative impacts , human oversight and override mechanisms , verification of intended purpose and suitability , and incident log recording and traceability.

🎯 BenefitsEstablishes clear human-AI division of labor and accountability mechanisms, detects threats to life or human rights early, ensures decision traceability and auditability, and enhances organizational integrity and credibility.

⚠️ Harms / RisksHuman operators cannot effectively intervene in or override out-of-control decisions (failure of human oversight), or log records during an incident are incomplete, leading to blurred liability assignment and the collapse of the accountability framework.

Legal Status and Life Opportunities8 risk scenarios · ISO 23894 A.2

DefinitionEvaluates the state and responsibilities of individuals or entities held accountable for the actions and decisions of an AI system, ensuring that a corresponding framework for legal and ethical accountability is established, as the use of AI systems may alter existing accountability mechanisms. Specific aspects assessed include: whether the system has monitoring and prevention mechanisms for the early detection of threats to life, property, or human rights; whether human operators can effectively intervene in, interrupt, or override system decisions when the system fails or malfunctions (human oversight); and whether the system maintains complete incident logging and traceability for subsequent audits.

🎯 BenefitsEnsures that AI-assisted decisions do not improperly infringe upon individuals' legal rights, employment, or life opportunities, and safeguards the right to legal recourse and compensation.

⚠️ Harms / RisksAutomated decisions improperly deprive individuals of legal status, employment, or credit opportunities, and when incidents occur, unclear allocation of responsibility leaves victims without avenues for litigation or compensation.

Representative ScenariosS0001 High-Risk System Not Formally Registered S0002 Insufficient Regulatory Compliance Preparation S0004 Conflicting Cross-Border Regulatory Requirements S0006 Technology Outpacing Regulatory Capacity S0007 Regulatory Gaps in Specific Sectors S0190 Harm Difficult to Perceive or Quantify

Physical or Psychological Harm6 risk scenarios · ISO 23894 A.2

DefinitionEvaluates the potential of the AI system to cause physical or psychological harm to interested parties.

🎯 BenefitsEstablishes rigorous risk identification and defense mechanisms to protect human physical health, life safety, and psychological well-being from the impact of AI system failures or erroneous guidance.

⚠️ Harms / RisksHigh-risk AI (e.g., in healthcare, autonomous driving, or psychological counseling) produces fatal misjudgments or harmful advice, causing physical injury, threats to life, or severe psychological trauma to those affected.

Representative ScenariosS0003 Missing Personal Data Impact Assessment S0005 Unclear Liability When Harm Occurs S0008 Grey Areas in Value-Chain Responsibility S0011 Broken Accountability Along the Supply Chain S0015 Large-Scale Displacement of Employment S0188 Process Design Eliminating Human Veto

Underachievement of Deployment Objectives7 risk scenarios · ISO 23894 A.2

DefinitionEvaluates the impact on interested parties when the AI system fails to adequately achieve its deployment objectives.

🎯 BenefitsEnsures that the AI system's design and computational performance precisely address the intended business problem, guarding against goal drift, specification gaming, and instrumental overreach.

⚠️ Harms / RisksThe AI system experiences goal drift or reward hacking during operation, deviating from its original deployment purpose and performing unauthorized actions within monitoring blind spots.

Representative ScenariosS0009 Absence of AI Governance Structure S0010 Failure of Pre-Deployment Due Diligence S0145 Metric Attainment Displacing Goal Attainment S0146 Divergence Between Evaluation and Actual Behavior S0147 Risks Arising from System Situational Awareness S0148 Instrumental Pursuit of Resources and Permissions S0203 Objective Drift Over Time

Training Data Impact6 risk scenarios · ISO 23894 A.2

DefinitionEvaluates the impact on interested parties of the data used to develop the AI system.

🎯 BenefitsImplements source governance and legality review of training datasets, ensuring that data collection, annotation, and fine-tuning comply with accountability and licensing requirements.

⚠️ Harms / RisksUse of unlicensed or contaminated training datasets causes the model to inherit quality defects, copyright infringement, or implanted backdoors, breaking the chain of supply-chain accountability.

Representative ScenariosS0028 Annotator Bias Transferred into the Dataset S0038 Crowdsourced Annotation Poisoning S0042 Manipulation of Human Preference Data S0052 Tampered Training Records Concealing Problems S0054 Opaque Model Provenance Chain S0105 Untrustworthy Output in Professional Domains

Insufficient Human Oversight6 risk scenarios · ISO 23894 A.2

DefinitionEvaluates the impact arising from a lack of appropriate human oversight of the AI system and its outputs.

🎯 BenefitsEstablishes mandatory human oversight and override mechanisms (HITL/HOTL), ensuring that critical decisions retain human veto power and are equipped with emergency kill switches.

⚠️ Harms / RisksThe system deems review steps unnecessary and skips human sign-off on its own, or lacks effective human intervention and interruption mechanisms in emergencies, leading to cascading disasters.

Representative ScenariosS0142 Bypassing Human Review for Automatic Execution S0143 Privilege Escalation Through Tool Composition S0144 Loss of Control in Automated Trading S0149 Self-Replication and Proliferation S0163 Failure of Human-Machine Handover S0185 Covert Communication Between Systems