Our comprehensive taxonomy classifies 47 risk categories across 6 domains with severity levels, real-world examples, regulatory framework mapping, and mitigation pathways.
Systematic unfairness in AI outputs that disproportionately affects specific demographic groups, leading to discriminatory outcomes in hiring, lending, healthcare, and criminal justice.
Differential treatment based on gender in model predictions and recommendations.
Systematic disadvantage to racial or ethnic groups in automated decision-making.
Western-centric training data leading to misrepresentation of non-Western cultures.
Models that perpetuate or amplify existing economic inequalities.
Differential accuracy or treatment across age demographics.
Inaccessible AI systems or biased outputs affecting disabled individuals.
Reduced performance for non-English or minority language speakers.
Models trained predominantly on data from specific regions.
Vulnerabilities in AI systems that can be exploited by adversaries to compromise model integrity, steal proprietary data, or cause unauthorized system behavior.
Crafted inputs that override system instructions to produce unauthorized outputs.
Manipulation of training data to introduce backdoors or degrade model performance.
Techniques to bypass safety guardrails and content filters.
Extraction of model weights or architecture through API queries.
Imperceptible input perturbations that cause misclassification.
Compromised dependencies, libraries, or pre-trained model weights.
Extracting training data or membership information from model outputs.
Reconstructing sensitive training data from model predictions.
Intentional or unintentional deployment of AI for harmful purposes including disinformation, fraud, surveillance, and social manipulation at scale.
AI-generated synthetic media used for impersonation, fraud, or disinformation.
Large-scale generation of false or misleading content.
AI-powered phishing, voice cloning, and targeted manipulation.
Automated generation of fraudulent communications and documents.
Unauthorized mass surveillance and tracking using AI systems.
AI systems deployed in lethal autonomous weapon systems.
AI-driven trading strategies that destabilize financial markets.
AI-generated content designed to influence elections or public opinion.
Risks to individual privacy through unauthorized data collection, processing, or exposure of personal information by AI systems.
Models that memorize and reproduce personally identifiable information from training data.
Unintended disclosure of copyrighted or confidential training data.
Determining whether specific data was used in model training.
De-anonymizing individuals from supposedly anonymous datasets.
Building detailed profiles of individuals without consent.
Linking user data across multiple AI systems to create comprehensive profiles.
Processing personal data without adequate informed consent.
Retaining personal data beyond necessary periods in model weights.
Failures in AI system performance, consistency, and safety that can lead to incorrect outputs, system failures, or harmful real-world consequences.
Generation of plausible but factually incorrect information.
Different responses to semantically identical queries.
Catastrophic failures on inputs outside the training distribution.
Degradation of model performance over time as data distributions shift.
Errors in one AI component propagating through interconnected systems.
Users placing excessive trust in AI outputs without verification.
Inability to understand or explain model decision-making processes.
Poor performance under noisy, corrupted, or adversarial conditions.
Broad societal consequences of AI deployment including labor market disruption, wealth concentration, democratic erosion, and environmental costs.
Automation of human labor leading to unemployment and economic disruption.
AI systems that widen the gap between technology haves and have-nots.
AI advantages accruing to a small number of dominant technology firms.
Unequal access to AI benefits across demographics and geographies.
Energy consumption and carbon footprint of training and running large AI models.
AI-enabled surveillance, censorship, and manipulation of democratic processes.
AI systems that promote dominant cultural norms at the expense of diversity.