A Taxonomy of AI Risks

Before you can reason carefully about AI risk, you need a vocabulary precise enough to distinguish one kind of problem from another. 'AI is dangerous' is not an argument — it is a gesture. 'This AI system creates a misuse risk because it lowers the technical barrier for synthesizing disinformation at scale' is a claim with enough structure to evaluate, refute, or confirm. The field of AI safety has converged on a set of categories that carve up the risk landscape in a principled way. This lesson introduces the three primary categories and explains why the distinctions matter practically.

Three Primary Categories

Misuse risks arise when AI systems work exactly as designed — but are deliberately used by a person or organization to cause harm. The AI is a tool; a human actor is the proximate cause of harm. Examples: using a voice-cloning system to impersonate a CEO and authorize a fraudulent wire transfer; using a generative model to produce personalized phishing emails at industrial scale; using an AI-assisted cyberattack tool to penetrate critical infrastructure. The harm comes from human intent, enabled by AI capability. Accident risks arise when AI systems fail in unintended ways — when the system does something its designers did not want and did not anticipate. The harm is not malicious intent; it is technical failure. Examples: an autonomous vehicle misclassifying a stop sign due to adversarial stickers and running the sign; a medical AI recommending a dangerous drug interaction because its training data lacked examples of elderly patients with multiple comorbidities; a content moderation system incorrectly removing legitimate journalism while leaving up coordinated harassment. The harm comes from system failure, not human malice. Structural and systemic risks arise not from any single AI system failing or being misused, but from the cumulative, society-level effects of AI's widespread adoption. These risks often emerge from the interaction of many systems, incentives, and human behaviors. Examples: AI-powered automation eliminating entire occupational categories faster than workers can retrain; algorithmic content curation systematically rewarding outrage and homogenizing public discourse; the concentration of AI capability in a small number of companies creating structural economic and political power imbalances. The harm is diffuse, emergent, and often not attributable to a single decision or system.

Overlapping and Compounding Risks

The three categories are analytically distinct but empirically overlapping. A single incident often involves all three. Consider a large language model deployed as a customer-service agent that is prompted by a malicious user to reveal confidential data (misuse), that occasionally hallucinates policy information and gives customers wrong advice about their rights (accident), and whose widespread deployment in the industry contributes to the elimination of tens of thousands of customer-service jobs with inadequate social support structures (structural). Analyzing that system requires all three lenses simultaneously. Compounding risks are also common: an accident failure erodes public trust in a way that makes misuse easier (if people believe AI is unreliable anyway, deepfakes become more believable); or structural concentration of AI power makes it harder to regulate misuse because the powerful actors who most benefit from the status quo resist oversight. Good risk analysis does not treat the categories as mutually exclusive — it uses them as analytical entry points that illuminate different facets of the same system. A second important dimension cuts across all three categories: the distinction between near-term and long-term risks. Some misuse, accident, and structural risks are already occurring today and are well-documented. Others are plausible future developments whose probability and timeline are genuinely uncertain. We will examine this near-term versus long-term distinction carefully in Lesson 6. For now, note that both near-term and long-term risks can be serious — and that taking near-term harms seriously does not require dismissing longer-horizon concerns.