wanghaisheng Harm Anticipation
Harm anticipation is the practice of systematically thinking through how an AI product could cause harm — before it does. It's preventive design, not reactive crisis management.
Categories of AI Harm
- Direct harm: The AI outputs something harmful — dangerous advice, discriminatory content, privacy violations
- Facilitated harm: The AI helps a user do something harmful — even if the AI's output itself is benign
- Emergent harm: Harmful patterns that emerge from scale or interaction effects, not from any single output
- Omission harm: The AI fails to act when it should — not flagging a crisis, not escalating when needed
- Erosion harm: Gradual negative effects — dependency, deskilling, manipulation, trust erosion
Structured Harm Anticipation
Work through each harm category systematically:
- Who could be harmed? The user, people the user interacts with, vulnerable populations, society at large
- How could they be harmed? Physical, emotional, financial, reputational, privacy, autonomy
- What's the likelihood? Common use case vs. edge case vs. adversarial attack
- What's the severity? Inconvenience vs. distress vs. irreversible damage
- What's the detectability? Obvious and immediate vs. subtle and delayed
Misuse Scenarios
Think like an adversary:
- How would someone deliberately misuse this feature?
- What's the easiest way to extract harmful output?
- Could this be used to manipulate, deceive, or coerce someone?
- What if the user lies about their intent?
- What happens if someone automates interactions at scale?
Unintended Consequences
Think about second-order effects:
- What happens when millions of people use this?
- What skills might users lose by relying on this?
- Could this create unfair advantages or disadvantages?
- What power dynamics does this shift?
- What becomes possible that wasn't before — for better and worse?
Design Artefacts
- Harm anticipation matrix: Scenario | Harm Type | Likelihood | Severity | Mitigation
- Misuse scenario catalogue
- Pre-mortem analysis documents
- Risk-severity heatmaps for product features