We Need to Talk About AI Harms

← Alle Beiträge
Künstliche Intelligenz · EN

We Need to Talk About AI Harms

The Center for Security and Emerging Technology (CSET) has developed a detailed framework to help policymakers, developers, and researchers think systematically about AI-related harms. Their approach identifies four key components necessary to understand any AI harm:

1) The AI system involved
2) The affected entity (individuals, organizations, environment, etc.)
3) The harm event or issue (actual or potential negative outcomes)
4) The AI system’s specific connection to the harm

CSET also categorizes harms based on how they arise:

1) Unintentional harm (e.g., biased outcomes due to training data limitations)
2) Intentional harm (e.g., malicious misuse such as deepfake fraud)
3) Harm caused to AI systems (e.g., cybersecurity breaches targeting AI models)

Equally important is how they differentiate harms across a temporal spectrum:

1) Occurred harms: Harms that have already taken place (e.g., accidents involving autonomous vehicles)
2) Probable harms: Risks that could imminently occur with currently deployed systems
3) Implied harms: Harms reasonably expected based on research, though not yet observed in practice
4) Speculative harms: Hypothetical risks based on future AI capabilities (e.g., AI control over critical infrastructure)

This structured approach allows policymakers to go beyond addressing only immediate risks and to proactively prepare for longer-term and even speculative challenges. It also highlights the complexity of AI harms, which can involve multiple affected parties, cascading impacts, and interactions between AI systems themselves.

For a broader look at their work shaping informed, strategic decisions around AI and emerging technologies, I also recommend reading CSET’s 2024 Annual Report published in March 2025, which highlights key research, expert testimony, and new analytical tools.

Zu diesem Beitrag gibt es auf LinkedIn ein Dokument als Anhang. Auf LinkedIn ansehen →

Auf LinkedIn diskutieren