This week we summarize “TASRA: A Taxonomy and Analysis of Societal-Scale Risks from AI” by Andrew Critch and Stuart Russell (2023).
The framework organizes risks from AI into six risk types:
1️ Diffusion of Responsibility
Societal-scale harm can arise from AI built by a diffuse collection of creators, where no one is uniquely accountable for the technology’s creation or use
2️ "Bigger than Expected" AI Impacts
Harm that can result from AI that was not expected to have a large impact Example(s): a lab leak or unexpected repurposing of a research prototype.
3️ "Worse than Expected" AI Impacts
AI intended to have a large societal impact can turn out harmful by mistake Example(s): a product that creates problems and partially solves them only for its users.
4️ Willful Indifference
As a side effect of a primary goal like profit or influence, AI creators can wilfully allow it to cause widespread social harms Example(s): pollution, resource depletion, mental illness, misinformation.
5️ Criminal Weaponization
One or more criminal entities create AI to intentionally inflict harm Example(s): terrorism or combatting law enforcement
6️ State Weaponization
AI deployed by states in war, civil war, or law enforcement
Uses a decision tree to classify societal-scale harms from AI in an exhaustive manner Explores a taxonomy based on accountability: whose actions lead to the risk, are the actors unified, and are they deliberate?
Provides stories to illustrate how the various risk types could play out, including risks arising from unanticipated interactions of many AI systems, as well as deliberate misuse
References/further reading
Critch, A., & Russell, S. (2023). TASRA: a taxonomy and analysis of societal-scale risks from AI. arXiv preprint arXiv:2306.06924.