Conversation

New paper with an exhaustive taxonomy of societal-scale AI risks, based on accountability: arxiv.org/pdf/2306.06924 Extinction, injustice, and other widespread harms are considered. Additional taxonomies are needed for a more diverse and robust perspective on risk. Meanwhile, you might appreciate: * this self-fulfilling pessimism story: arxiv.org/pdf/2306.06924 * this figure depicting industries that could eventually get out of control in a closed loop: arxiv.org/pdf/2306.06924 ...as in this "production web" story: arxiv.org/pdf/2306.06924 * these two "bigger than expected" AI impact stories: arxiv.org/pdf/2306.06924 * this email helper story and corrupt mediator story, which go together: arxiv.org/pdf/2306.06924 arxiv.org/pdf/2306.06924 * this harmful A/B testing story: arxiv.org/pdf/2306.06924 * concerns about weaponization by criminals and states: arxiv.org/pdf/2306.06924 The point of the paper is not just to tell stories, but to illustrate a strategy for uncovering new risks and novel ways to address them: exhaustive taxonomy. It's a powerful thinking strategy, similar to classical fault tree analysis, and essential to mitigating AI risk.
I prefer links to /abs/ when referencing the entire paper. But Critch also includes links to specific pages of the PDF here. Gotta use /pdf/ for those.
1