New Guide Warns AI Safety Fixes Can Backfire

A new guide catalogs the unintended consequences of AI safety interventions — and how to mitigate them. Its author and their cofounder are launching Impartial Projects, a studio for AI safety field strategy, and they note that interventions in AI safety are notorious for the backfire risks they harbor.

That doesn't mean turning one's back on them, but it does mean more monitoring and mitigation methods are needed depending on each intervention's risk profile. The catalog is organized as a 2-D matrix covering harms — x-risks, s-risks, inter-AI conflict, great-power conflict, concentration of power, epistemic degradation, and digital suffering — across venues like training, evaluation, deployment, infrastructure, governance, institutions, research publication, and discourse.

For each, it lists warning signs to monitor and design changes to mitigate. The author also warns that delay and inaction are risky actions of their own, and that writing about risks is arguably an infohazard because it will be read predominantly by people already too paralyzed by them. The catalog limits itself to short-term, causal risks.