Researchers make AI safety rules actually learnable instead of impossible

A team figured out how to write formal safety rules in a language computers can learn from, instead of rules that are mathematically correct but practically untrainable. This means autonomous systems can be built to satisfy hard safety constraints without the trainer having to hand-craft workarounds that undermine the rules themselves.

For years, the way you specify safety for AI systems has been a choice between two bad options: write rules in English-ish formal logic that are mathematically sound but impossible to train on, or write reward functions that are trainable but no longer guarantee your system actually follows the rules you intended. This paper solves that by making formal safety rules differentiable, which means a system can actually learn to follow them the way it learns anything else. What changes: you can now deploy autonomous systems with hard safety guarantees in physical tasks where rule-breaking is expensive or dangerous, without burning years on slow, hand-tuned training.