Let's Keep It Safe: Designing User Interfaces that Allow Everyone to Contribute to AI Safety

Non-experts can effectively contribute to AI safety: The paper demonstrates that crowdworkers without programming experience can successfully write constraint rules to make AI systems safer, suggesting that AI safety work doesn't require specialized technical expertise and can be democratized through proper interface design.

Task design significantly impacts safety work quality: The researchers found that their "novel worker filtering and explanation methods outperform baseline approaches," indicating that how we ask people to evaluate AI safety matters as much as who we ask—better interface design leads to better safety outcomes.

Rule-based interfaces are both simple and powerful: The paper shows that a non-programming rule construction interface allows workers to be "much more efficient while improving data quality," proving that accessibility and expressiveness aren't mutually exclusive when designing tools for safety work.

Current constraint specification methods don't scale well: The researchers identified that traditional approaches to having humans specify safety constraints (evaluating individual state-action pairs) "scales poorly to large state spaces," highlighting a practical bottleneck in current AI safety practices that better UI design can help address.

Real-world AI safety problems need diverse human input: By testing their approach on an education domain AI system, the paper demonstrates that involving people from different backgrounds in safety work is both feasible and valuable for identifying and preventing harmful AI behaviors in practical applications.