Human-AI Safety: A Descendant of Generative AI and Control Systems Safety

Current AI safety approaches are incomplete because they ignore human behavior: The paper argues that today's dominant safety methods—fine-tuning AI outputs based on human feedback—fail to account for how people actually respond to and interact with AI systems over time, which can lead to unintended consequences.

Human-AI interactions form feedback loops that need safety analysis: Rather than treating AI outputs as isolated events, the paper emphasizes that meaningful safety requires understanding how AI responses influence human behavior, which then influences future AI outputs, creating dynamic cycles that can drive interactions toward harmful outcomes.

Control systems engineering offers valuable lessons for AI safety: The paper identifies synergies between traditional control systems safety (used in aviation, nuclear power, etc.) and AI safety, suggesting that proven methods from these high-stakes fields could help address gaps in current AI safety approaches.

We need new frameworks to reason about dynamic human-AI interactions: The authors propose developing formal methods and technical approaches specifically designed to analyze safety-critical scenarios where humans and AI systems continuously influence each other, rather than treating them as separate components.

Responsible AI requires a shift toward human-centered safety thinking: Moving beyond output-focused safety measures, the paper calls for a new paradigm that explicitly models and reasons about the joint human-AI system as an integrated whole to provide meaningful safety assurances.