Responsible AI for General-Purpose Systems: Overview, Challenges, and A Path Forward

General-purpose AI systems pose fundamentally different risks than traditional AI: The paper identifies that modern large language and vision models have a "non-deterministically high Degree of Freedom in output" (DoFo), meaning they can generate vastly more varied outputs than task-specific systems, making risks like hallucinations, toxicity, and stereotypes harder to predict and control.

Eight key responsible AI principles need rethinking for modern systems: The paper evaluates general-purpose AI against widely accepted responsible AI principles—fairness, privacy, explainability, robustness, safety, truthfulness, governance, and sustainability—and finds that current approaches developed for traditional AI are insufficient for addressing these risks in systems like ChatGPT and similar models.

A new framework (C2V2) is needed to guide responsible AI development: Rather than applying existing safeguards, the authors propose four new desiderata—Control, Consistency, Value, and Veracity—specifically designed to address the unique challenges posed by general-purpose AI systems with their unpredictable output ranges.

Emerging technical approaches show promise but need formal integration: Recent efforts in AI alignment, retrieval-augmented generation, and reasoning enhancements are moving in the right direction, but the paper suggests these need to be formally modeled and integrated into domain-specific applications to effectively meet responsible AI requirements.

One-size-fits-all solutions won't work: The paper emphasizes that responsible general-purpose AI requires application- or domain-dependent approaches rather than universal solutions, acknowledging that different use cases have different risk profiles and requirements.