A Different Approach to AI Safety: Proceedings from the Columbia Convening on Openness in Artificial Intelligence and AI Safety

Openness can enhance AI safety through independent scrutiny and decentralized oversight: The paper finds that transparent model weights, interoperable tools, and public governance mechanisms enable broader communities to identify and address safety issues, rather than relying solely on centralized corporate oversight.

Critical safety gaps exist in open-source AI systems: The research identifies three major deficiencies: a shortage of benchmarks for evaluating multimodal and multilingual models, limited protections against prompt-injection and compositional attacks in AI agents, and insufficient mechanisms for input from communities most harmed by AI failures.

Content safety filters need significant improvement: The paper maps the current ecosystem of content filters and calls for "future-proof" approaches, suggesting that existing safety measures are inadequate and require coordinated research and development across the field.

Participatory processes are essential but underutilized: The convening emphasizes that affected communities must have meaningful input into AI safety decisions, yet current mechanisms for this participation remain insufficient—a gap that needs urgent attention.

A coordinated research agenda is needed across the AI development workflow: Rather than isolated safety efforts, the paper proposes five priority research directions that address safety holistically, from model development through deployment, requiring collaboration between academia, industry, civil society, and government.