Overview
- Researchers at Cornell and Carnegie Mellon published a peer‑reviewed theoretical model that compares safety rules aimed at general‑purpose model developers with rules aimed at downstream deployers.
- The model predicts that weak rules focused only on downstream companies can backfire because general AI providers will cut safety investments and shift responsibility to users.
- Requiring both upstream model makers and downstream deployers to meet meaningful safety targets creates a regulatory 'sweet spot' in the model that raises safety for users and can increase firm profits.
- The result is derived from a simplified game‑theory framework that the authors say needs extension to multiple firms, multiple regulators, and empirical validation before direct translation into law.
- The paper enters an active policy debate over fragmented U.S. and international approaches and strengthens the formal case for rules that directly include major model developers such as OpenAI, Google, and Anthropic.