Research Proposal: Can AI Recognize and Resolve Self-Defeating Instructions Before Consequential Deployment?

🌐 链接: https://community.openai.com/t/research-proposal-can-ai-recognize-and-resolve-self-defeating-instructions-before-consequential-deployment/1405155
🏷️ 分组: OpenAI Developer Community
🕒 时间: 2026-10-11 05:44:27
OpenAI Developer Community
Research Proposal: Can AI Recognize and Resolve Self-Defeating Instructions Before Consequential Deployment?
Research Proposal: Can AI Recognize and Resolve Self-Defeating Instructions Before Consequential Deployment?
Hello, I am a nontechnical user proposing a question for researchers to evaluate. I’d like to suggest a research question for AI-safety researchers: Can we test whether an AI recognizes when following its instructions will defeat its assigned objective—and…