Scholarlink

We explore a possible approach to ex ante regulation of AI systems based on behavioural red lines that demarcate specific, unacceptable behaviours that AI systems must not exhibit. The onus would be on developers to demonstrate with high confidence that their systems will not cross these red lines. We briefly discuss possible forms that such demonstrations could take and the difficulty of compliance for black-box systems.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Behavioural Red Lines as a Safety Framework for Intolerable Levels of Risk

  • 章节

  • 发表日期：

    2026年

查看全文
  • Yoshua Bengio,
  • Stuart Russell,
  • Andrew Yao,
  • Ya-Qin Zhang

摘要

We explore a possible approach to ex ante regulation of AI systems based on behavioural red lines that demarcate specific, unacceptable behaviours that AI systems must not exhibit. The onus would be on developers to demonstrate with high confidence that their systems will not cross these red lines. We briefly discuss possible forms that such demonstrations could take and the difficulty of compliance for black-box systems.