Frontier AI developers, including Anthropic, OpenAI, and Alibaba, have begun publishing “model specs”: documents that describe how their models are intended to act, communicate, navigate conflicting instructions, and more. But there’s no independent guidance for writing or implementing them.
A new paper from the @Safe_AI_Forum aims to address this gap by:
• Outlining the advantages of writing a model spec
• Introducing a taxonomy for specs
• Proposing guidance for writing and implementing a spec
• Highlighting open problems related to writing specs for safe AI behaviour
• Recommending best practices for AI companies, governments, and others
Its co-authors include LawAI’s Founding Director @Christophkw, alongside researchers from the China Academy of Information and Communications Technology (CAICT), the Universities of Cambridge and Oxford, and others.
Read the paper: saif.org/research/model…
saif.org Emerging International Best Practices for AI Model Specs - Safe AI Forum A new paper from the @Safe_AI_Forum aims to address this gap by:
• Outlining the advantages of writing a model spec
• Introducing a taxonomy for specs
• Proposing guidance for writing and implementing a spec
• Highlighting open problems related to writing specs for safe AI behaviour
• Recommending best practices for AI companies, governments, and others
Its co-authors include LawAI’s Founding Director @Christophkw, alongside researchers from the China Academy of Information and Communications Technology (CAICT), the Universities of Cambridge and Oxford, and others.
Read the paper: saif.org/research/model…