Newshot Frontier AI developers from around the world have begun to publish documents that describe how their AI models are intended to behave. Although these ‘model specifications’—or ‘model specs’, for short—are increasingly important in AI development and governance, little independent guidance exists for writing them. This paper addresses this gap in five main sections. First, we outline the advantages of writing a model spec for AI developers and society at large. Second, we introduce a taxonomy for model specs that developers can take inspiration from. Third, we propose guidance for writing and implementing a spec in practice. Fourth, we note eight open problems related to writing model specs for safe and desirable behaviour. We conclude with recommendations for frontier AI companies, governments, and research and civil society organisations.
Frontier AI developers from around the world have begun to publish documents that describe how their AI models are intended to behave. Although these ‘model specifications’—or ‘model specs’, for short—are increasingly important in AI development and governance, little independent guidance exists for writing them.
This paper addresses this gap in five main sections. First, we outline the advantages of writing a model spec for AI developers and society at large. Second, we introduce a taxonomy for model specs that developers can take inspiration from. Third, we propose guidance for writing and implementing a spec in practice. Fourth, we note eight open problems related to writing model specs for safe and desirable behaviour.
We conclude with recommendations for frontier AI companies, governments, and research and civil society organisations.