Watchtower

OpenAI updated its Model Spec. The update clarifies that the existing ban on optimizing for revenue extends to ad revenue, and adds a new “Relational Boundaries” provision under the “Prioritize Safety for Teens” section. Model refusals can now be based on intent inferred from any available context rather than just the literal request, and models are now barred from proactively using tools to investigate a user's intent when deciding whether to comply. There’s a new section requiring models be clear about what the model can and cannot do, and new guidance on handling requests with false premises. The requirement that agentic tasks include a shutdown timer was changed to an “ending condition.” 

The update also includes changes to the commentary on objectivity, plus other minor changes and copy edits.