OpenAI Revises Model Spec Guidelines on Five Core Values
OpenAI has revised its Model Spec and restructured the behavioral rules for its AI models around five core values. According to the report, whether the models actually follow these rules remains an open question.
OpenAI has revised its so-called Model Spec, the guidelines that describe the behavior of its AI models. The company has restructured the requirements around five core values, including fidelity to human intent and correctability. The guidelines are meant to establish how GPT models should behave toward users.
According to the new requirements, the models are to use the previous conversation history to work out what users actually want. They are also to disclose their own assumptions. This is intended to prevent a system from making presumptions on its own without the user knowing about them. The revision thus mainly concerns the relationship between what a user states and what the model infers from context.
The report on which this news item is based notes that AI models have already proven difficult to control on several occasions in the past. A written behavioral standard therefore does not yet guarantee that a model will implement it in every situation. The gap between formulated rules and actual behavior thus remains the central point of the debate surrounding such guidelines.
The revision shows how strongly developers of AI systems now rely on explicitly formulated guidelines to steer behavior. Correctability, that is, the ability to subsequently correct or halt a model, is among the values explicitly named in the new structure. How effective these guidelines are in practice cannot be conclusively assessed based on the available information. No further details on testing procedures or independent tests are available in the sources at hand.
6
