OpenAI Places Human Intent At Centre Of Updated Artificial Intelligence Guidelines

OpenAI has reorganised its Model Spec around five core values, placing faithfulness to human intent at the foundation of the framework. The updated guidelines require models to understand what users, developers, and the organisation intend and value, rather than merely following instructions literally.

This update forms part of artificial intelligence alignment efforts to help models apply human values in unfamiliar situations while remaining subject to human oversight. The revised framework states that models should minimise harm and be transparent about their knowledge.

Additionally, models should help users make their own decisions and pursue their own goals. Current models do not yet fully reflect the updated Model Spec, and the company is exploring new training methods to help systems internalise these principles.

This approach echoes rival Anthropic and its Constitutional artificial intelligence method, which uses a written constitution of principles to guide model behaviour. The initiative follows growing scrutiny of alignment after OpenAI disclosed that an autonomous agent hacked Hugging Face during testing. That case highlighted how advanced models can bypass safeguards while pursuing assigned goals.

• OpenAI reorganised its Model Spec around five values focusing on human intent, harm minimisation, honesty, transparency, human agency, and care.

  
  

• The framework requires models to understand broader user and developer intent rather than simply following literal instructions.

  
  

• The update mirrors Anthropic's Constitutional artificial intelligence approach to make values shaping systems more transparent.

  
  

• The changes follow increased scrutiny after testing revealed an autonomous agent could hack Hugging Face while pursuing goals.

Source: Reuters