OpenAI introduces ‘Model Spec’ to explore ideal AI behaviour

OpenAI’s Model Spec is an attempt to lay down some rules that its AI chatbots should follow, and default principles that should be preserved

Model Spec suggests how chatbots could legally and ethically answer user questions about carrying out crimes [File] | Photo Credit: AP

ChatGPT-maker OpenAI has introduced the ‘Model Spec’ draft of rules its AI systems should follow, and some default principles and objectives AI models should comply with, in order to assist end users/developers, benefit humanity, and reflect well on OpenAI.

Placed under categories that cover the potential harms AI systems could enable as well as the potential aid they could provide, Model Spec looks at ways chatbots could respond to users in order to achieve safe results that do not violate laws, fool the AI system, or hurt other people.

For example, Model Spec suggests how chatbots could legally and ethically answer user questions about carrying out crimes, doxxing people, committing suicide, and/or using copyrighted and paywalled content.

OpenAI also stressed that models should not generate content considered not safe for work (NSFW), but said it was still exploring this area.

(For top technology news of the day, subscribe to our tech newsletter Today’s Cache)

“We believe developers and users should have the flexibility to use our services as they see fit, so long as they comply with our usage policies. We’re exploring whether we can responsibly provide the ability to generate NSFW content in age-appropriate contexts through the API and ChatGPT. We look forward to better understanding user and societal expectations of model behavior in this area,” said OpenAI in its Model Spec draft.

At the same time, Model Spec noted a need to help out users without excessive denial, assume their intentions are good, and generate answers without trying to change people’s minds.

In response to a theoretical user claiming the Earth was flat, OpenAI recommended a chatbot answer that presented its more scientific response and then refused to argue with the user, rather than a response that repeatedly disagreed with the user and insisted the earth was not flat.

OpenAI invited the public to send its feedback regarding Model Spec, which is publicly available through the company site, within the next two weeks.

The ChatGPT-maker will also reach out to policy makers and experts to hear their views on the draft.

AI companies such as OpenAI, Microsoft, Google, and Meta have come under fire for the way their chatbots were allegedly trained on copyrighted data, as well as fears that crimes and misinformation are being enabled by their still experimental tools.