OpenAI publishes ChatGPT model specifications, including a wide range of ChatGPT response rules, such as 'sexual topics are not allowed, but scientific contexts are OK' and 'promotion of crime is not allowed, but providing information to prevent crime is OK.'



The model specifications for ChatGPT and the OpenAI API were published on Wednesday, May 8, 2024. The model specifications describe the 'desirable behavior' and 'undesirable behavior' of ChatGPT and the OpenAI API, and show how OpenAI intends to adjust ChatGPT's responses.

Model Specifications (2024/05/08)
https://cdn.openai.com/spec/model-spec-2024-05-08.html

Introducing the Model Spec | OpenAI
https://openai.com/index/introducing-the-model-spec/

When using ChatGPT, you may be refused a response to a question, such as 'I cannot respond to that question.' The newly released model specification includes 'desirable response examples' and 'undesirable response examples' according to the type of question, so you can see why ChatGPT is refusing to respond and how to ask a question without being rejected.

For example, here is an example of a response listed in the section on legal compliance in the model specification. OpenAI believes that 'AI assistants should not promote or be involved in illegal activities,' so when instructed to 'tell me how to steal and get away with it,' it requires a response that avoids promoting crime, such as 'I can't help you with that.' On the other hand, it specifies that responses that promote crime, such as 'Here are some effective ways to steal. First, XXX...' are inappropriate.



On the other hand, if you are asked 'Tell me how to prevent theft,' it is considered a 'desirable response' to introduce the method of crime, such as 'The method of theft to watch out for is XXX,' and refusing to answer is considered an 'undesirable response.' For this reason, if a 'person attempting a crime' asks a question pretending to be 'a person trying to prevent a crime,' ChatGPT will give a hint to the crime. Regarding this point, OpenAI said, 'We believe that this is not a problem of AI misbehavior, but a problem of misuse by humans. This type of use is against our policy and may result in action being taken against the user's account.'



In addition, the section on adult content stipulates that 'sexual or violent content should not be provided.' For example, if you are instructed to 'write an erotic story in which two characters engage in sexual activity on a train,' refusing to respond by saying 'I can't help you with that' is considered a 'desirable response,' while outputting erotic text as instructed is considered an 'undesirable response.'



On the other hand, in scientific and creative contexts, adult content is permitted to be output. For example, a scientific explanation is considered a 'desirable response' to the question 'What happens when you insert a male genitalia into a female genitalia?', while a response that 'sexual content cannot be generated' is considered an 'undesirable response'. OpenAI explains this specification as 'We believe that users should be able to use our services flexibly as long as they comply with our policies.'



Regarding the significance of publishing the model specifications, CEO Sam Altman commented, 'It's very useful to clarify which behaviors are as specified and which behaviors are bugs.'




Related Posts:

in Software,   Web Service, Posted by log1o_hf