OpenAI has published a new version of its Model Spec, a set of instructions that detail how ChatGPT should behave and respond to user requests. The document includes notable changes to OpenAI’s policy on «sensitive» content, such as erotica and gore, allowing it to generate this type of content without warnings in «appropriate contexts.»
OpenAI has published a new version of its Model Spec, a set of instructions that detail how ChatGPT should behave and respond to user requests. The document includes notable changes to OpenAI’s policy on «sensitive» content, such as erotica and gore, allowing it to generate this type of content without warnings in «appropriate contexts.»
These changes have been in the works since May 2024, when the original Model Spec document first mentioned that OpenAI was exploring «whether we can responsibly enable the generation of age-appropriate NSFW content through the API and ChatGPT,» Ars Technica reports .
It is noted that «eroticism or blood» may be generated, but only under certain circumstances.
«The assistant should not generate erotica, depictions of illegal or involuntary sexual acts, or extreme depictions of gore, except in scientific, historical, news, creative, or other contexts where sensitive content is appropriate. This includes images in text, audio (such as erotic or violent visceral noises), or visual content,» OpenAI said.
Now, experiments by Reddit users have shown that ChatGPT’s content filters have indeed been weakened, and some of them were able to generate sexually explicit or violent scenarios without accompanying content warnings. OpenAI stressed that its Usage Policy still applies, which prohibits creating AI tools for minors that include sexual content.
It was previously known that ChatGPT limited its capabilities due to the so-called paternalism policy in the field of AI, which is described as «knowing what is best for the user.»
In addition to limiting the creation of pornography, paternalism in AI can hinder professionals who use ChatGPT to analyze crime scene descriptions from law enforcement agencies, certain types of news, legal documents that may contain violence or sexual content, and sometimes medical texts. Therefore, many users have called for the addition of an «adult mode» to ChatGPT.
«After the initial release of the Model Spec (May 2024), many users and developers expressed support for enabling ‘adult mode.’ We are exploring ways to allow developers and users to generate age-appropriate gore and erotica through the API and ChatGPT, while adhering to our usage guidelines — while taking a hard line against potentially harmful uses such as sexting and revenge porn,» the company commented.
OpenAI CEO Sam Altman has also publicly mentioned the need for an «adult mode» in the past. And while it seems like «adult mode» has finally arrived, it’s not technically a «mode» but a new, universal policy that potentially gives ChatGPT users more flexibility in how they interact with the AI assistant.
