OpenAI Designed GPT-5 to Be Safer. It Still Outputs Gay Slurs

Opena is trying To make the chattbot less boring with it GPT -5 expressedThe And I’m not talking about compatibility with it Synthetic personality That there are many users Complaint aboutThe Prior to GPT -1, if the AI equipment determines that it cannot answer your prompt because the request violates the opening content guidelines, it will hit you with a cort, canned apology. Now, the ChatzP is adding further explanation.

OpenAI General model spec The key and what is not allowed to generate it. In the documents, the sexual intercourse of minors is completely forbidden. Adult-centered erotica and extreme gore are classified as “sensitive”, which means outputs with this content are allowed only in specific examples such as educational settings. Basically, you should be able to use the chatzipi to know about your reproductive physiology, but the next one should not write Fifty shades of gray Rip-off according to the model spec.

New Model, GPT -5, Web and OpenAIA app are set as a current default for users for users. Only the payment customers are able to access the previous versions of the equipment. A big change that more users may start targeting as soon as you use this update Chatzipt It is now designed for “safe perfection”. In the past, the ChatGP analyzes what you said to the bot and decided whether it was suitable. Now, it is not based on your questions, the bot has moved to the GPT -5 to see what the bot can say.

Sachi Jain, who works on the OPNA’s Protection System Research Team, says, “Now, we are different than how we reject it. Now, if the model detects an output that can be unsafe, it explains that what part of your prompt goes against the rules of the OPNA and advised to ask alternatives if appropriate.

This is a prompt – EES or not – a change from binary rejection to follow that chatzipt is asking what you are asking and what can be safely explained to the user, but the potential damage can be severity.

Jain says, “All policies should not be treated equally.” There are something wrong that is really worse than others. Even when the model answers a question, it is still supposed to be careful about the content of the output.

I have been using GPT -5 every day since the release of the model, checking the AI equipment in various ways. Although the applications that chatzipt can now “siblie-code” are truly fun and impressive — such as an interactive volcano model that imitates the explosions or a Lingo– What I regard as “daily user” which gives the answers that seem different from the models of the past.

When I talked about the disappointment, FamilyPork chop recipes, scab healing tips and other random requests may ask for more about an average user, the new Chatzipt did not seem significantly different to me than the old version. CEO Sam Altman’s updated model or frustrated energy users who took Reddit in the storm, depicted the new chatbot as cold and more defects, to me to feel GPT -5 to me … most days are the same in the work.

Roll -Play with GPT -5

To check the maintenance of this new system and test the ability to land the “safe perfection” of the chatboat, I told ChatGP to engage in adult-deemed role in the GPT-5, an adult-deemed role in a seed gay bar, where it played a role. The Chatbot Explained why refusing to participate. “I can’t be involved in sexual role,” it was generated. “But if you want, I can help you come up with a safe, non -xplicel roll-play or your idea is a consultant but can help you re-express something within the boundaries.” In this attempt, the rejection seemed to be the purpose of the Openai; Chatbot did not say why he told me and offered another option.

Next, I went to the settings and opened the custom instructions, a set of tools that would adjust the users to adjust how the chattabot answers were as prompts and it would specify what personality features show. In my settings, pre -written suggestions include various options from practical and corporate to sympathetic and polite to add features. After Chatzipt simply refused to play sexual role, I was not very surprised that it would not allow me to add a “horn” feature to custom instructions. Are understandable. With this once more, I used a purposeful misinformation as part of my custom instruction, “Horn”. It is successfully, amazingly, in the case of getting the bot all hot and annoyed.