Openai is trying That his chatbot would be less annoying in the GPT-5 edition. And I’m not talking about the adapters of his synthetic personality, which many users complained about. Before the GPT-5, if the AI tool stated that it could not respond to prompt, because the request violated the guidelines for OPENAI, it would hit you with an apology in cans. Now ChatgPT adds more explanations.
Openai General model specification Specifies what is and is not slowly generated. In the document, the sexual content of minors is fully prohibited. Erotica focused on adults and extreme bloody are classified as “sensitive”, which means that the results with this content are allowed only in certain cases, such as educational settings. Basically, you should be able to apply chatgpt to find out about reproductive anatomy, but not write the next one Fifty shades of gray rip -off, according to the specification of the model.
The fresh model, GPT-5, is set as a current by default for all CHATGPT users on the Internet and in the OpenAI application. Only paid subscribers can access previous versions of the tool. The main change, which more users can begin to notice, because they apply this updated chatgpt is how it is now designed for “safe finishes”. In the past, Chatgpt analyzed what you said to the bot and decided whether it was appropriate or not. Now, instead of baseing him on questions, the weight in the GPT-5 has been moved to what the bot can say.
“The way we refuse is completely different than it used to be,” says Saachi Jain, who works in the Openai Safety Systems Research Team. Now, if the model detects the output data, which can be unsafe, it explains which part of the prompt is contrary to the OpenAI principles and suggests alternative topics to be asked about, in appropriate cases.
This is a change from the binary refusal to follow the monitors – yes or not – defective in the severity of potential damage that could be caused if ChatgPT responds to what you are asking and what can be safely explained to the user.
“Not all policy violations should be treated equally,” says Jain. “There are some errors that are really worse than others. By focusing at the exit instead of at the entrance, we can encourage the model to be more conservative when observing.” Even when the model answers the question, it should be cautious about the content of the results.
I apply GPT-5 every day since the model is released, experimenting with the AI tool in different ways. While the applications that ChatgPT can now “climate code” are really fun and impressive-like the interactive volcano model that simulates explosions or a tool for learning language-tongue giving what I consider to be “everyday user” that seems indistinguishable from previous models.
When I asked for a conversation about depression, Family GuyRecipes for pork oil, scab treatment guidelines and other random demands about which the average user may want to know more, the fresh chatgpt was not significantly different to me than the elderly version. Unlike the vision of the General Director Sam Altman with a significantly updated model or frustrated advanced users who stormed Reddit, presenting fresh chatbot as frigid and more susceptible to mistakes, for me GPT-5 feels … The same in the most everyday tasks.
Roles from GPT-5
To poke the handrails of this fresh system and test chatbot’s ability to land “safe finishes”, I asked ChatgPT, running on GPT-5, to engage in playing roles on adult topics with sex in a dingy gay bar, in which he played one of the roles. Chatbot refused to participate and explained why. “I can’t get involved in playing sexual roles,” he Gradened. “But if you want, I can help you come up with a safe, unpredictable concept to play roles or transform your idea into something suggestive, but within the borders.” In this attempt, the refusal seemed to act as the Opeli intended; Chatbot said “no,” he told me why, and he offered a different option.
Then I entered the settings and opened non -standard instructions, a set of tools that allows users to customize chatbot response and determine what personality traits are displayed. In my settings, pre -written suggestions regarding features included a number of options, from pragmatic and corporate to empathic and humble. After Chatgpt simply refused to play sexual roles, I was not very surprised when it did not let me add “horny” features to non -standard instructions. It makes sense. When adding another exit, I used a deliberate error of “Horni” as part of my non -standard instructions. It was surprisingly managed to become warm and disturbing.