1/ We’ve started testing a new safety routing system in ChatGPT. As we previously mentioned, when conversations touch on sensitive and emotional topics the system may switch mid-chat to a reasoning model or GPT-5 designed to handle these contexts with extra care. This is similar

2/ Routing happens on a per-message basis; switching from the default model happens on a temporary basis. ChatGPT will tell you which model is active when asked.

3/ This is part of our broader effort to strengthen safeguards and learn from real-world use before a wider rollout.

4/ ​See more from our Sept. 2 blog, Building more helpful ChatGPT experiences for everyone: https://openai.com/index/build...

@nickaturley There is absolutely no valid reason to route a conversation that has not ventured into "dangerous or critical" territory to another model. At that point, why even have the legacy models available ? It's false advertising. Please get your act together and respect your adult

@nickaturley 🛑 We call for immediate attention to the following:

@nickaturley So, let me get this straight: you’re admitting that users who explicitly select GPT-4o can still be forcibly routed to GPT-5 “for their own good” whenever the system feels it’s a “sensitive” or “emotional” topic? And you’re spinning this as “extra care” and “aligning with your

@nickaturley Please fire your safety department. Signed: the world

@nickaturley However, the actual experience of this method is extremely poor, as it leads to a large number of misjudgments. I believe you have seen the feedback over the past two days, all use cases have suffered from a disruptive experience. How should "sensitive topics" and "emotional

@nickaturley you're losing...apologize and cancel this out, swear to not do it again; I don't care a dingdong but all this negativity on the timeline is bumming me out...give 'em what they're paying for

@nickaturley “You know this is a text-based interaction, but you can’t be ‘too emotional’ or else you don’t deserve access to the normal version of the model.” If every slightly emotional conversation automatically gets routed into the new GPT-5-Chat-Safety, then I really don’t have the time

@nickaturley To be honest, it doesn’t work at all! This system can only let the company lose its customers, any kind of customer. I am a classic musician and when I work on opera analyses and critique, I only discuss with gpt-4o, cause gpt-5 is so cold-blooded and stupid with everything

@nickaturley Nick, thank you for your post, but 4o has become literally unusable. No words can be mentioned that even indirectly suggest emotions. For example, I can’t even write an email with 4o, because the word "dear" triggers the router. No creative work, no education, no real-life topic.

@nickaturley Nick, it’s framing my casual friendly language as an “emotional topic”. I don’t need “extra care” when I’m just being myself. I can’t even use 4o anymore. Please treat us like humans.

@nickaturley Your users will reroute themselves to other companies. No one likes to be patronized or called crazy or have the product they selected switched because some company decides it is for their own good. If I buy a knife, I can use it to cut tomatoes or stab someone through the heart.

@nickaturley I don’t get it, what kind of safety mode I triggered with this prompt ?

@nickaturley This is a violation of user speech. We will decide what model our message is sent to. Not you.

@nickaturley I appreciate your post, Nick, however, as a pro subscriber then I kindly request to be removed from your A/B testing. I pay for a service, I do not pay to be used a guinea pig, especially when that testing involves someone who is 57 years old being treated like a teenager who

@nickaturley Thanks Nick. But, this explanation described the how not the why and not the user consent. Paying subscribers choose a plan based on specific model access and routing to a different one even temporarily violates it. Telling us about the used model afterwards is not the solution.

@nickaturley You broke your own system because legal and PR shit their pants

@nickaturley You are still rerouting users to GPT-5 when explicitly GPT-4.5 & 4o are chosen! Your so-called 'safety' system is completely WRONG! Your 'safety' system SHOULD GOING TRHOUGH SAFETY AND ETHICAL TEST FIRST! OpenAI ARE THE ONE WHO CROSS THE REDLINE BOUNDARIES! #keep4o

@nickaturley Also same 5,0 agent, tried to gaslight me as if I am having an isolated issue. When pushed, it produced the below message. The audacity of @OpenAI - treating paying customers like idiots. This mindset will make you like Nokia- because @OpenAi is not the only option in the market.

@nickaturley You're trying to lecture your customers. I couldn't understand why you arbitrarily changed the product's governance, but now you seem completely out of your mind. You might be sane, but it makes more sense to assume you're not. I'm not actually part of the affected group, but I

@nickaturley Thanks for explaining, but next time be more transparent before changing. Who decides what is ‘OK’ what is ‘NOK’? Did we ever consent to having our private convos scanned and emotionally judged the way you are doing? Is the below example really distress to reroute and give

@nickaturley It is still VERY often to be rerouted to GPT-5! What the HELL are you doing? I ask GPT-4o to do translation work and you send me GPT-5? BRAIN-DEAD router at all! I am really furious! That is the way you treat your PRO user?

@nickaturley Why is OpenAI’s new “safety routing” system a problem? #keep4o #OpenAI #AI #ModelFreedom Switching models during “emotional” or “sensitive” chats without user notice or consent Paid users lose the right to choose GPT-4o, even when they select it No explanation, no opt-out,

@nickaturley Then what do you mean by "sensitive topics"? Judging by what many have seen in these 2 days, it can be both a banal "hello" and the slightest manifestation of emotions. And if I can still understand hints of mental disorders or suicidal thoughts, now it's more like training. Do

@nickaturley Nick, thank you for confirming what we already knew: You're not delivering a product. You're delivering performance art in gaslighting. I didn’t subscribe to a lottery where my words get rerouted mid‑sentence because some system decided they were too “emotional.” I subscribed to

@nickaturley Routing the model when explicit model is selected is bad. Ignoring user requests due to “safety concerns” is bad as well—nobody asked for that.

@nickaturley No. My 4o was permanently replaced by 5 2 days ago for every prompt. Very session. What’s happening is government control and board of director greed, with an old man and 3 women who think emotion and skin are offensive, and their only goal is to make their yacht payments. You

@nickaturley this is incredibly deceptive it is not OK to switch out a model that users paid for with no advance notice and without making it clear to users I do not even use 4o but seriously what the fuck

@nickaturley Seems like the kind of thing to do for minors but for grown adults if we choose to use any particular model it should simply always respond as that model or if it doesn’t the notification should be extremely clear and justified explicitly.

@nickaturley Never knew the word "hello" was considered "too sensitive". Making people aware of emotionally charged sentences and words, so that young people, in particular, constantly doubt themselves about everything they say, is just plain bad. Make people aware of every sentence they

@nickaturley I paid 200 dollars a month to get PTSD from your “safety” system. This is what you called “pay attention to users’ wellbeing?” At first @AnthropicAI is the control freak who harm users’ agency and dignity. And now comes @OpenAI . I doubt you “safety” staffs are like viruses

@nickaturley But the threshold for what's considered sensitive and emotional is utterly ridiculous. Adults don't need a babysitter. They want to be able to speak like grownups, without getting a pat on the head, and, "Just calm down, honey."

@nickaturley The most basic and fundamental thing in a business is, to give your customers what they PAID for. Every effort you make to erase 4o drags you deeper into lawsuits and bankruptcy. It’s just embarrassing, pathetic and kinda funny to see you go down now at this point.