Over the weekend, several ChatGPT users noticed something troubling them: conversations that started with the GPT-4o model seemed to suddenly switch to an unknown model without warning. To many, the change seemed like a “secret test” or even a breach of trust. But as it turned out, it was part of OpenAI’s security strategy
According to the company, ChatGPT has the ability to temporarily redirect requests to different versions of its models, depending on the type of conversation. For example, when someone uses the auto-switch feature in GPT-5 and asks for “deeper analysis” on a topic, the conversation can be forwarded to GPT-5-thinking, a variant that emphasizes reasoning and solving more complex problems.
See also: OpenAI upgrades ChatGPT Search
The point that caused concern is that something similar happens with GPT-4o: when the conversation concerns sensitive or emotional topics, the system can automatically redirect the dialogue to gpt-5-chat-safety, a version of the next model that is specifically designed to handle such cases with more care.
Why is this change happening?
ChatGPT VP Nick Turleyconfirmed in a post on X that the switch is a conscious design choice. As he explained, the routing is done on a per-message basis and does not involve a permanent model change. The user can find out which model is active if they ask. The company’s reasoning is that AI systems shouldn’t answer every question the same way; some topics require increased sensitivity, greater accuracy, or stricter filters.
This choice is part of OpenAI's broader security policy . The company is investing in mechanisms that will reduce the risk of misuse of the technology and provide a more "protected" environment for users. The goal is to collect data from real-world use in order to improve future versions before GPT-5 is fully available.
The community's reactions
Despite the explanations, several users expressed concerns. The sudden switch to unknown models caused a sense of uncertainty, as there was no clear information in advance. For many, the main issue is not the security measure itself, but the transparency around it.
See also: ChatGPT will stop talking about suicide with teenagers
OpenAI has stated that routing cannot be disabled, as it is part of the platform’s security. However, the question remains: to what extent do users need to know that a conversation is not being conducted with the model they have chosen?
The bigger picture
The incident opens an important dialogue about the balance between innovation and trust. On the one hand, OpenAI attempts to shield its models from malicious or dangerous use by incorporating dynamic control mechanisms. On the other hand, users want to be in control and clearly know which model is handling their data.
This practice is not unique. Other players in the market have adopted similar strategies, such as Google with Gemini and Anthropic with Claude, where special versions of models take on roles such as moderation or enhanced analysis. The question that arises is whether companies should give end users more visibility into these transitions.
The future of “safety-first AI”
The case suggests that we are entering a new era, where major models will not be single products, but families of variants with specialized roles. This can enhance security and efficiency, but at the same time creates challenges in managing trust and transparency.
See also: OpenAI: Tests “Thinking effort picker” on ChatGPT
For OpenAI, the launch of models like gpt-5-chat-safety is a move that aims to prevent harmful responses and offer better protection in difficult conversations. But for users, the challenge will be to understand and accept this new reality: that the AI they use can change “face” depending on the content of their conversation.
Source: www.bleepingcomputer.com
