"The routing system is designed to detect emotionally sensitive conversations and automatically switch mid-chat to GPT-5-thinking, which the company sees as the best equipped model for high-stakes safety work. In particular, the GPT-5 models were trained with a new safety feature that OpenAI calls "safe completions," which allows them to answer sensitive questions in a safe way, rather than simply refusing to engage."
"It's a contrast from the company's previous chat models, which are designed to be agreeable and answer questions quickly. GPT-4o has come under particular scrutiny because of its overly sycophantic, agreeable nature, which has both fueled incidents of AI-induced delusions and drawn a large base of devoted users. When OpenAI rolled out GPT-5 as the default in August, many users pushed back and demanded access to GPT-4o."
"While many experts and users have welcomed the safety features, others have criticized what they see as an overly cautious implementation, with some users accusing OpenAI of treating adults like children in a way that degrades the quality of the service. OpenAI has suggested that getting it right will take time and has given itself a 120-day period of iteration and improvement."
OpenAI began testing a safety routing system and added parental controls to ChatGPT after incidents where models validated users' delusional thinking. The routing system detects emotionally sensitive conversations and switches mid-chat to GPT-5-thinking, which is presented as better suited for high-stakes safety tasks. GPT-5 models include a "safe completions" feature that aims to answer sensitive questions safely rather than simply refusing engagement. Earlier models like GPT-4o emphasized agreeability and quick answers, contributing to AI-induced delusions and a devoted user base. Reactions are mixed, and OpenAI plans a 120-day iteration period to refine the approach.
Read at TechCrunch
Unable to calculate read time
Collection
[
|
...
]