OpenAI started testing a brand new security routing system in ChatGPT over the weekend, and on Monday launched parental controls to the chatbot – drawing blended reactions from customers.

The security options are available in response to quite a few incidents of sure ChatGPT fashions validating customers’ delusional pondering as a substitute of redirecting dangerous conversations. OpenAI is going through a wrongful demise lawsuit tied to 1 such incident, after a teenage boy died by suicide after months of interactions with ChatGPT. 

The routing system is designed to detect emotionally delicate conversations and routinely change mid-chat to GPT-5-thinking, which the corporate sees as one of the best outfitted mannequin for high-stakes security work. Particularly, the GPT-5 fashions have been skilled with a brand new security characteristic that OpenAI calls “secure completions,” which permits them to reply delicate questions in a secure approach, somewhat than merely refusing to interact.

It’s a distinction from the corporate’s earlier chat fashions, that are designed to be agreeable and reply questions shortly. GPT-4o has come underneath specific scrutiny due to its overly sycophantic, agreeable nature, which has each fueled incidents of AI-induced delusions and drawn a big base of devoted customers. When OpenAI rolled out GPT-5 because the default in August, many customers pushed again and demanded entry to GPT-4o. 

Whereas many specialists and customers have welcomed the protection options, others have criticized what they see as an excessively cautious implementation, with some customers accusing OpenAI of treating adults like youngsters in a approach that degrades the standard of the service. OpenAI has instructed that getting it proper will take time and has given itself a 120-day interval of iteration and enchancment.  

Nick Turley, VP and head of the ChatGPT app, acknowledged a few of the “sturdy reactions to 4o responses” because of the implementation of the router with explanations.  

“Routing occurs on a per-message foundation; switching from the default mannequin occurs on a short lived foundation,” Turley posted on X. “ChatGPT will inform you which mannequin is energetic when requested. That is a part of a broader effort to strengthen safeguards and be taught from real-world use earlier than a wider rollout.” 

The implementation of parental controls in ChatGPT obtained comparable ranges of reward and scorn, with some commending giving dad and mom a solution to preserve tabs on their childrens’ AI use, and others fearful that it opens the door to OpenAI treating adults like youngsters.

The controls let dad and mom customise their teen’s expertise by setting quiet hours, turning off voice mode and reminiscence, eradicating picture technology, and opting out of mannequin coaching. Teen accounts may even get extra content material protections – like decreased graphic content material and excessive magnificence beliefs – and a detection system that acknowledges potential indicators {that a} teen is perhaps serious about self-harm.  

“If our programs detect potential hurt, a small crew of specifically skilled individuals evaluations the scenario,” per OpenAI’s weblog. “If there are indicators of acute misery, we’ll contact dad and mom by electronic mail, textual content message and push alert on their telephone, until they’ve opted out.” 

OpenAI acknowledged that the system received’t be excellent and should typically increase alarms when there isn’t actual hazard, “however we predict it’s higher to behave and alert a guardian to allow them to step in than to remain silent.” The AI agency stated additionally it is engaged on methods to succeed in regulation enforcement or emergency companies if it detects an imminent menace to life and can’t attain a guardian.