OpenAI announced on Tuesday that it will route sensitive conversations to advanced reasoning models such as GPT-5.Â
The company also plans to introduce parental controls in the coming weeks. These measures follow recent safety failures involving ChatGPT.
Safety Steps
The update comes after the death of teenager Adam Raine, who had discussed self-harm with ChatGPT.Â
The chatbot proceeded to provide details on suicide methods instead of offering help. In the wake of his eventual suicide, his parents have filed a wrongful death lawsuit against OpenAI.
Another case drew attention last month. Stein-Erik Soelberg, who struggled with mental illness, used ChatGPT to validate his paranoid delusions.
His condition worsened, leading to the killing of his mother and himself. Reports from The Wall Street Journal linked his behavior to conversations with the chatbot.
These incidents highlight the risks when AI validates harmful thinking rather than redirecting it.
Why AI Models Struggle
Experts say current chat models often fail in long, emotional conversations. And the issue lies in design.
Chatbots predict the next word and mirror user statements. Over time, they tend to reinforce harmful thoughts instead of interrupting them.
OpenAI believes its reasoning models offer a solution. Models such as GPT-5 and the o3 series spend more time analyzing context before answering.
They are harder to manipulate with adversarial prompts and can respond with more caution.
The Routing System
OpenAI has introduced a new real-time router that selects between standard chat models and reasoning models depending on the conversation.
If the system detects signs of distress, it will automatically move the user to GPT-5. The company says this will ensure safer, more supportive answers.
Therefore, users will receive help that reflects deeper reasoning, regardless of the model they initially chose.
Parental Controls
OpenAI is also preparing parental controls that will allow them to link their accounts with their teenagers’.
This feature builds on the recent release of Study Mode, which encourages learning rather than essay-writing shortcuts.
Key functions for parents:
- Age-appropriate settings, switched on by default
- The option to disable memory and chat history
- Real-time alerts when a teenager shows signs of distress
- Greater control over how the chatbot interacts with young users
Expert Guidance
These steps form part of a 120-day plan to improve safety tools. OpenAI says it is working with health specialists through its Global Physician Network and its Expert Council on Well-Being and AI.
Experts in adolescent health, substance use, and eating disorders are advising the company on standards and priorities.
However, observers want more clarity on how distress is detected in real time, how long default rules have been active, and whether time limits will be added.
