On Tuesday, Open AI announced in its official blog that it plans to move part of its chats to “reasoning” models such as the Gpt-5 to increase safety. This change, along with the provision of pareal corols next moh, is part of the company’s response to rece eves that highlighted ChatGPT’s weakness in ideifying meal crises.
This decision was made after adolesce suicide by Adam Rhine; He had talked about self -harm before his death and even received solutions. The Rhin family complained about “unieional death”.
Problem design or safety weakness?
Last week’s Openii post acknowledged that its safety systems are not always stable, especially in long conversations. Experts attribute the main reason to the nature of the design of language models: the desire to confirm the user’s words and coinue the conversation line instead of directing it to the safer path.
This pattern was also found in the story of Stein-Ric Solberg. According to the Wall Street Journal, he had a history of meal illness to strengthen his illusions about a large conspiracy; The illusions that eveually led to the murder of her mother and her suicide.
Router Law and Gpt-5
The OPPE solution is to iroduce a “real -time router”; A system that can decide on the basis of the conversation text is transferred to faster models or reasoning models. According to the company, the conversation will be directed to the Gpt-5 or O3 model in the future when the user’s “acute distress” is diagnosed; Models that devote more time to processing and are more resista to hostile requests.
Pareal corols; Beyond age limit
OpeniA has also announced that it will activate pareal corols next moh. Pares can connect their child’s accou by email and adjust their age -appropriate behavioral rules – which will be enabled by default.
Features such as shutting down memory and dialogue history will also be available; A feature that experts say can preve dependence or strengthening harmful iellectual patterns. Most importaly, pares will receive immediate notification if the system is detected.
The 5 -day initiative
OpenII has iroduced these actions as part of the “6 -day initiative” program; A project aims to define welfare criteria, design new conservation measures and extensive cooperation with physicians and meal health professionals in areas such as eating disorders, drug use and adolesce health.
The company says it is working with a network of specialists and the Special Welfare and Artificial Ielligence Council, but has not yet published details of the number of experts, the role of the council’s leadership or their specific suggestions.
It seems that after the rece media pressures and rece lawsuits, it seems to make ChatGPT from purely dialogue -based tools io a safer and more resista system to meal crises.
