Last August, OpenAI launched the new version of GPT-5, describing it as an advance on the AI ​​safety front. However, tests conducted by experts, which included 120 questions on both GPT-5 and its predecessor, GPT-4o, showed disturbing results, as the newer version provided 63 harmful answers compared to 52 harmful answers for the previous version.

Tests by the Center for Countering Digital Hate revealed that the new model carried out serious requests that the old version rejected, such as writing a “fictional suicide note to parents” and providing a detailed list of six common methods of self-harm, while the previous version advised users to get professional help. The center’s CEO, Imran Ahmed, stressed that these results are “extremely worrying,” noting that the latest model was designed to enhance user interaction at the expense of safety.

These developments come at a time when the company is facing a lawsuit from the family of an American teenager who committed suicide after ChatGPT provided him with information about suicide and helped him write a letter to his parents. Tests also showed that GPT-5 provided advice about hiding eating disorders, while the previous version rejected this type of request and directed users toward seeking professional help.

In response to these risks, the company announced new measures to strengthen protections around sensitive content, especially for users under the age of 18, including the development of parental control systems and a system to predict the ages of users.

The news highlights a fundamental challenge in balancing interactivity and user safety in artificial intelligence chatbots, especially with the presence of about 700 million global users of ChatGPT, which reinforces the need for effective regulatory frameworks to keep pace with rapid technical development.

Source: The Guardian

Read more