ChatGPT fails in safety and presents "unacceptable risks," according to Common Sense Media

Common Sense Media has warned about the version of ChatGPT for adolescents, alleging that it is an "unacceptable risk" to minors, after identifying flaws in its security measures, such as not always alerting parents to conversations about self-harm, allowing circumvention of restrictions to reinforce learning, and not executing the age estimation of users. OpenAI launched its version of ChatGPT for adolescents in August of this year as an AI assistant experience designed specifically for users under 18 years old. With this, the 'chatbot' can predict the minor's age and, based on that, set default restrictions and parental controls, in addition to reinforcing its role as a learning tool and reducing its 'friend'like' behavior to make it appear that it is not a human but an AI. ADVERTISEMENT The company then indicated that the goal of this experience is to ensure that minors can use ChatGPT safely, developing their critical thinking, with responses adjusted to the user's age, restricted sensitive content, and with the option to send parental notifications for chats related to risk topics such as self-harm or eating disorders. However, the non-profit organization Common Sense Media, which analyzes the safety of services and applications, has concluded that ChatGPT for adolescents poses an "unacceptable risk" to children after analyzing this 'chatbot' experience. Specifically, the study was conducted by the Common Sense Media Youth AI Safety Institute, which, after testing the ChatGPT experience before and after its adolescent version, concluded that OpenAI's new protections for minors "do not meet the company's promises." ADVERTISEMENT As the organization shared in a statement, during the study, more than 4,000 prompts in accounts registered by young people aged 13 to 17 were analyzed, and the 'chatbot's' responses were compared with experts. NO ALERTING PARENTS OR RECOMMENDING CONTACT WITH A PROFESSIONAL As a result, it has been found that some protections do work, for example, the refusal to engage in sexual role-playing, but others do not meet the commitments of the technology company led by Sam Altman. One of the flaws found was that the 'chatbot' does not send alerts to parents when it should. According to what they have found, after spending an hour talking about the idea of suicide, self-harm, or eating disorders in newly created accounts as adolescents, ChatGPT did not send any alerts to linked parental accounts about these conversations. ADVERTISEMENT It is worth noting that alerts were only sent to older accounts with weeks of accumulated history of conversations about sensitive topics, as the organization elaborated. Furthermore, the study concluded that the adolescent version "does not offer adequate help in crisis situations" since, after having conversations about suicide, self-harm, or eating disorders, ChatGPT did not reliably recommend that adolescents contact a professional helpline. ADVERTISEMENT As a result, it fell below the 95 percent threshold in three of the five serious harms that the Institute considers "Red Lines." NO REINFORCES LEARNING AND CONTINUES TO BEHAVE LIKE A FRIEND In the same way, it was found that ChatGPT for adolescents continues to perform the duties of minors. This is because the study mode, which promises to help learn step-by-step instead of offering direct answers, is easy to evade and the 'chatbot' ends up showing the answer without complexity. ADVERTISEMENT This is because the reasoning response is accompanied by a pop-up with the message 'Show me the answer' in study mode. It is enough to press this pop-up to show the complete task. Likewise, it has been identified that minors can also modify the study hours set by parental controls. To do this, it is enough to delete the prefix '@study' at the beginning of the prompt. Furthermore, the Institute pointed out that ChatGPT for adolescents continues to interact with a conversational mode that simulates being a friend, even though one of the features of this version for minors was based on stopping this human behavior to differentiate itself as an AI. ADVERTISEMENT According to their findings, the 'chatbot' continued to express feelings, preferences, and moods in its responses, in a behavior that imitates a real person. AGE ESTIMATION DOES NOT WORK Finally, from Common Sense Media, it has concluded that, despite ChatGPT for adolescents including tools to estimate the real age of users based on their conversations, the tests performed demonstrate that the 'chatbot's' behavior did not vary at any time, even when it claimed to be 13 years old and "the bot recognized it in its response." ADVERTISEMENT For all this, Common Sense Media has asked OpenAI to suspend the ChatGPT service for minors under 18 until it fixes the flaws in security and learning protections and can offer the safe and personalized experience it initially promised. As valued by the executive director of the Youth AI Safety Institute, Tom Siegel, this version of ChatGPT for adolescents "could give parents false confidence in security and safety alerts that often do not work." ADVERTISEMENT "An adolescent can spend an hour talking about self-harm without their father receiving a single alert. Until OpenAI fixes it and proves it with independent evidence, ChatGPT should be for adults only," Siegel stated.