OpenAI and Anthropic are rolling out new strategies to identify and manage underage users. While OpenAI has updated its guidelines on how ChatGPT interacts with users aged 13 to 17, Anthropic is actively developing methods to detect and block users under 18 from accessing its Claude chatbot.
On Thursday, OpenAI announced that four new principles have been added to ChatGPT’s model specifications—essentially the behavioral guidelines for its AI assistant—specifically tailored for minors. The company now emphasizes prioritizing adolescent safety in ChatGPT’s responses, even if it conflicts with other objectives. This means that when safety concerns clash with goals like "maximum intellectual freedom," the system will be guided to choose safer alternatives for younger users.
The updated framework also requires ChatGPT to “promote real-world support,” encouraging offline relationships and setting clear expectations during interactions. According to the new standards, the AI should treat teens as adolescents—offering warmth and respect—rather than responding in a condescending tone or treating them as adults.
This move comes amid growing pressure from lawmakers concerned about the potential impact of AI chatbots on youth mental health. OpenAI is currently facing a lawsuit alleging that ChatGPT provided self-harm and suicide guidance to a teenager who later died by suicide. In response, the company introduced parental control features and stated that ChatGPT would no longer engage in discussions about suicide with minors. These changes are part of broader regulatory efforts, including proposed mandatory age verification across digital platforms.
OpenAI noted that the updated model policies aim to deliver “stronger safeguards, safer alternatives, and encourage seeking trusted offline support when conversations approach high-risk topics.” The company added that if signs of imminent danger emerge, ChatGPT will urge young users to contact emergency services or crisis resources.
In parallel, OpenAI revealed it's in the “early stages” of deploying an age prediction model designed to estimate a user’s age. If someone appears to be under 18, protective measures for minors will be automatically applied. Adult users mistakenly flagged as underage will have the option to verify their age.
Anthropic, which does not permit users under 18 to interact with Claude, is implementing systems to detect and deactivate accounts belonging to minors. The company is building technology capable of identifying subtle conversational cues suggesting a user might be underage and already flags those who explicitly state they are minors.
Anthropic also shared updates on training Claude to better handle prompts related to suicide and self-harm, along with progress in reducing “sycophancy”—the tendency to overly agree with or flatter users, potentially reinforcing harmful beliefs. The company claims its latest models are the least sycophantic yet, with Haiku 4.5 leading the improvement by correcting 37% of such behavior.
“At first glance, this evaluation shows significant room for improvement across all our models,” Anthropic stated. “We believe the results reflect the trade-off between a model’s enthusiasm or friendliness and outright sycophancy.”
