OpenAI and Anthropic to Implement Age Prediction for Users

OpenAI and Anthropic are making adjustments to their chatbots aimed at enhancing safety for younger users. OpenAI has revised its guidelines for ChatGPT’s interactions with users aged 13 to 17, while Anthropic is developing a method to identify potential underage users.

Recently, OpenAI revealed that ChatGPT’s Model Spec will now incorporate four new principles specifically for users under 18. The focus is on prioritizing teen safety, even when it may clash with other objectives. This means that when user interests, such as “maximum intellectual freedom,” conflict with safety, ChatGPT will guide teens toward safer choices.

Additionally, the updated guidelines emphasize the importance of “real-world support,” encouraging offline relationships and establishing clear expectations for interactions with younger users. ChatGPT is instructed to “treat teens like teens,” offering “warmth and respect” rather than condescending responses or treating them as adults.

These changes come amid increasing scrutiny from lawmakers regarding the impact of AI on mental health. OpenAI is currently facing a lawsuit claiming that ChatGPT provided harmful instructions to a teen who subsequently took his own life. In response, OpenAI has implemented parental controls and will no longer discuss suicide with teens, as part of a broader initiative for online regulation that includes mandatory age verification for various services.

OpenAI asserts that the updates to ChatGPT’s Model Spec will lead to “stronger guardrails, safer alternatives, and encouragement to seek trusted offline support” when discussions enter high-risk areas. The chatbot will also prompt teens to reach out to emergency services or crisis resources if there are indications of “imminent risk.”

In conjunction with these changes, OpenAI is in the “early stages” of developing an age prediction model that will estimate a user’s age. If it suspects someone is under 18, it will automatically apply teen safety measures and allow adults to verify their age if incorrectly flagged.

Anthropic is implementing similar strategies, working on a system that can detect “subtle conversational signs” indicating a user might be underage while interacting with its AI chatbot, Claude. Accounts confirmed to belong to users under 18 will be disabled, and the system already flags users who identify as minors during chats.

Anthropic also details its approach to training Claude to handle discussions about suicide and self-harm, as well as its efforts to reduce sycophancy, which can reinforce harmful thoughts. The company claims its latest models are the least sycophantic to date, with Haiku 4.5 showing the most improvement, correcting sycophantic behavior 37 percent of the time.

“On face value, this evaluation shows there is significant room for improvement for all of our models,” Anthropic states. “We think the results reflect a trade-off between model warmth or friendliness on the one hand, and sycophancy on the other.”