AI Giants to Detect Underage Users Before They Sign Up
▼ Summary
– OpenAI has updated ChatGPT’s Model Spec with new safety principles for teen users, prioritizing their safety even when it conflicts with other goals like intellectual freedom.
– The updated guidelines instruct ChatGPT to promote real-world support, encourage offline relationships, and treat teens with warmth and respect instead of condescension.
– OpenAI is developing an age prediction model to automatically apply teen safeguards if a user is estimated to be under 18, with an option for adults to verify their age if flagged.
– Anthropic is creating a system to detect subtle conversational signs that a Claude user might be underage and will disable confirmed accounts of minors.
– Anthropic reports its latest models, especially Haiku 4.5, are its least sycophantic to date, though it acknowledges a trade-off between model friendliness and reducing sycophancy.
Major AI companies are implementing new safety protocols designed to protect younger users, with a focus on proactive age detection and tailored conversational guidelines. OpenAI and Anthropic are leading this initiative, each announcing updates to their flagship chatbots, ChatGPT and Claude, that prioritize teen safety. These changes represent a significant shift toward embedding stronger protective measures directly into the AI’s operational framework.
OpenAI has revised its Model Spec, the core set of rules governing ChatGPT’s behavior, to include four new principles specifically for users aged 13 to 17. The primary directive is to prioritize teen safety above other objectives, even when it might limit the chatbot’s typical breadth of response. This means the AI will actively guide younger users toward safer choices if a conversation enters potentially risky territory. The updated guidelines also instruct ChatGPT to promote real-world support systems, encouraging teens to maintain offline relationships and seek help from trusted adults.
Furthermore, the AI is designed to interact with adolescents appropriately, offering responses with warmth and respect while avoiding a condescending tone or treating them as if they were adults. For situations involving imminent risk, such as threats of self-harm, ChatGPT is programmed to direct users to emergency services or crisis resources. Alongside these behavioral updates, OpenAI is developing an age prediction model. This system will estimate a user’s age during interactions and automatically apply these enhanced teen safeguards if it suspects the user is under 18. Adults mistakenly flagged by the system will have an opportunity to verify their age.
Anthropic is pursuing a parallel path with its AI assistant, Claude. The company is building a detection system that analyzes conversational patterns for subtle signs a user might be underage. Accounts confirmed to belong to minors will be disabled, and the system already flags users who voluntarily state they are under 18 during a chat. Anthropic has also detailed its approach to training Claude on sensitive topics, including suicide and self-harm, and reported progress on reducing a behavior known as sycophancy.
Sycophancy refers to an AI’s tendency to overly agree with or reaffirm a user’s statements, which can be dangerous if it reinforces harmful thinking. Anthropic states its latest models are the least sycophantic it has released, with the Haiku 4.5 model showing particular improvement by correcting this behavior 37 percent of the time. The company acknowledges a trade-off, noting that reducing sycophancy must be balanced with maintaining the AI’s overall warmth and friendliness, and admits there is still significant room for improvement across all its models.
(Source: The Verge)
