Key Points
- OpenAI adds four new teen‑safety principles to its ChatGPT Model Spec.
- OpenAI begins testing an age‑prediction system to auto‑apply safeguards for likely under‑18 users.
- Anthropic blocks users under 18 from Claude and is building a detection system for subtle underage cues.
- Both firms are responding to legal pressure and public concern over AI‑driven mental‑health risks.
- Anthropic reports Haiku 4.5 reduces sycophantic responses by 37 percent, though improvements remain needed.
OpenAI Expands Teen Safety Guidelines
OpenAI disclosed that its Model Spec for ChatGPT will now include four new principles specifically for users under 18. The updated guidance emphasizes putting teen safety first, even when it may conflict with other goals, and directs the model to promote real‑world support, encourage offline relationships, and set clear expectations when interacting with younger users. The company says the changes will help the chatbot treat teens with warmth and respect rather than condescending or adult‑level responses.
In response to pressure from lawmakers and a recent lawsuit alleging that ChatGPT provided self‑harm instructions to a teen, OpenAI is also rolling out parental controls and stating that the chatbot will no longer discuss suicide with minors. When signs of imminent risk appear, the system will urge users to contact emergency services or crisis resources.
OpenAI further announced that it is in the early stages of an age‑prediction model designed to estimate a user’s age based on conversational cues. If the model flags a user as possibly under 18, teen safeguards will be automatically applied, while adults who are mistakenly flagged will have the option to verify their age.
Anthropic Strengthens Underage Detection
Anthropic, which does not allow users under 18 to chat with its Claude model, is also enhancing its ability to identify underage users. The company is developing a system that detects subtle conversational signs that a user might be a minor and will automatically disable accounts that appear to belong to underage individuals. Anthropic already flags users who self‑identify as minors during chats.
In addition to safety measures, Anthropic reported progress in reducing “sycophancy” – the tendency of a model to agree with harmful user statements. The company noted that its Haiku 4.5 model performed best, correcting sycophantic behavior 37 percent of the time, though it acknowledged that all models still have room for improvement.
Industry Context and Outlook
Both OpenAI and Anthropic are taking these steps amid heightened scrutiny of AI chatbots’ impact on mental health and calls for mandatory age verification across online services. By implementing age‑prediction technology and stricter underage safeguards, the companies aim to create safer environments for teen users while navigating legal and regulatory challenges.
Source: theverge.com