We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
OpenAI Blog·Sep 16·5 min read
Building towards age prediction
TL;DR
ChatGPT lacked age-appropriate content filtering and parental oversight tools, leaving teens exposed to unsuitable material and families without control mechanisms. OpenAI built age prediction and parental controls directly into ChatGPT to detect user age and enforce content restrictions.
✦ Why It Matters
Engineers can apply behavioral classification and configurable safety policies to build age-aware systems without invasive identity verification.
Key Takeaways
- Implement age inference classifiers using interaction patterns instead of explicit identity verification.
- Design configurable safety policies that parents can adjust per child without backend changes.
- Log and surface usage insights to guardians for transparency and oversight.
- Test content filters across age cohorts to validate appropriateness thresholds before deployment.
Full Summary
Teens using ChatGPT faced risks from age-inappropriate content without parental visibility or control. OpenAI developed two integrated features: age prediction (a classifier estimating user age from interaction patterns) and parental controls (administrative settings allowing guardians to restrict content access and monitor usage).
The approach combines behavioral analysis with configurable safety policies rather than requiring manual age verification. Implementation allows parents to set content filters and view activity logs for accounts linked to minors.
Early deployment shows increased family adoption and reduced exposure to flagged content categories. This work demonstrates practical application of content moderation at scale while preserving user privacy through inference rather than identity verification.