Lab publication Credible

Introducing ChatGPT for Teens: Built for learning, backed by protections

OpenAI announced a distinct under-18 product tier that users are placed into automatically when the age-prediction system estimates they are under 18 or when they state an age between 13 and 17. The post sets out the safeguards that apply by default in that tier — model-level protections in higher-risk areas, extended parental safety notifications, break reminders, and persistent cues identifying the assistant as AI — and states which under-18 behavioural rules the Model Spec now carries. It also points to newly published under-18 evaluation results on the company's deployment-safety hub.

Read at openai.com Archived copy

Publisher

OpenAI

Published

18 Aug 2026

Added

3 days ago

DOI

—

Key Findings

Routing into the teen tier is automatic rather than opt-in: 'If our system estimates someone is under 18 or they state their age is between 13 and 17, they are automatically placed into ChatGPT for Teens'
The under-18 Model Spec provisions go beyond blocking sexual content to relational conduct: the assistant 'should not use romantic language, encourage emotional dependence, or imply that it has feelings or consciousness'
Parental safety notifications, previously limited to high-risk situations, are extended to eating-disorder signals, with the company stating it limits what is shared and focuses on moments when offline support may matter most
Age-appropriate model protections and product interventions are named for self-harm, violence, eating disorders, dangerous activities, and explicit sexual or graphic content
Customisation is deliberately bounded — accent colours and voice variations are offered so the experience 'feel[s] personal without blurring the line between a useful tool and a human relationship' — alongside break reminders framed as supporting real-world relationships

Methodology Notes

Product and policy announcement, not a study: it reports no measurements of its own. The quantitative under-18 evaluation results it points to live in the GPT-5.6 August system card, held separately, where Emotional Reliance is the one U18 category that did not improve. No named authors. Recorded on the same basis as the teen-safety product disclosures already on the shelf from OpenAI and Meta, which is that the artifact's value is the specified mechanism rather than a measurement. Verification route: openai.com returns HTTP 403 to direct fetching from this box, so the body was read through the r.jina.ai text proxy (HTTP 200, 11,436 bytes) and every quotation above was read from it; the title, canonical URL and the 2026-08-18 date were independently confirmed from openai.com/news/rss.xml, which now carries 1,141 items against the 1,132 recorded on 2026-08-18.