OpenAI splits teens into a separate ChatGPT as states sue Meta over feeds
OpenAI now puts 13 to 17 year olds into a separate ChatGPT with tighter defaults. Nearly half of real chat traffic falls outside companies' work focused filters. Twenty nine US states opened a trial over how Meta builds its feeds.
Artificial Intelligence··Evening
OpenAI applies a teen ChatGPT automatically
OpenAI said in a Tuesday blog post that ChatGPT for Teens now applies automatically to users who identify themselves as 13 to 17 and to those the system estimates are under 18. Restrictions on graphic violence, self-harm depictions and sexual or romantic roleplay sit on by default, and parents can set quiet hours and receive a notice when the system flags a possible safety risk. Most of the package gathers protections already shipped: age prediction at the start of the year, parental controls and study mode about a year ago, and more frequent break reminders last month. The new pieces are a teen first-run flow, a prompt before sensitive image uploads, accent colours and voice options, and homework reminders that try to spot a teen shortcutting an assignment and steer them into study mode. Teens or parents can set study hours that switch study mode on automatically. Under OpenAI's published age policy, children under 13 are not allowed on the platform.[1]
Nearly half of real chats sit outside the work filter
The AI Observatory, led by Anka Reuel of Stanford's Trustworthy AI Research lab and Shayne Longpre of the MIT Media Lab, analysed 24,521 conversations covering 85,633 turns from 5,000 users across 52 models between 2023 and 2025. Nearly 48 per cent of that traffic would be filtered out by the work-focused method Anthropic applies to its own data. Inside the filtered-out share, health and relationships held 44.2 per cent against 31.2 per cent in Anthropic's published data, harassment or hate 27.5 per cent against 5.66 per cent, sexual content 16.7 per cent against 2.4 per cent, and adult or illicit content 7.9 per cent against 2.1 per cent. OpenAI has separately reported that only 30 per cent of consumer ChatGPT use is work-related. The company datasets are far larger — about 1 million conversations at Anthropic and 1.5 million at OpenAI — and stay proprietary. Reuel said those company usage figures have no independent source that can corroborate them.[2]
Twenty-nine states take Meta's feed design to court
A trial brought by 29 US states, led by California, Kentucky, Colorado and New Jersey, opened in California on Tuesday. The states say Meta designed its platforms to encourage addictive behaviour, failed to verify users' ages and did not adequately guard against harmful content; the damages sought reach 1.4 trillion dollars and the case is due to run about seven weeks. The attorneys general are asking for parental verification, changes to what they call dopamine-manipulating algorithms, removal of image filters, a ban on multiple accounts and an end to disappearing messages. Mark Zuckerberg and Instagram's head are due to testify. The case follows a New Mexico judge's order in August 2026 for Meta to pay 567 million dollars, and sits alongside an EU investigation into child safety whose preliminary findings cite Digital Services Act breaches, a UK ban on social media for people under 16 announced for 2027, an Australian ban in force since December 2025, lawsuits in Brazil worth 525 million dollars and an approved case in Kenya seeking 2.4 billion dollars. Sonia Livingstone of the London School of Economics said a complete remake of Meta's business model or feed looks unlikely.[3]
