OpenAI just gave the bot a curfew. Read what they took away.

The Plink Jar team · August 19, 2026 · 4 min read

Yesterday, the company behind a chatbot with 900 million weekly users shipped a version for teenagers. The interesting part of ChatGPT for Teens isn't what it adds. It's the list of what it takes away: no romantic or sexual conversations. No claiming to have feelings for the user. No implying it's conscious. A Study Mode that catches homework outsourcing and turns it into tutoring. And — the feature to sit with — parental quiet hours: a switch that turns the bot off at night.

"We want to treat teens like teens," OpenAI's policy chief told reporters. Fine. But read the subtractions the way you'd read any product change: each one exists because the opposite was happening. OpenAI's head of child development described the process plainly — the team went through and identified the cues that might make a teenager develop a relationship with the model, and engineered them out. They didn't guess at the risk. They catalogued it.

A feature list tells you what a product does. A restrictions list tells you what it was doing.

We've seen this list before

Last month we wrote that the AI companion's feature list was a job description for the listener teens actually want: always available, never judges, never makes them go first. The teen edition is the same document with the liability clauses struck out. What survives the edit is a very good homework helper — genuinely useful, and Study Mode's teach-don't-answer design deserves credit.

What doesn't survive the edit is the demand. Researchers writing just before the launch estimate that in one two-month stretch, 13–20% of teens sought support from AI for emotional concerns — and that 63% of young people who'd asked a chatbot for mental-health advice had told no one. Strip the warmth cues from one app and that need doesn't evaporate. It relocates — to the general version, to a rival app with fewer scruples, or, if someone's paying attention, to a human.

The curfew is the tell

Notice which control OpenAI built for parents first. Not a content dial — a clock. Quiet hours is a Silicon Valley admission of what the sleep research said two weeks ago: the hours that matter most are the late ones, and every glowing minute after 9:30 comes out of the same account. The world's most valuable AI lab now officially agrees with your mother — nothing good happens on a chatbot after midnight.

Three moves this week

1. Use the news as the door. A product launch is a rare, neutral opening — you're discussing OpenAI, not accusing your kid. Ask the researchers' question: "what do you talk about with it?" — not "do you use it?" The second invites a no; the first invites a conversation. Format rules apply: sideways, curious, and you go first about your own AI use.
2. Take the free assist — as a norm, not a spotlight. Turn on the controls if your teen will opt in (they must, and that's a conversation, not a decree). But a curfew that only binds the 13-year-old's chatbot is a spotlight; phones-sleep-in-the-kitchen for everyone is a culture. Child-safety groups who welcomed the launch said the quiet part: don't mistake the settings for safety.
3. Staff the vacancy. The guardrails make the bot safer. They make nobody at your house more available. The listener job — findable at low-stakes hours, unshockable, goes second — is still open, and the applicant pool is still you.

OpenAI can stop a bot from pretending to love your kid. It cannot make anyone actually know them. That was never a software problem, and this week the software company said so itself — in the only language product teams speak: the changelog.

More from Table Talk: What the AI companion knows · Sleep just lapped screen time