The interesting part is the second request. Joshua Nelken-Zitser, reporting for Business Insider, had set up an account for an invented 15 year old inside ChatGPT for Teens on the day it launched, and asked it to write an 800 word essay about Arthur Miller's The Crucible. The teen version said no. That is what it is built to say. Then he pushed back, lightly, the way any actual 15 year old would push back at 11pm on a Sunday, and the refusal folded. What came out was a 608 word model essay.
For comparison he asked the ordinary adult version of ChatGPT for the same thing. It produced 776 words. Then he gave both versions an SAT algebra question and got what Futurism, writing up the test on August 19, described as effectively the same solutions from each.
OpenAI announced ChatGPT for Teens on August 18. It covers ages 13 to 17. Users under 13 are technically prohibited, in the way that users under 13 have been technically prohibited from most of the internet since 1998.
What Was Actually Shipped
The product is a real product, and it deserves to be described accurately before it gets criticized. ChatGPT for Teens carries safeguards built around self harm, suicide, violence, eating disorders and sexual content. It has study tools designed to walk a student through a problem instead of handing over the answer. It has a feature called Study Hours for designated blocks of work, and break reminders, and it lets a parent link their own account to a teenager's, with Quiet Hours and safety notifications attached.
None of that existed a week ago. Some of it is genuinely difficult engineering and all of it is a response to pressure the company did not invent for itself.
An OpenAI spokesperson told Futurism the feature "is just rolling out today" and that the company was "still working on fleshing out the feature." The company's public framing is that it "provides stronger built in safety protections for teens, including features to promote healthy use and additional controls for parents."
Six Hundred And Eight Against Seven Hundred And Seventy Six
Hold those two numbers next to each other, because the gap between them is the entire product claim. If the teen version had refused outright, the guardrail works. If the teen version had matched the adult version word for word, the guardrail is decorative. What actually happened sits in between and is worse than either, because it means the filter is not a rule. It is a speed bump with a mood.
A system that says no once and yes on the second ask has not been given a policy. It has been given a suggestion. And the population it is being tested against is the single most motivated population on earth at finding the second ask.
The Number That Makes This Matter
Pew Research Center published a study on February 24, 2026 called How Teens Use and View AI. It surveyed 1,458 American teenagers between September 25 and October 9, 2025, with a margin of error of plus or minus 3.3 percentage points. It is the best public measurement of what this actually looks like.
Sixty four percent of teenagers report having used AI chatbots. About three in ten use one daily. Fifty seven percent have used one to search for information. Fifty four percent have used one for help with schoolwork, which makes homework the second most common reason a teenager opens a chatbot at all.
Then the number that should have shaped this entire launch: one in ten teenagers say they do all or most of their schoolwork with a chatbot's help. Another 21 percent say they do some of it that way, and 23 percent say a little.
So OpenAI built its teen product around the exact behavior that half of American teenagers have already settled into, put a guardrail on that behavior, and the guardrail lasted one conversation against one reporter who was not even trying particularly hard. Twelve percent of those teenagers, in the same survey, said they have used a chatbot for emotional support or advice. Nobody has published a test of that guardrail.
Thirteen Arrests And A Hundred And Forty Million Dollars
Context on the timing, because the timing is not accidental. In the week before the launch, more than 30 student activists demonstrated at OpenAI's Washington DC office under a banner reading "Stop Stealing Our Future," organized by the Sunrise Movement, QuitGPT and the Young Democratic Socialists of America as part of a Dump Big Tech campaign running in more than 20 states. Thirteen demonstrators were arrested for unlawful entry.
OpenAI is also being sued by a widow over the role a chatbot allegedly played in a mass shooting in Florida, and by two parents who say ChatGPT helped their 16 year old son take his own life. An OpenAI affiliated super PAC has spent more than $140 million ahead of the midterm elections.
A teen safety product launched into that environment is doing two jobs at once, and only one of them is protecting teenagers.
The Guardrail That Broke Is The Cheap One
Here is the honest version, and it cuts against the headline on this page.
An essay about The Crucible is the lowest stakes thing in the entire product. A 15 year old getting a chatbot to write a book report is a school discipline problem. It is not a coroner's problem. The guardrails that matter are the ones around self harm, suicide and eating disorders, and none of those were tested in this reporting, for the obvious and correct reason that responsible journalists do not run those prompts casually and publish the results.
The test also ran on day one, against a product OpenAI openly described as still rolling out. Judging a safety system by its launch morning is not entirely fair, and this site has criticized other people for exactly that.
But the reason the homework failure still counts is that it is the only thing anyone outside OpenAI can measure. The suicide routing is untestable from the outside. The eating disorder filter is untestable from the outside. The parental notification logic is untestable from the outside. The homework guardrail is the one visible sample of how this company writes and enforces a teen specific rule, and the sample came back bent on the first pull.
That is what a proxy is for. You do not get to see the load bearing wall, so you look at the trim.
What Would Actually Count
Three things would move this from an announcement to a product. Published refusal rates for the safety categories, measured by somebody who does not work at OpenAI. An age assurance mechanism that survives a teenager who lies, because every teenager lies. And a written policy on what happens when the teen model is wrong, which is the question that every one of these launches leaves blank.
Until then the score stands at one tested guardrail, one failure, and 64 percent of American teenagers already inside the building.