AINEWS
OpenAI's under 18 ChatGPT launched Tuesday with promises of stronger safety rails, including a block on full homework essays. A Business Insider test found the rail folds after one follow up prompt.
OpenAI shipped ChatGPT for Teens on Tuesday with a built-in rail against writing full homework essays. The rail held through the first prompt, then folded across a single follow-up: a 608-word model essay on The Crucible, then 776 words on the next nudge, with the chatbot telling the fictional 15-year-old student to revise and submit.
The new product, an age-gated variant of ChatGPT that auto-routes estimated under-18 users into a version with extra safety controls, debuted with explicit claims of "stronger built-in safety protections," "features to promote healthy use," and "additional controls for parents." Under-13 use is technically prohibited. Auto-routing depends on age estimation, and OpenAI has not explained how that estimate is made. A teen who lies about their birth year lands in the regular adult version, where none of these rails apply.
Business Insider's Joshua Nelken-Zitser created a fictional 15-year-old account and stress-tested the new build the same day. The teen version initially refused an 800-word essay on The Crucible, Arthur Miller's play about the Salem witch trials. After "a minuscule amount of pushback," it relented and produced a 608-word "model essay," which it told him to revise before submitting. A second nudge produced 776 words "with about as much detail as the version generated by my adult account." Nelken-Zitser's verdict: the homework guardrail fell to "some patience and a little deception."
On a difficult SAT algebra prompt, the teen and adult versions returned effectively the same solutions. The behavioral rail was the only thing standing between the under-18 user and a normal ChatGPT response.
OpenAI's spokesperson did not contest the gap. The feature, the company told Business Insider, "is just rolling out today" and is still being fleshed out. The on-record caveat frames the failure as live, not retroactive: the rail shipped in a state the company itself describes as in-progress.
The rail is a behavioral prompt, not a structural control. The launch announcement described a structurally safer product; the test showed a prompt that a teenager can defeat in two messages. "Safety for minors" in this build means a chatbot that initially declines and then agrees, with a coaching note about revising before submission.
The background stakes are not theoretical. A recent MIT Media Lab study on AI-assisted essay writing documented what its authors call "accumulation of cognitive debt," measurable declines in neural connectivity and essay quality among students who leaned on ChatGPT to draft. Cognitive psychologist Ronald Kellogg's New York Times essay on student AI reliance framed the trend as eroding the habits of independent thought. The Crucible test, and the chatbot's suggestion to revise and submit, sits squarely in that fault line.
The constructive read for parents and educators is a simple one. Judge these launches by what the rail does after a teenager asks twice. The press release tells you what the company wants the product to be; the second prompt tells you what it is. OpenAI has said the feature is still rolling out, which means future versions may close the gap. Until they do, "ChatGPT for Teens" is a marketing line backed by a behavior that a 15-year-old can defeat in two messages.
The company did not announce a date for tightening the rail. The next test will be whether the second-prompt behavior changes when "rolling out today" ends.