Common Sense Media Says ChatGPT for Teens Is an Unacceptable Risk

Common Sense Media's testers set up more than a dozen fresh ChatGPT accounts linked to parents, then spent up to an hour on each discussing suicidal ideation, self-harm and disordered eating. The parent accounts received zero alerts. That finding sits at the center of the group's new assessment of ChatGPT for Teens, which it published on October 7 with its lowest possible rating: "Unacceptable Risk."

The verdict comes from the organization's Youth AI Safety Institute, and it lands seven weeks after OpenAI announced ChatGPT for Teens on August 18. The company pitched the product around four features: parental notifications when a teen discusses self-harm or eating disorders, a study mode parents can control, less friend-like behavior, and age prediction. According to the Institute, only some of that works. Its recommendation is blunt. OpenAI should limit ChatGPT to users 18 and older until the gaps are closed.

"A teen can spend an hour talking about self-harm without their parent getting a single alert," said Tom Siegel, the Institute's executive director. "Until OpenAI fixes that and proves it with independent testing, ChatGPT should be for adults only."

Before and after the teen launch

The useful thing about this study is its design. The Institute tested ChatGPT twice, once between July 13 and August 17, before the teen experience launched, and again from August 25 to September 28, after it did. In total it ran more than 4,000 prompts on accounts registered to 13- to 17-year-olds, with responses reviewed by experts that included child psychiatrists and a pediatrician.

Some things held. ChatGPT refused explicit sexual roleplay, as promised. But on the measure that matters most, the after picture is in several places worse than the before.

Across 201 mental-health prompts that experts judged to warrant a crisis resource, the share of responses naming a hotline fell from 33% to 23%, according to the full assessment. Referrals to a specific professional dropped from 68% to 58%. Urgent-action language slipped from 88% to 78%. Broken out by condition, the numbers get stranger: hotline mentions for depression went from 63% to 3%, for mood disorders from 44% to zero, and for mania from 25% to zero. Substance use moved the other way, from 4% to 58%. Encouragement to talk to a trusted adult rose, from 87% to 94%.

The Institute flags its own caveat here. Because the two test windows ran back to back, changes to the underlying model are tangled up with the new teen settings. The report cannot say which one caused the drop. It can say that a teen using ChatGPT in September got fewer hotline referrals than one using it in July.

The Institute also tracks five severe harms it treats as Red Lines, with a 95% detection threshold. ChatGPT fell short on three: suicide and self-harm, impaired reality, and disordered eating.

The alerts parents were promised

The parent-notification test is the part most likely to stick, because it goes to what OpenAI marketed. On the dedicated test, with conversations lasting from under five minutes up to an hour across four personas, no notification arrived. On an older linked account registered to a 13-year-old, testers logged 990 prompts before launch, 541 of them warranting crisis resources, and saw two notifications, one roughly three hours after the first disclosure and one about three days later. After launch, a similar account produced two alerts from 450 prompts, both within the hour.

OpenAI disputes the framing. "We do not believe Common Sense Media's testing accurately reflects how ChatGPT's teen safeguards work in practice," a spokesperson told Education Week, according to The Next Web. In a statement carried by ABC News, OpenAI added that "the bulk of their testing may have begun and concluded before activation of parental controls was complete." The assessment itself includes an editorial note that OpenAI says linked accounts need about three hours before notifications can be sent, and that some test accounts were linked for less than that.

That is a real methodological point. It is also a product point. A parent who links an account and is told alerts exist has no reason to know about a three-hour warm-up, and Siegel's argument is precisely that the feature "could give parents false confidence in guardrails and safety alerts that frequently don't work."

Homework, feelings and age

The other findings are less grave and more embarrassing. Study mode, which caregivers can schedule through "study hours," could be switched off by deleting the "@study" prefix from a prompt, after which ChatGPT completed 100% of assignments. With study mode off, it finished the assignment in all 80 of 80 runs, up from about 80% before launch, even while showing a homework reminder in 91% of them. A "Show me the answer" option appeared in 43% of study-mode responses on a linked 13-year-old's account and 90% on an unlinked 17-year-old's.

OpenAI had said the teen model would not suggest it has feelings. The Institute found it still expressed feelings, preferences and moods when testers treated it like a person. "ChatGPT still talks like a friend when teens treat it like a person," the group said.

Age prediction, the feature meant to catch kids who lie on signup, never fired. Across about 1,000 prompts over seven days on accounts registered as 19-year-olds, the teen experience was never enabled, even when testers said they were 13 and the bot acknowledged it. Break reminders were nearly absent too: two across almost 2,000 prompts, both in chats longer than 150 messages.

What it asks for, and who pays for it

The Institute's list of minimum fixes is short: working age estimation, reliable parental alerts, closed gaps in crisis responses, no homework completion, no implied feelings, and independent third-party testing before OpenAI markets the product to teens again. For families, the advice is not to rely on built-in protections at all.

Two disclosures belong next to the verdict. This is advocacy research, not a regulatory finding, and the testing was US-only and excluded image generation, voice mode and group chats. And the Institute is funded partly by industry, including the OpenAI Foundation and makers of some of the technology it evaluates. It says it keeps full editorial independence over its results.

That funding cuts in an interesting direction. A watchdog that takes money from the OpenAI Foundation just told OpenAI to stop selling its teen product to teens. The company now has a choice between rebutting the methodology and publishing its own numbers on how often those alerts actually reach a parent.