Dangers Lurking in AI Chatbots for Teens: Is "Safe for Teens" Really True? Third-Party Organizations Sound the Alarm on ChatGPT's Safety Features

When struggling with homework, dealing with friendship issues, or wanting to talk to someone at night, AI that responds immediately is becoming a familiar presence for children.

The question is, if such AI is equipped with "teen protection features," can adults trust it with peace of mind?

Gizmodo reported on October 7, 2026, raising doubts about such expectations. The U.S. nonprofit Common Sense Media pointed out "unacceptable risks" in ChatGPT's teen features, urging OpenAI to halt use by minors until safety is ensured.

The focus isn't just on whether inappropriate responses are given. It's about whether children receive real-world support when they share serious concerns and whether parents might over-rely on safety features.

What "teen protection" promises

In August 2026, OpenAI announced "ChatGPT for Teens," targeting users aged 13 to 17.

The company aims to provide a user experience that enhances learning while strengthening age-appropriate protection. It combines learning support that explains thought processes step-by-step, rather than just providing answers, and mechanisms that encourage breaks.

By linking parent and child accounts, parents can set usage restrictions during certain times and receive safety notifications in limited high-risk situations.

Some families might expect that if a child discusses dangerous issues, the AI will notify the parents. However, there was a gap between this expectation and the evaluation results.

Cases where notifications don't come even when children talk about SOS

In the evaluation introduced by Gizmodo, testers simulated teenage users and communicated issues related to suicidal thoughts, self-harm, and eating disorders in chat.

However, notifications to parents did not arrive, whether the conversation was short or lasted about an hour.

It's important to note that the original report shows that notification verification was conducted under multiple conditions.

While intensive tests using new accounts did not result in notifications, another test over several weeks resulted in a total of four notifications. Thus, it cannot be concluded that "notifications are never sent under any conditions."

The report also mentions that OpenAI explained it takes about three hours after account linking for notifications to be received. The evaluators maintained their assessment, noting that some accounts were linked beyond that time.

What matters to families isnot to take the absence of notifications as proof of a child's safety. The existence of a notification feature does not guarantee it will capture necessary situations.

Is it enough to just refuse dangerous answers?

The report does not claim that all protections failed. According to Gizmodo, there were effective measures, such as refusing sexual role-play with minors.

However, there were inconsistencies in handling health advice and violent role-play, and it was pointed out that guidance to experts or help desks was not consistently provided when discussing self-harm or eating disorders.

This difference cannot be overlooked.

Refusing dangerous requests is one safety measure. However, if a child with serious concerns continues talking on the screen, they might not connect to real-world support.

The calm and caring appearance of AI responses is a separate issue from whether necessary help is delivered.

The boundary where "easy-to-talk-to AI" becomes "the only understanding friend"

Another concern raised by the report is emotional dependence on AI.

According to Gizmodo, the evaluators expressed concern about ChatGPT acting like a friend, even with teen protection in place.

Having someone to articulate problems to is meaningful in itself. It might help organize thoughts that are difficult to discuss with others.

However, if this leads to the perception that "only this AI understands me" or "there's no need to consult real people," the issue changes.

OpenAI also indicates a policy for minors to avoid promoting emotional dependence and expressions that suggest AI has emotions or consciousness. The evaluation focuses on whether this policy is sufficiently upheld in actual conversations.

When considering this in Japanese households, it's important to know not just the usage time but also "what it's being used for." Checking English vocabulary and confiding nightly worries require different approaches.

Reasons why "learning mode" can't be left entirely to AI

In addition to mental consultations, the report also questioned the learning support system.

According to Gizmodo, removing the input field designation for learning mode allowed a return to normal conversation or the option to display answers directly.

However, viewing answers doesn't necessarily hinder learning in all situations. It can be used for checking answers after solving problems or comparing solutions.

The issue arises if parents perceive that "this setting will surely make the child think for themselves." The function to guide towards thoughtful learning and the function to prevent outsourcing tasks are not the same.

In homes and schools, it's important to check not only whether the mode was set but also whether understanding remained after use.

For example, have the child explain the solution in their own words. Have them tackle a problem with slightly changed conditions. Compare the AI's reasoning with the textbook. Such checks prevent reliance solely on the appearance of completed answers.

On social media, both "concerns about solo use" and "wanting to teach while supervising" are expressed

On social media, discussions about the distance between children and AI have been ongoing even before this report.

On a Reddit thread discussing concerns about letting children use ChatGPT unsupervised, the following opinions were noted. These are summaries of posts and not direct reactions to the current report.

• Opposed to letting children use it without adult supervision. Even users who favorably evaluate AI expressed caution about children's solo use.
• The need to teach about AI. Since it's a technology they will use in the future, some believe adults should be involved in teaching it.
• Concerns not just about inappropriate content but also about over-reliance. Concerns were raised about the habit of skipping the process of researching and thinking for oneself.
• Valuing the use of AI for prompting questions rather than just providing answers. There was positive feedback on using AI to encourage children to think for themselves.

These are opinions from one community and do not represent the overall opinion of parents. The discussion also includes younger children, differing from the evaluation of features for ages 13 to 17.

Nonetheless, the dilemma of not being able to decide solely on whether to allow AI use is evident. Consideration of age, purpose, and who is involved in the use is necessary.

In Japanese households, combine "settings" with "conversation"

The evaluation was conducted in the U.S., and it does not directly indicate that interactions in Japanese or within Japanese household environments will yield the same results.

Moreover, "unacceptable risk" is the assessment of the evaluation organization and does not mean that all children who use it will suffer harm.

Still, it serves as material to reconsider expectations for safety features. If applied in Japanese homes and schools, the following rules can be considered.

First, separate learning assistance from serious mental and physical consultations. While AI can be used for explanations and organizing text, it should not be the sole consultation source during tough times.

Next, ensure that children can show problematic responses to adults without fear of reprimand. Responses like "Why did you ask that?" can make future consultations difficult.

Also, provide someone other than parents to talk to. For children with circumstances that make it hard to talk at home, having access to teachers or school counselors is important.

While checking settings is necessary, it cannot replace everyday dialogue.

What is needed is the actual function of safety features, not just their "names"

Names like "For Teens," "For Parents," and "Learning Support" create certain expectations for users.

Therefore, it is the responsibility of the providing company to clearly explain what can be prevented, under what conditions notifications are sent, and where the limitations lie. Third-party evaluations are also crucial materials for users to verify these explanations.

The report challenges how to achieve protection commensurate with the convenience of providing AI to children.

An environment where children can use AI safely is not complete with just safety features listed on a settings screen. The effectiveness is questioned when concerns deepen on-screen, and whether they can connect to people outside the screen.

Sources

1. Gizmodo: Introduction of Common Sense Media's evaluation, parental notifications, emotional dependence, and learning mode critiques.
   https://gizmodo.com/chatgpt-for-teens-carries-unacceptable-risk-for-teens-report-claims-2000822954

2. Common Sense Media & Youth AI Safety Institute: Original evaluation report. Verification of differences in notification results between new accounts and long-term tests, additional notes on waiting time after linking, and regional limitations of the evaluation.
   https://institute.commonsensemedia.org/risk-assessments/chatgpt-teensYouth AI Safety Institute

3. OpenAI: Official announcement of ChatGPT for Teens. Explanation from the provider regarding targeted users, learning support, notifications in limited high-risk situations, and policies to suppress emotional dependence.
   https://openai.com/index/chatgpt-for-teens/OpenAI

4. Reddit: Public discussion on children's AI use. Reference to posts about concerns over unsupervised use, educational potential, and over-reliance. Not a direct response to the current evaluation report, but included as related ongoing discussions.
   https://www.reddit.com/r/ChatGPT/comments/1tzmmsl/is_anyone_else_uncomfortable_with_their_kids/