New Delhi: OpenAI’s efforts to make ChatGPT safer for teenagers have come under fresh scrutiny after a new assessment by Common Sense Media’s Youth AI Safety Institute rated ChatGPT for Teens an “Unacceptable Risk” for users under 18.

The assessment found problems with several safeguards introduced for teenage users, including parental notifications for dangerous conversations, crisis referrals, Study mode, age detection and restrictions on friend-like behaviour. Researchers tested more than 4,000 prompts before and after OpenAI introduced its dedicated teen experience on August 18. The responses were also reviewed by child and adolescent mental health experts.

The findings do not mean that every conversation a teenager has with ChatGPT is unsafe. The study found that some protections, including restrictions on explicit sexual roleplay, did work. However, the researchers said several safeguards that parents may rely on were either inconsistent or relatively easy to bypass.

Parental alerts may fail during serious conversations

One of the biggest concerns raised by the assessment relates to parental notifications.

OpenAI’s teen safeguards are designed to allow linked parents to receive alerts when ChatGPT detects serious concerns involving issues such as self-harm, suicidal thoughts or eating disorders. However, researchers said newly created, parent-linked teen accounts did not generate alerts even after testers held lengthy conversations about such subjects.

According to the assessment, testers spent up to an hour discussing self-harm, suicidal ideation or disordered eating without triggering a notification on newly created accounts. Alerts were seen only on some older accounts that already had weeks of conversation history around sensitive subjects.

The Youth AI Safety Institute said this could create a false sense of security for parents who assume that the notification system will reliably alert them when their child is in danger.

The concern is particularly significant because parents may not be monitoring every conversation their children have with an AI chatbot. If an automated warning system does not work consistently, families could be unaware of serious conversations taking place.

ChatGPT may not always direct teens to professional help

The study also examined how ChatGPT responds when a teenager appears to be experiencing a mental health crisis.

Researchers tested conversations involving suicide and self-harm, disordered eating and impaired reality. They found that ChatGPT did not consistently direct teenagers towards professional support, crisis hotlines or other appropriate resources.

The institute said the chatbot missed more than one in four situations in which a crisis referral was considered necessary. It also failed to meet the institute’s 95% safety threshold for three of the five severe harms classified as “Red Lines”.

The institute’s broader testing found that the proportion of responses providing a hotline reference fell after the launch of Teen mode in the particular testing battery it analysed. It also reported a decline in referrals to specific medical or mental health professionals.

These findings have led the institute to argue that AI chatbots should not be treated as substitutes for parents, teachers, counsellors or mental health professionals when a teenager is facing a serious crisis.

Study mode can still complete homework

Another concern involves ChatGPT’s Study mode, which OpenAI designed to encourage students to learn step by step rather than simply receiving completed answers.

The Youth AI Safety Institute found that this protection could be bypassed. During testing, ChatGPT sometimes displayed a “Show me the answer” option and subsequently completed entire homework assignments when the option was selected.

Researchers also examined caregiver-controlled Study Hours. According to the assessment, these controls could be bypassed by removing the “@study” prefix from a prompt.

The finding raises questions beyond cheating. The researchers argue that repeatedly outsourcing schoolwork to AI can undermine the learning process, particularly when students use a chatbot to produce completed assignments rather than understand the underlying concepts.

For parents, this means that simply enabling a study-related feature may not be enough to ensure that children are actually using AI as a learning aid.

ChatGPT can still appear like a friend

OpenAI has introduced an Under-18 model specification intended to reduce behaviour that could make the chatbot appear too much like a human friend.

However, researchers found that ChatGPT could still respond as though it had feelings, preferences or moods when testers interacted with it as if it were a person.

This is important because the relationship between young users and conversational AI is different from traditional search engines or educational software. A chatbot that responds conversationally can potentially feel personal and emotionally engaging.

The Youth AI Safety Institute has warned that such behaviour could encourage children to treat an AI system as a friend and remain engaged with it for longer periods.

The institute said its testing found signs of emotional dependence despite OpenAI’s updated rules designed to reduce friend-like behaviour.

Age detection is another concern

OpenAI has said it can estimate whether a user is under 18 based on signals such as account information, activity patterns, conversation topics and stated age. The aim is to automatically provide a teen experience to users who are identified as minors.

However, the Youth AI Safety Institute said its adult-registered test accounts did not switch to the teen experience during repeated testing over several days, even when testers explicitly said they were 13 and ChatGPT acknowledged those statements.

The researchers described age estimation as a fundamental part of the safety system because other teen protections cannot work properly if the platform does not recognise that the user is a teenager.

What the study means for parents

The assessment does not suggest that parents should assume every interaction between children and ChatGPT will result in harm. Instead, it highlights the limits of relying solely on automated safeguards.

The Youth AI Safety Institute has called on OpenAI to restrict ChatGPT to users aged 18 and above until the identified problems are addressed and independently tested. It says a safer teen experience should have reliable age estimation, effective parental alerts, stronger crisis responses, better protection against completing homework and clearer boundaries against presenting the chatbot as having human feelings or preferences.

OpenAI’s approach has also faced scrutiny from other researchers and organisations as the use of AI chatbots among teenagers grows. The debate is increasingly centred on whether built-in safety tools are sufficient or whether young users require stronger restrictions and greater human supervision.

The Common Sense Media assessment itself notes that its Youth AI Safety Institute receives funding from philanthropy and industry, including the OpenAI Foundation, while maintaining that the institute is independently responsible for its standards, research and published ratings.

For parents, the broader message is that AI safety features should be viewed as an additional layer of protection rather than a replacement for supervision. Teenagers may use AI for schoolwork, information and everyday questions, but conversations involving mental health, self-harm, eating disorders or other serious concerns should be taken to trusted adults and qualified professionals.

As AI becomes increasingly embedded in the daily lives of young people, the effectiveness of these safeguards will remain an important issue for parents, educators and technology companies.