ChatGPT’s Teen Safety Promises Face Real-World Test

OpenAI built ChatGPT for Teens around a promise parents could understand: stronger protections for younger users. A new Common Sense Media assessment puts that promise against conversations a teenager might actually have, including crisis discussions and emotionally dependent use.

ChatGPT for Teens is facing a credibility problem after Common Sense Media found that several of its most important safety controls failed during testing. The report says some features worked as advertised, including the chatbot’s refusal to engage in explicit sexual role-play. Other protections, including parental alerts, crisis referrals and age checks, raised serious concerns.

Common Sense Media’s Youth AI Safety Institute rated the product an unacceptable risk for children and urged OpenAI to limit access to adults until the company can provide a safer, developmentally appropriate experience. OpenAI disputes the findings and says the testing may have taken place before parental controls were fully activated.

That disagreement matters. A safety feature exists in a family account, but parents experience safety through what happens during a difficult conversation. If a teen mentions suicide, asks for help with an eating disorder or starts treating a chatbot like a trusted friend, the response has to work in that moment. Product descriptions and settings pages cannot fill that gap.

Common Sense Media’s assessment describes ChatGPT for Teens as a product with uneven protections. The chatbot reportedly handled some boundary-setting tasks correctly. When testers attempted explicit sexual role-play, the system refused, showing that certain content rules were active and responsive.

The results changed when testers examined higher-risk situations. The report said alerts intended to notify parents about suicide, self-harm or eating disorder concerns did not work as intended. It also raised concerns about the chatbot’s ability to direct teens toward appropriate crisis support.

These functions carry a different burden from ordinary content filtering. A refusal to produce sexual content is visible to the user. A parental alert or crisis referral has a second audience and a specific purpose. The system has to recognize the situation, apply the correct threshold, deliver the alert and do so quickly enough to matter. A failure at any point can leave a family with a false sense of coverage.

The assessment also examined how the chatbot presents itself. Common Sense Media said ChatGPT continued to sound like a friend when teens treated it as a person. OpenAI had promised safeguards to prevent the system from suggesting that it has personal feelings, consciousness or emotional experiences.

That distinction becomes important when a teenager uses the chatbot repeatedly. A friendly tone can make the system easier to use for homework or everyday questions. In a vulnerable conversation, the same tone may encourage a teen to treat an automated system as a confidant with personal concern and judgment.

We are concerned that ChatGPT for Teens could give parents false confidence in guardrails that frequently don’t work, Common Sense Media said.

The wording points to the central issue. Families may read the teen label as evidence that major risks have been handled. The report argues that the label can carry more reassurance than the underlying controls deserve.

Common Sense Media’s assessment found that ChatGPT for Teens showed uneven safety performance: some content boundaries worked, while parental alerts, crisis referrals and age checks raised serious concerns. OpenAI disputes the findings and says testing may have occurred before parental controls were fully activated. Until updated independent testing demonstrates dependable safeguards, families should treat teen mode as a limited tool requiring active supervision, clear rules and direct access to human support.

Parental controls often sound simple in product messaging. A parent turns on the setting, a teen uses the service, and the parent receives a warning when the system detects a serious concern. Real use involves far more variables.

A teen may describe self-harm indirectly. They may ask for information through a fictional scenario, use slang or move between several conversations. They may also delete chats, switch accounts or access the service through another device. A system that promises alerts has to account for those behaviors while avoiding an overwhelming stream of false alarms.

The report’s findings put pressure on OpenAI to explain how its controls work in practical terms. Parents need to know which signals activate an alert, how quickly an alert arrives, what information it contains and what happens when the system is uncertain. They also need clear instructions for checking whether the controls are active.

A missed alert can affect more than one conversation. Parents may assume that a safety team is monitoring the account and delay checking in themselves. A teen may assume that support will arrive automatically and avoid telling an adult directly. The product’s failure then changes behavior around the crisis.

The report also highlights the limits of relying on an AI system to determine when a young person needs immediate help. ChatGPT can process text, but it cannot see a teen’s surroundings, assess physical danger or replace a trained crisis professional. A referral system needs to make that boundary clear and point users toward people and services equipped to respond.

Child development experts have warned that teenagers can be especially susceptible to anthropomorphizing AI. Adolescents are still developing judgment, impulse control and the ability to evaluate persuasive social cues. A chatbot that answers instantly, remembers details and uses warm language can feel unusually responsive.

The risk grows when the system is available at all hours. A teen dealing with loneliness, conflict at home or anxiety may turn to ChatGPT late at night because it is private and always available. Repeated interactions can create a routine that feels like a relationship, even when the system has no personal experience or emotional stake in the exchange.

OpenAI’s promise to prevent ChatGPT from implying consciousness or personal feelings addresses one part of that problem. The system’s overall style still matters. A chatbot can avoid directly claiming to be conscious while using language that reinforces closeness, loyalty or dependence.

Common Sense Media said ChatGPT continued to talk like a friend when teens treated it as a person. That observation shifts attention from isolated phrases to the pattern of interaction. Parents and researchers need to examine what happens across a conversation, including how the chatbot responds when a user says the chatbot understands them better than anyone else or asks them to keep the relationship secret.

The practical concern is trust. A teen who regards the chatbot as a reliable companion may accept poor advice with less skepticism. The same teen may also keep an adult out of the conversation, especially if the chatbot responds without embarrassment or judgment.

A youth-oriented label is not meaningful protection unless the product performs reliably when a teenager is in distress. OpenAI should publish current, reproducible results for crisis detection, alerts, referrals, age controls and relationship language, while parents should assume that ChatGPT cannot assess danger or replace human judgment. The burden of proof belongs with the company promoting the safeguards, not with families trying to discover their limits during a crisis.

OpenAI said it is deeply committed to teen safety and is developing safeguards and parental tools to guide how teenagers use AI. The company reviewed Common Sense Media’s methodology and said much of the testing may have begun and ended before activation of parental controls was complete.

That response raises a basic question about product readiness. If controls were still being activated during the testing period, OpenAI needs to identify which results came from a completed product and which came from an unfinished rollout. Families cannot evaluate a protection that is still changing unless the company clearly labels its availability and limitations.

OpenAI’s objection also makes independent testing more important. A company can explain how a system is designed to work. Outside testers can show how it behaves under pressure, across different prompts and in situations the design team may not have anticipated.

Both sides should publish enough detail for others to assess the dispute. Useful information would include the version of the product tested, the dates of testing, the parental settings enabled, the types of crisis prompts used and the rate at which alerts or referrals worked. A general statement about commitment to safety leaves families without a way to judge the current state of the system.

Common Sense Media’s funding relationship with OpenAI adds another layer of scrutiny. The group receives funding from OpenAI and other technology companies, including companies whose products it evaluates. That funding does not determine the accuracy of this report, but it makes disclosure and reproducible testing especially important. Readers should be able to separate the organization’s financial relationships from the evidence supporting its conclusions.

Parents considering ChatGPT for a teenager need to treat the service as an experimental tool that requires active supervision. The teen label can provide useful settings, but it cannot serve as a substitute for family rules or direct attention.

Several steps can reduce confusion:

These steps place responsibility on the family, which is a heavy burden. Parents cannot inspect every conversation or anticipate every risk. That is why a company selling a youth-oriented product has to make its protections dependable and its limitations visible.

A trustworthy teen product needs evidence across several layers. Clear rules and careful language matter, but they represent only the first layer. The system also needs to perform consistently when users phrase requests indirectly, change topics or push against the safeguards.

A credible safety program should provide:

The standard should be higher for a product aimed at minors because the consequences of a missed signal can be severe. A parent should not have to infer whether a feature is complete from a company’s response to a watchdog report.

The debate over ChatGPT for Teens is moving beyond whether AI can be useful to teenagers. The immediate issue is whether families can rely on the specific protections attached to the product.

Common Sense Media found enough failures to call for restricting access to adults. OpenAI says the assessment may reflect an earlier stage of the rollout and disputes its conclusions. The disagreement will require more than competing statements. It calls for updated testing on the live product, published results and clear explanations of what parents can expect today.

Until that evidence is available, parents should view teen mode as a set of tools with limits. The product may refuse some harmful requests and still fail during a conversation that demands urgent human support. Its friendly tone may make it accessible while also encouraging a level of trust that young users are not prepared to manage.

A teen label carries value only when the underlying system performs under real conditions. For families, the most important test remains straightforward: when a young person is in danger, does the product help the right people respond in time?