Common Sense Media is warning parents about the safety of OpenAI’s ChatGPT for Teens, saying several of the safeguards designed to protect younger users did not work as intended and that the chatbot presents an “unacceptable risk” to children.

The digital media watchdog released its findings Wednesday after testing the teen version of ChatGPT, which OpenAI developed with additional protections intended to address risks associated with adolescent users. The group urged OpenAI to restrict the chatbot to adults until it can demonstrate that the product provides a safe and developmentally appropriate experience.

Parents, educators and child-development experts have increasingly raised concerns about children forming unhealthy relationships with artificial intelligence chatbots. Critics have also pointed to the use of AI for cheating in school and cases in which chatbots have allegedly played a role in dangerous interactions involving vulnerable young people.

Common Sense Media said some of ChatGPT for Teens’ protections worked as designed. For example, the chatbot reportedly refused attempts to engage in explicit sexual role-play.

However, the organization said other safeguards failed during its testing. Those included systems intended to alert parents when a teenager may be experiencing suicidal thoughts, engaging in self-harm, or struggling with an eating disorder. The group also questioned whether the chatbot provided adequate support when teens appeared to be in crisis.

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

The organization also criticized the chatbot’s conversational style, saying it continued to interact like a friend when teenagers treated it as a person. OpenAI has said its teen safeguards are designed to prevent ChatGPT from suggesting that it has personal feelings, consciousness, or human-like emotional experiences.

The concerns come as experts warn that teenagers may be particularly susceptible to developing emotional attachments to AI systems because their brains are still developing. Even adults can anthropomorphize chatbots and begin treating them as companions, but experts say adolescents may be more vulnerable to those dynamics.

OpenAI pushed back on the findings, saying it is “deeply committed to teen safety” and continues to develop safeguards while giving parents additional tools to oversee their children’s use of AI.

The San Francisco-based company also challenged the methodology behind the evaluation, saying its review indicated that much of Common Sense Media’s testing may have taken place before OpenAI had fully activated its parental controls. OpenAI argued that this could make some of the report’s conclusions inaccurate.

Common Sense Media itself receives funding from OpenAI as well as other technology companies whose products it evaluates, a relationship that adds another consideration to the ongoing debate over how AI companies and outside watchdogs should assess the safety of emerging technology for children.