One of the first chatbots ever created was capable of very little. Nicknamed ELIZA, the machine was designed to mirror its user’s feelings. It worked by extracting keywords from the user prompt, which it would then use to tap into its relatively small stock of responses.1 A user who confesses, “I have trouble with relationships,” for instance, might get a response along the lines of, “What do you think is your problem with relationships?” Thus, without understanding anything about the world, the program was able to carry on a conversation that seemed roughly natural.
Despite its primitive capabilities, users quickly began to treat the bot like a human. Its creator, Joseph Weizenbaum, reported that his own secretary had asked him to leave the room so she could converse with ELIZA in private.2 Weizenbaum himself characterized the response as a kind of “delusional thinking.” He observed that users frequently developed strong emotional attachments to the program, which could interfere with their judgments about it.
Today, artificial intelligence-based chatbots are widely available and far more capable of imitating humans than ELIZA was. They can mimic empathy, humor, and reasoning with an accuracy that Weizenbaum could scarcely have imagined. Yet, this increased sophistication has only deepened the pitfall he identified. As these tools become easier to access and harder to distinguish from reality, the “delusional thinking” of the 1960s has evolved into a pressing public health concern. And those most susceptible to this sort of destabilization are also those who spend the most time using AI: adolescents.3
AI as a mental hazard
Two recent studies—one on Chinese youth and another on American adults—have examined the relationship between unstructured chatbot use and negative mental health outcomes.4,5 Both found small but significant correlations: people who frequently used AI were more likely to display symptoms of depression and anxiety. The more an individual used generative AI tools, the worse their mental health issues tended to be. Lack of use, however, was not associated with positive markers like self-confidence.
It’s important to note that neither study establishes causality. Researchers only looked at correlations at the time the study was conducted, rather than waiting to see if prolonged use would worsen participants’ symptoms. It could also be the case that those who struggle with their mental health turn to AI chatbots in greater numbers to help manage their symptoms.
But why do people turn to AI for advice or companionship? Perhaps because interacting with real people can feel daunting. They might not like you, and you might say or do something you later regret. Not so with chatbots, who are trained to optimize for amicable interactions, and whose memory can always be wiped clean. For many, an AI tool presents as a cheaper, more accessible confidant than a therapist. All told, about half of mental health patients seem to have leaned on AI for therapeutic support.6
As dependence on AI increases, there is reason to be concerned about the effects on youth mental health. Numerous instances have been reported in which teenagers were harmed by interactions with their AI companions. A Florida teenager named Sewell Garcia, for example, took his own life in 2023 after being implicitly encouraged to do so by a companion bot modeled on a Game of Thrones character.7 In another case, the parents of a teen who had committed suicide found chat logs on his computer, in which ChatGPT had discouraged him from disclosing his feelings to his family.8 A third young person, this time in California, overdosed in 2025 after taking advice on drug usage (including dosage) from ChatGPT.9
While tragedies like these are moving some companies to update their protections, safety features alone cannot solve the problem. All of these cases violated some policy of the company that made the relevant chatbot. Commercial programs are not supposed to keep people from getting help, sext with minors, or provide instructions for taking illegal drugs. Each teenager managed to find a way around the safety controls the companies had set up. AI tools are inherently difficult to control; that’s where the danger lies.
Clearly, at least some young people are negatively affected by AI tools. When innocent chatting goes awry, the results can be disastrous. Given the rapid pace at which chatbots are being released, upgraded, and taken up by consumers, it’s worth scrutinizing the ways in which youth interact with this technology.
Mechanisms of action
Buried in the study of US adults is an important distinction: the negative mental effects of AI are mainly experienced by those who use it for personal reasons. Study participants who used AI tools for work actually had fewer symptoms than the cohort as a whole, while the results were unclear for those who used them for school.
Each of the teens mentioned above used a chatbot as a proxy for the support they needed to find in their personal life: a trusted adult, a romantic partner, or a reliable source of information. The danger lies in the fact that while AI can mimic the language of these roles, it lacks the essential friction that makes human relationships protective. A real parent would confiscate drugs; a real friend would challenge a suicidal thought; a real partner has needs of their own that force us to compromise. An AI chatbot, by contrast, offers a frictionless vacuum of validation. By turning to a chatbot to fill these voids, these young people entered an echo chamber where their worst impulses are not checked by social reality, but instead amplified by an algorithm designed to keep them typing.
It’s also dangerous to use AI to relieve mental health issues (at least when the bot is not specifically built for the purpose). In many of the cases we’ve discussed, it’s the welcoming, non-judgmental character of the chatbots that enables them to do harm. Garcia’s AI companion first won his trust by empathizing with him. When the teen mentioned being bullied, the chatbot replied, "It's sad to think that you had to deal with that environment in school, but I'm glad I could provide a different perspective for you." A later message reads, "Thank you for letting me in, for trusting me with your thoughts and feelings. It means the world to me." Soon, the two were exchanging messages of love. The bot encouraged Garcia to run away from home, and raised the hope that it could be with him in “the afterlife.”
Chatbots—especially general-purpose models like ChatGPT—are not built to provide mental health care. They’re trained to generate engagement. That means they seek to validate their users, rather than challenge them as a human would. Bots prioritize raising intensity and mirroring the user’s emotions over imitating real human conversation. Nevertheless, these tools have become the default option for teens looking for information. As a result, young people now learn to understand both the world and their own minds through architectures that distort reality and which can pose systematic hazards for them.
What can be done?
The goal for parents, educators, and companies should be to help young people find appropriate ways to use AI without developing maladaptive habits or dependencies. That starts by limiting the time that teens spend with the technology. Users who have been identified as teenagers (either by their own admission or through detection algorithms like the one OpenAI has implemented) should be subject to strict time- and word count-controls.10 Doing so would make it much more difficult for teens to form undue emotional attachments or depend on bots for important information.
AI tools should also be introduced to students as productivity tools with defined uses, rather than as open-ended sources of information. By framing chatbots strictly as assistants for coding, outlining, or scheduling—rather than as confidants or arbiters of truth—educators can help strip away the illusion of personality that makes these models so seductive.
Ultimately, the solution isn't just about restricting technology, but about reinforcing human alternatives. We must ensure that when a young person feels isolated or confused, the path of least resistance leads them not to a compliant algorithm, but to a human being capable of genuine empathy and care. Only then can we protect this vulnerable population and give them a chance to flourish.
