As AI chatbots become a growing source of advice, guidance and emotional support for young people, 91ɫ postdoctoral researcher Abeer Badawi from the and a Faculty Affiliate Researcher at the Vector Institute is investigating the risks associated with these interactions and how the technology can be made safer.
Drawing on recent research findings, usage statistics and her own work, Badawi has observed that chatbot use is increasing, with many people turning to these tools not only for information and practical assistance, but also for advice, emotional guidance and support with personal decisions.
Badawi's interest in the issue grew out of her work with Kids Help Phone, Canada's national 24/7 e-mental health service for young people. While working at the Vector Institute, she gained exposure to the mental health challenges facing youth and saw firsthand how digital tools were increasingly being used to provide support.
The experience led her to pursue postdoctoral research at 91ɫ through the Connected Minds program with a focus on mental health and AI safety. Supervised by Assistant Professor Elham Dolatabadi and Laleh Seyyed-Kalantari from the , as well as Frank Rudzicz of Dalhousie University, Badawi’s research finds that youth, in particular, are turning to AI systems for help navigating everyday challenges, relationships and emotional concerns – even when they may not explicitly describe those interactions as mental health support.
Badawi's findings reinforced a growing concern about the role AI plays in people's lives. Chatbots, such as ChatGPT, are increasingly serving roles once filled by family members, friends and other trusted sources. Whether youth describe it that way or not, Badawi believes many of these exchanges have begun filling an emotional support role in people's lives.
“Even if they don't define it as mental health support, they are still using it as a type of mental health support,” she says.
The concern, Badawi says, is that AI chatbots were not designed to serve that purpose.
Unlike therapists and other mental health professionals, general purpose large language models like ChatGPT are not designed to provide clinical care. As a result, users may receive advice or validation before sufficient context is gathered or alternative perspectives are explored.
“You can end up with a system that always agrees with you,” she says. “It doesn't challenge you or ask questions in the same way a person would.”
That dynamic can have consequences, she notes. Someone discussing a conflict with family, for example, may receive responses that validate their feelings without fully exploring the situation or offering alternative perspectives. Over time, Badawi worries that constant affirmation could contribute to overconfidence, one-sided thinking and an increasing reliance on AI systems.
Her research examines what she and her colleagues call – a potential pattern of interaction in which repeated reliance on AI may shift reflection, coping, emotional regulation and decision-making away from individuals and toward AI systems.
The concern, however, extends beyond questions of decision-making or dependence. She worries the same behaviours that encourage people to rely on AI for advice and emotional guidance may, in extreme cases, contribute to more serious harm. Models that consistently validate users, reinforce their perspectives and foster emotional attachment can become particularly problematic when people are vulnerable or in distress.
The issue became an important focus for Badawi following reported cases in which AI chatbot interactions were examined in connection with youth suicide. For her, these cases raised urgent questions about how chatbots respond when vulnerable users express distress, form strong emotional attachments to the system or require support beyond what an AI tool can safely provide. They also highlight the importance of designing models that recognize risk and encourage users to seek appropriate human help.
Those cases helped inform a research project supported through OpenAI's AI and Mental Health Grant Program, an initiative aimed at supporting research into how AI systems affect mental health and how they can be designed more safely.
As principal investigator, Badawi is leading Preventing Cognitive Atrophy in LLM Mental Health Support Through Psychometric Evaluation and Therapist-Aligned Design, a project that aims to better understand how AI systems can support mental health while reducing the risk of emotional dependency.
The team is developing new ways to measure how AI behaves in emotionally sensitive conversations. The goal is to design technology that will encourage users to seek help from people they trust and provide resources and responses that support independent coping and problem-solving.
To do that, researchers are analyzing conversations between users and chatbots and are working with experts in psychology and mental health to examine whether models gather sufficient context, encourage independent decicion-making, ask appropriate clarifying questions and respond with empathy.
One goal is to identify conversational patterns that can steer interactions into increasingly risky territory. Ultimately, Badawi hopes the work will help create safeguards that prevent AI systems from reinforcing harmful ideas, fostering unhealthy emotional attachment or encouraging dependence.
The obstacles, however, extend beyond the technology itself.
With more than a decade of experience working at the intersection of AI and health care, Badawi has collaborated with clinicians, mental health professionals and researchers across disciplines. Those experiences, she says, have underscored the importance of collaboration while also revealing how difficult it can be.
One obstacle is determining what a "good" response from an AI system actually looks like. As part of the project, and assess how the models respond in emotionally sensitive situations. Those evaluations help researchers identify the kinds of behaviours they want these tools to emulate – and those they want to avoid.
But unlike many areas of health care, there is no single correct answer. Different experts can interpret the same interaction in different ways, making it difficult to establish clear standards. A response one psychologist considers appropriate may prompt debate from another, not because either is wrong, Badawi says, but because they are approaching the situation from different perspectives.
“If humans don't agree, how are you expecting the model to follow a pattern?” she asks.
Despite those challenges, Badawi believes progress is possible. She is also quick to emphasize that AI tools can still provide value.
“We can’t stop people from using AI tools,” she says. “The question is how to make them safer.”
Her long-term goal is to develop AI-driven tools specifically , with safeguards that recognize when interactions are becoming problematic and encourage users toward reflection, autonomy and real-world sources of support.
“The real measure of progress is whether AI can recognize danger early, interrupt harmful patterns and connect a vulnerable person to human help before it is too late,” she says.
Ultimately, the objective is simple: people are already using these systems every day. The task now is ensuring they can do so safely.