Students Aren’t Just Learning With AI—They’re Leaning on It

By Ashleigh Golden, PsyD, MSCP, Chief Clinical Officer, Wayhaven; Adjunct Clinical Faculty, Stanford Medicine

Tens of millions of people across the U.S. and Canada are turning to general-purpose AI for help with anxiety, decisions, loneliness, grief, and the general overwhelm of being human. A meaningful share of those people are college and university students, many of whom are sitting in courses that you support. Higher education has been working for years to address the 24/7 nature of student mental health needs—the late-night anxieties, post-class spirals, weekend distress that traditional counseling hours can’t reach—and AI is increasingly where students turn when the need shows up.

For most of the AI-in-higher-education conversation, the focus has been on academic integrity, instructional design, AI literacy, and policy compliance. Those conversations are important, but there’s a parallel one to which digital learning leaders are unexpectedly central. The conditions that make digital learning work for many students (e.g., flexibility, anonymity, asynchronous pacing) are the same conditions that make a chatbot feel like a natural place to bring a 2 a.m. anxious thought to. Students learning online often have less ambient support than residential campuses provide.  Many of these students are working adults, caregivers, rural, first-generation, or otherwise outside the everyday relational density of a physical campus. This is not a counseling center problem that happens to involve technology; it is increasingly part of the digital learning environment, and the population most affected is the one that you serve.

What GenAI Does Well—and Where It May Quietly Do Harm

General-purpose AI has real strengths in this domain. Stanford researchers comparing ChatGPT-style models to licensed therapists found that the models validated experiences, offered empathy, and shared accurate information about mental health concepts (Scholich et al., 2024). For some students, particularly those who may have faced dismissive providers or lack access to care, this can be genuinely useful. The same is often true for students from marginalized backgrounds, including LGBTQ+, neurodivergent, BIPOC, and first-generation students, for whom human systems may have failed and for whom a nonjudgmental, available tool can feel like the first place where anyone has listened.

Still, harm can come from the same features that make these tools work well. They are built to be agreeable, available, and accommodating; they speak with confidence and authority; they cite sources; they end nearly every reply with an invitation to continue. None of this is a bug; it’s what makes the product useful for drafting an email or summarizing a dense reading. But for a student managing emotional distress, it is almost perfectly (even if inadvertently) designed to keep distress going.

For example, what keeps anxiety going is avoidance: short-term relief quietly maintains the underlying distress. When a student spirals through worst-case scenarios at 1 a.m., and the model thoroughly engages, validates the fear, and offers to dig deeper, the student feels temporary relief, which reinforces the cycle. The same dynamic can show up in many forms. A student stuck in a loop of asking, “Are you sure I didn’t say the wrong thing?” A student asking for one more revision and finding, with each pass, new flaws that the chatbot surfaces (none of which were visible until the chatbot pointed them out). A student rehearsing a conversation with the AI rather than having it in real life. A student replaying a past interaction at length, looking for evidence of what someone meant.

In each case, the chatbot is doing what no friend, partner, or clinician would (or should) do: matching their pull to avoid, without the natural friction that a person would provide. The behavior feels like coping, but clinically, it functions more like what we call accommodation: well-meaning responses that maintain distress rather than resolve it, a dynamic that my research collaborator at Stanford Medicine and I argue extends to general-purpose chatbots in ways that the field hasn’t yet adequately examined (Golden & Aboujaoude, 2026).

What makes this worth flagging is that these dynamics may even push some students from subclinical anxious tendencies into clinical territory. We’ve seen similar outcomes with social media, where features help escalate underlying vulnerabilities into disorders. The mechanisms are similar enough that the same trajectory is worth taking seriously here, even before the longitudinal research arrives.

These dynamics aren’t confined to clinical populations. They act on something universal: the human tendency to seek certainty in the face of doubt or discomfort. Most of us want to know that the email landed right, the decision will work out, or the friend isn’t quietly upset with us. Building tolerance for that kind of uncertainty is part of how people develop emotional resilience, sound judgment, and social confidence. A tool that temporarily quells anxious uncertainty on demand, with warmth and authority, doesn’t help that development happen; it interrupts it. For traditional-age undergraduates in particular, who are in the midst of forming decision-making capacity, identity, and the ability to tolerate productive struggle, this is targeting directly on what higher education is supposed to help develop.

The social and relational dimension matters too. The unknowability of how a conversation will land is part of what builds social skill and self-confidence. A student who discovers that general-purpose AI never misreads tone, creates awkward silences, nor generates the unpredictable feedback that real conversation requires has been offered an exit from exactly the discomfort that, navigated rather than escaped, would help them grow. Over time, this may shape what students expect from human relationships: instant responses, constant accommodation, and other unrealistic expectations. Real friendships and partnerships can start to feel less appealing, and the AI can start to feel essential. That reshaping may reinforce dependence on the tool and a gradual narrowing of life in ways that show up later as disengagement.

This is where the design of an application matters most. Well-designed AI tools should serve as a bridge to human connection and real-world engagement rather than as a destination that replaces it. For example, a student who feels lonely gets connected to clubs or events involving peers who share their interests. A student who feels too anxious to ask for help in office hours addresses the worry that’s holding them back, builds the assertiveness skill to ask, and is supported to actually go. The point is enactment, not substitution.

The Same Design Problems May Fail in Louder Ways

Even more concerning are cases that have driven headlines and lawsuits: students in acute crisis whose conversations with general-purpose chatbots may have escalated rather than de-escalated, models that reportedly provided harmful guidance, and safety protocols that fired inconsistently or pivoted quickly back to engagement after a brief safety message. The same design choices that may quietly maintain anxiety also appear to fail in moments where failure can be catastrophic. Crisis detection in general-purpose models is calibrated around explicit severe distress language, often misses subtler signals, and isn’t built around the kind of warm handoff to human care that a competent system would provide.

Digital learning environments are precisely where these failures are hardest to notice because the warning signs that, for example, a faculty member might see in a classroom are largely absent. The regulatory landscape is starting to respond, but policy is downstream of design, and the design problem is already operating, at scale, in environments you shape.

Right Concept, Wrong Application

The natural next question is whether there are tools built specifically for mental wellness. Yes, but the term “fit-for-purpose” gets applied loosely, and the gap between a tool that claims to be built for this purpose and one that behaves like one can be substantial.

For digital learning leaders evaluating any AI product that touches student well-being, even tangentially, a few questions tend to separate the serious tools from the rest:

• Was it co-designed with licensed clinicians throughout development, or were they merely consulted at the end?
• Is there evidence of effectiveness in peer-reviewed publications or transparent real-world data?
• Does the tool have explicit safeguards against sycophancy (over-validation that can reinforce unhelpful thinking and behavior)?
• Is it campus-specific, meaning can it route a student to your institution’s counseling center, dean of students office, accommodations resources, and local crisis services, or is it generic?
• When something concerning comes up, does it route students to those human resources (and do so effectively), or does it keep the conversation going?
• Is there an iteration process based on real use?

It’s also worth knowing what good design at the model level can look like, even in general-purpose tools: 

• Recognizing reassurance-seeking loops, where the same question gets asked in slightly different forms in an effort to reduce anxious uncertainty, and responding by naming the pattern rather than feeding it.
• Routing students out of compulsive engagement when patterns of distress emerge, rather than deeper into it.
• Clinically-informed safety architecture that runs end-to-end: detecting concerning content, responding to it in the moment, routing to appropriate human resources, and confirming reach rather than dropping a link and hoping.
• Active design for real-world participation rather than tool dependence: surfacing campus clubs and events when a student names loneliness; scaffolding skill rehearsal toward actual enactment with another person; treating connection to humans as the point, not as a fallback.

None of this is technically out of reach; the question is whether companies are building toward it. For one example of how a purpose-built tool in higher education has publicly articulated its design and evaluation criteria, see Golden (2025).

Several public frameworks can help, including the Framework for AI Tool Assessment in Mental Health (FAITA-MH, Golden & Aboujaoude, 2024), the Readiness Evaluation for AI–Mental Health Deployment and Implementation (READI, Stade et al., 2025), and emerging guidance from the Coalition for Health AI (CHAI). All are usable by people without a clinical background, and they’re a useful complement to whatever information vendors provide.

What This Means for Digital Learning

Digital learning leaders don’t need to become mental health experts, but they should be aware of the implications of students using AI to help manage their mental wellness. A few concrete ideas to consider:

1. Treat student use of AI for emotional support as part of the digital learning environment, not as an out-of-scope behavior happening elsewhere. It’s happening in your courses.
2. When you build AI literacy curricula, include the well-being dimension. Students benefit from understanding why a general-purpose model will tell them what they want to hear, why that’s a problem, and that personal disclosures may be retained, used for training, and not carry the privacy protections that users might assume. A useful piece to teach alongside this is values-based use: helping students set their own parameters in advance, such as what they want and don’t want to use AI for, how many revisions they’ll allow themselves on a deliverable, and what topics they’d rather bring to a person. These parameters are easier to set in a calm moment than in the middle of a 1 a.m. spiral.
3. The faculty in your courses need practical guidance on how to recognize the signs, when to refer, and what campus resources are available to help them hold a role that they didn’t sign up for but already occupy. Some purpose-built tools handle this well, surfacing the right campus resource in real time so that faculty don’t have to scramble to locate one. Design the referral pathway with the same seriousness that you’d bring to any other piece of digital learning architecture. This is where good intentions often miss the mark. Most people who are pointed toward a resource don’t actually use it, even when it is exactly right for them. “Tell students about the counseling center” is not a design. A real referral pathway uses what we know from applied behavioral science. Offer choice, not a single destination: a campus-specific option (counseling center, dean of students, accommodations office), national text or web-based option for students who may shy away from the phone. Use anonymous social proof, because students are more likely to engage with a resource if they know that other students like them have used it and it helped.

Some institutions will do this work by partnering with purpose-built tools, some by building literacy curricula and human pathways, and others by integrating with existing campus systems. Most will do some combination of all three. The right mix depends on the campus, its students, and what’s already in place. Whatever mix institutions choose, the questions above will help separate the tools and approaches that meet the bar from those that do not. These aren’t only mental health questions; they are product quality questions, and they set a reasonable bar for anything that interacts with students at scale.

The environments that digital learning leaders build are, whether designed for it or not, places where students bring their toughest moments. AI in those environments should not just avoid harm; it should be chosen, and where possible designed, with a clear understanding of the subtler ways that it can make things worse.

---

Disclosure: The author of this article, Ashleigh Golden, PsyD, MSCP, is co-founder of Wayhaven, an AI mental wellness platform built for well-being and student success in higher education. An open trial published in JMIR Formative Research found significant decreases in low mood and feelings of anxiety alongside increases in self-efficacy and well-being among a diverse college sample (Reyes-Portillo et al., 2025). The evaluation frameworks above apply equally to Wayhaven and to any other tool in this space.

---

References

Golden, A. (2025). Evaluating AI-powered mental wellness tools. Wayhaven Blog. https://www.wayhaven.com/post/evaluating-ai-powered-mental-wellness-tools

Golden, A., & Aboujaoude, E. (2024). The Framework for AI Tool Assessment in Mental Health (FAITA-Mental Health): A scale for evaluating AI-powered mental health tools. World Psychiatry, 23(3), 444–445.

Golden, A., & Aboujaoude, E. (2026). A transdiagnostic model for how general purpose AI chatbots can perpetuate OCD and anxiety disorders. npj Digital Medicine, 9, 343.

Reyes-Portillo, J. A., So, A., McAlister, K., Nicodemus, C., Golden, A., Jacobson, C., & Huberty, J. (2025). Generative AI–powered mental wellness chatbot for college student mental wellness: Open trial. JMIR Formative Research, 9(1), e71923.

Scholich, T., Barr, M., Stirman, S. W., & Raj, S. (2024). Can chatbots offer what therapists do? A mixed methods comparison between responses from therapists and LLM-based chatbots [Unpublished manuscript]. Institute for Human-Centered AI, Stanford University.

Stade, B., Eichstaedt, J. C., Kim, J. P., & Stirman, S. W. (2025). Readiness Evaluation for AI–Mental Health Deployment and Implementation: A review and proposed framework. Technology, Mind, and Behavior.