Answers are grounded in this paper and cite their sources.
Universities are adopting artificial intelligence (AI) conversational agents that offer students emotional support faster than they are building the rules needed to govern them. This position paper argues that the features that make these systems valuable are the same features that make them risky. Constant availability, anonymity, human-like warmth, and data-driven scale each offer potential benefits while also creating corresponding governance concerns. Availability widens access but may reduce opportunities or prompts for human help-seeking; anonymity lowers the barrier to disclosure but can separate disclosure from a person who can respond; warmth sustains engagement, but invites misplaced trust and dependency; and scale extends reach at low cost but concentrates privacy exposure and can amplify model limitations across student populations. Because benefit and risk originate in the same property, technical safeguards alone are unlikely to address the institutional responsibilities associated with these systems and must therefore be complemented by governance. Drawing on randomized trials, meta-analytic and review evidence, campus deployments, and established ethical and psychological theory, the paper shows that evidence of benefit is concentrated among students with mild to moderate distress over short periods. In contrast, evidence on dependency, crisis response, long-term outcomes, and equitable performance across student groups remains limited. The paper, therefore, advances a position in which universities adopt AI emotional support only as a bounded supplement to human care, within a proposed tiered governance framework that specifies triggering criteria, required actions, responsible roles, and escalation routes, and that keeps human professionals accountable for decisions involving clinical judgment, safety, escalation, or significant student well-being. The paper responds to five counterarguments and concludes with recommendations for policy, practice, and research.
No takes yet. Share an insight, caveat, or question.
Adebayo et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: