Can AI Replace Human Counselors at Scale? A Nationwide Experiment to Reduce Teacher Shortages

The effectiveness of many interventions depends on implementer quality; as programs scale and the pool of implementers broadens, average quality falls. Conversation-based interventions, such as counseling or peer mentoring, are particularly sensitive to implementer skill. LLMs offer a natural way out: they can hold personalized exchanges at zero marginal cost and uniform quality. Whether AI can replicate, or even improve on, the human skills thought to make such interventions effective remains an open question. We run a nationwide experiment in Chile (N=41,000) comparing an AI-powered chatbot, Kai, with trained human counselors. Both arms reach high school seniors via WhatsApp and aim to motivate them toward teaching careers, addressing future teacher shortages. Among students with baseline interest in teaching (the main target of the intervention) Kai significantly increases the probability of ranking an education major first as well as the share of education majors in students' choice sets; estimates in the human arm are smaller and not statistically significant. We use text analysis to characterize conversational attributes and test mechanisms. Both arms cover a comparable breadth of topics, but Kai concentrates on factual content while human counselors lean on subjective, socioemotional material. Kai's exchanges are substantially faster and more dynamic, with near-instantaneous responses and shorter student gaps, and more semantically coherent on both sides. Human counselors, in turn, score substantially higher on motivation and empathy. We interpret these results as evidence that AI relaxes the implementer-quality constraint in conversation-based interventions at scale.