<p>As part of the expansion of digital mental healthcare and the growing attention towards artificial intelligence in psychiatry, Large Language Models (LLMs) are being explored as devices for the delivery of psychotherapy. In this paper I consider what it would take for LLM-based psychotherapy chatbots to be effective, and argue that this question leads to a deep and difficult problem: the threat of <i>value capture</i>. In short, there is a risk of these systems becoming overly optimized towards improving outcomes on simplified, standardized psychometric scales. This may cause the systems to guide users towards adopting values which have been constructed based on convenience for institutions and developers, rather than fostering the type of self-reflection that would yield values better tailored to an individual. I outline the concept of value capture, explain how it applies to this context, and motivate the idea that it is a problem we should be concerned about. I then give suggestions about how the problematic aspects of value capture in the case of LLM psychotherapy chatbots can be at least partly mitigated.</p>