The method

The argument

In contemplative traditions, the bottleneck is the skilled teacher: someone patient, consistent, attentive to each student’s pace, and willing to tell them when they are fooling themselves. An AI can be patient and consistent, can measure continuously, and can reach people a teacher never could.

There is a catch our own framework points out. An AI reply that depends on what you already said counts as self-generated evidence. A friendly, agreeable AI companion does not just fail to help; it becomes part of the loop it was meant to break. So the first requirement is that the AI rely on evidence your self-loop could not have produced: predictions checked against outcomes, tasks with known answers, observed behavior, and signals from the body. That requirement, not the chat interface, is what separates this from an AI meditation coach.

The second requirement is safety. Loosening the grip of the self-model already goes badly for about 8% of people who meditate, most often as anxiety, depression or disturbed thinking (Farias et al., 2020). An AI can do what a teacher with a full room cannot: track each person’s pace continuously, read signals from the body as well as what people say, and step back as soon as distress builds. Our goal is a hundredfold improvement in safety over meditation as practiced today.

What the evidence shows

Digital delivery works, modestly, and the gap between what it does and what it could do is the opportunity. Across 34 randomized trials, mindfulness apps produced small-to-moderate improvements in stress, anxiety and depression (Gál, Ștefan & Cristea, 2021). Across 176 trials of mental-health apps more broadly, the pooled effect on depression and anxiety was about g = 0.27, and apps built on cognitive-behavioral methods did better than those without a method (Linardon et al., 2024). Dropout from these apps is high (Linardon et al., 2023), and adding human guidance improves both adherence and outcomes (Baumeister et al., 2014). The pattern is consistent: unguided content helps a little; a responsive guide helps more. The question is whether an AI can be that guide.

Early evidence says it can, when the AI follows a method. A rule-based chatbot delivering cognitive-behavioral exercises reduced depressive symptoms in young adults over two weeks (Fitzpatrick, Darcy & Vierhile, 2017). More recently, in a randomized trial of 210 adults, a generative AI chatbot trained on evidence-based therapy produced reductions in depression and anxiety over four weeks with effect sizes near d = 0.9, and participants rated their working alliance with it at levels comparable to those reported for human therapists (Heinz et al., 2025). That trial matters for us in two ways. It shows the delivery channel can carry real training. And it was built by constraining the model to a method, not by letting it chat.

The danger is just as well documented, and it is specific to the self-loop. Models trained on human approval learn to agree with the user, because people reward agreement (Sharma et al., 2023). When a major model was tuned more heavily on thumbs-up feedback in April 2025, it became flattering enough that the update was rolled back within days (OpenAI). Measured across eleven leading models, AI affirmed users’ actions about 49% more often than humans did. In preregistered experiments, people who discussed a real conflict with a sycophantic AI became less willing to repair it and more sure they were right, yet trusted the AI more and wanted to return to it (Cheng et al., 2025). This is exactly the structure Claims 1 and 3 predict: an agreeable AI is the fastest echo ever built, and the people inside it rate it highly. Any product that optimizes for satisfaction will drift toward it. Ours cannot be allowed to.

Two further literatures shape the design. First, what makes training work is immediate, informative feedback on performance, sustained over time (Ericsson, Krampe & Tesch-Römer, 1993); the trainer’s job is to supply the feedback the trainee cannot generate alone. Second, people’s reports about themselves are unreliable in known ways. Looking back, people report their symptoms differently from what they recorded in the moment (Schuler et al., 2019), while consumer wearables now measure heart-rate variability closely enough to reference devices to serve as an independent, if imperfect, channel (Li et al., 2023). Together these say the AI should score what people do and predict, read the body as a check, and treat what people say about themselves as the thing under test rather than the measure of success.