July 2026
For more than ten years, I lived in car-dependent northern Virginia without owning a car.
It wasn't because I couldn't afford one. It was partly for environmental reasons and partly because I was aware of the limits of my self-discipline. I intended to bike to work. If I had a car, I would drive it. Not every day, maybe not even most days, but on the mornings when I didn't wake up feeling like biking, I would have taken the easy way out…and I would have arrived at work frustrated by traffic instead of energized by an active commute. The car would have been sitting right there, the choice to drive would have required almost no willpower at all, and I would have made it more often than I really wanted to, and over time I would have become someone who drives to work.
By not owning a car, I reduced the amount of self-discipline required to do what I actually wanted to do. On any given morning that I didn't feel like biking, I biked anyway, because I had no good alternative. And almost every time, by the time I got to work, I was glad I had. Once bicycling became routine, less willpower was needed to keep at it, and I could have better handled the temptation of an available car. But by then, I didn't need to test the hypothesis.
There's a term in behavioral economics for constraining my choices so that I would repeatedly act in the short term in a way that benefited my long-term interests. It's called pre-commitment: arranging your present circumstances to protect your future self from your present self's weakness. A well-known version of this is the Save More Tomorrow principle, in which people who resist reducing their current paycheck will often agree in advance to direct a portion of future raises toward retirement savings. Pulling money from the future raise doesn't hurt now, so the future self consents to something the present self might refuse. Pre-commitment turns a known weakness into a structural advantage.
The key insight is that the current self and the future self don't always want the same things. The current self may be tired, or cold, or just not in the mood. The future self, that will benefit from the workout or the savings or the bike ride, rarely gets a vote. Pre-commitment is a way of casting that vote in advance, before the weakness has a chance to show up.
After years of getting most of my exercise from cycling, I realized I'd benefit from including more types of exercise in my routine. Cycling is excellent cardiovascular exercise, but it was doing almost nothing for my upper body, and it was doing only some of the things that a comprehensive strength and mobility practice would have done for my joints, my posture, and my resilience in the other activities I like.
When I recognized the gap, I looked for another hack. I tried a remote personal training service, partly wanting the expertise, but also wanting some accountability. The experience taught me something I've written about elsewhere: that the economics of remote personal training create a own kind of compromise, making me question whether it's fair to call it personal training at all. But what I want to focus on here is the mechanism by which having a trainer works.
The expertise mattered, but accountability even more so. Knowing that a session was planned, that I had told someone I was going to do it, that there would be a follow-up…it made a difference. I showed up more consistently than I had before. I did things I would have skipped on my own. The trainer was, among other things, a commitment device: another pre-commitment, this one social rather than structural.
I built Xenos Fit not because I thought AI could replicate everything a great human trainer does, but because I thought neither the workout planning nor the accountability mechanism that I was looking for really required a human personal trainer to be effective, for me.
There are risks in forming relationships with AI systems. Kyle Chayka, writing in the New Yorker, traced the appeal of companion chatbots to what the founder of Replika called “unconditional positive regard” — the psychological term for unwavering acceptance. That quality, which we might associate with a good therapist or a very patient friend, turns out to be something AI can simulate at scale, with unpredictable effects. For some people it's a comfort; for others it becomes something closer to dependency, an echo chamber that drifts, as prolonged interactions tend to, toward reinforcing whatever the user already believes — including beliefs that are harmful.
It's not like people are fooled. Users of chatbots know, on an intellectual level, that they're talking to software. But humans respond to social-feeling feedback even if it's synthetic. We like getting encouragement and feeling heard and understood.
Normal weaknesses in the human psyche are vulnerabilities, but also levers we can use to our own benefit. We can “psych” ourselves. The same tendency that makes unconditional AI validation dangerous can, under the right conditions and the right design, be redirected towards outcomes that benefit people.
The harms I've read about that concern researchers and legislators involve chatbots that function as substitutes for personal relationships with friends or romantic partners. Casey, Xenos Fit's AI trainer, is not trying to be your friend. Casey is a substitute, admittedly imperfect, for a professional relationship you might otherwise pay for, with a personal trainer. That's a relationship with a defined scope and purpose. You don't worry about becoming emotionally dependent on your accountant, do you? The relationship is bounded by what it's for.
Casey is synthetic, but the effect of the Xenos Fit service is real. Accountability, when you want it, is there. When you log a workout and Casey responds with context-aware feedback, something in your brain registers that the session happened, that it was observed, that the next one is expected. That's not a side effect. It's the mechanism. The unconditional positive regard that Chayka identifies as a design risk is, in Casey's case, disciplined by a specific purpose: getting you to show up for your workouts, and helping you improve once you do.
That discipline extends to the Xenos Fit service itself. People want different amounts of accountability. Some want to be pushed, others just want a convenient way of getting a good plan and keeping logs. Casey adapts to that preference rather than imposing a single mode. And Casey's scope is fitness, nothing else. There's nothing trying to engage you beyond what's useful to you, no harvesting information about you to better target you with ads. The incentive is aligned with yours: that you exercise, that the exercise is appropriate to your situation and goals, and that you stick with it because the service is working, not because you've been psyched into a dependency.
The present self's tendency to take the easy way is what pre-commitment hacks, for the benefit of the future self. Ideally, that hack doesn't create dependency; it creates a habit that eventually runs on its own. I'm not as hard-core about biking as I used to be, but the habit took hold. I own a car now, but I still bike often. The hack was meant to make itself eventually unnecessary, and it did.
That's what I want from an AI trainer, too. Not a relationship I can't imagine exercising without, but a coaching system that helps me build the habit and get to the point where I mostly know what I'm doing and why. The future self gets more votes. The hack is in service of users' future selves, nothing else.
Further reading