Can AI Texts Get You Walking? The DIAMANTE RCT

An AI system that learns which motivational texts you respond to may get you moving more than generic messages do. In the 24-week DIAMANTE randomized trial, adults with type 2 diabetes and elevated depression symptoms who received reinforcement-learning-guided texts took about 608 more steps per day than an app-only control group—a roughly 19% increase from baseline. The result was statistically significant. Whether that extra walking improved blood sugar control or depression symptoms remains less certain.

That distinction matters. A step-count increase is a real behavioral outcome, and behavior change is hard. But 608 additional daily steps are not automatically a treatment for diabetes or depression, no matter how sophisticated the notification engine sounds.

AI personalization beat random motivational texts

The 2024 DIAMANTE trial randomized 245 adults with comorbid type 2 diabetes and clinically significant depressive symptoms, defined as a PHQ-8 score of at least 5, to one of three groups for 24 weeks. The adaptive group received text messages selected by a reinforcement-learning algorithm. A second group received messages with randomly selected content. The control group used the app and completed a weekly mood check but received no motivational messaging program.

The adaptive group increased daily step counts by approximately 608 steps per day relative to control. The random-message group landed between the adaptive and control groups. That three-arm structure is the study's best feature: it tests whether messages help compared with passive app use, then asks whether adaptive selection adds something beyond sending motivational content at random.

The direction of the results supports both ideas. Any structured messaging appeared better than app-only participation, while adaptive messaging performed best. The bigger gap was between messaging and no messaging; the added advantage of reinforcement learning over random content was smaller.

DIAMANTE is among the first adequately powered completed randomized trials to compare reinforcement-learning personalization with both randomized messaging and passive control in a behavioral-health population. Reviews of just-in-time adaptive interventions describe reinforcement learning as an emerging method, not a settled standard. A 2024 systematic review of JITAIs and a 2025 scoping review of physical-activity JITAIs found that most adaptive programs still use rules or behavior-change theory rather than formal machine learning.

What the algorithm learned—and what the paper does not show

DIAMANTE's contextual-bandit algorithm chose messages across three dimensions: the kind of feedback, the motivational framing, and the time of delivery. It used information such as age, baseline PHQ-8 score, and the timing and categories of previous messages. In principle, that lets the system learn whether one person responds better to an early-afternoon progress update while another responds to an evening prompt built around a personal goal.

In practice, the trial report does not provide convergence diagnostics, say how long the algorithm took to settle on useful choices, or show whether it discovered individual preferences rather than broadly popular message types. That is a meaningful gap. An algorithm that learns that most people prefer one kind of reminder can outperform random messages without demonstrating deep individual personalization. It would still be useful; it would simply be a different claim.

Why steps are a sensible target

Step count is not a frivolous endpoint. A 2024 systematic review and meta-analysis found an inverse dose-response association between daily steps and both depression incidence and depression prevalence in adults. More steps tracked with lower depression risk, although those data do not establish that adding 608 steps will improve symptoms in someone who already has depression.

Higher device-measured daily step counts also correlate with more favorable cardiometabolic markers in prospective data. The ProPASS Consortium study supports step count as a reasonable intermediate marker for metabolic health, including markers relevant to diabetes. It does not provide a causal HbA1c threshold for an increase of about 600 steps among people with established type 2 diabetes.

Smartphone-derived step counts also capture useful physical-activity information at the population level. A 2025 measurement study found that simple step counting conveyed health-relevant information comparable to more complex accelerometer measures. That supports DIAMANTE's choice of endpoint, though it does not erase variation among phone models, carrying habits, and app behavior.

DIAMANTE shows that adaptive texts can change a measured behavior. It does not yet show that the size of that change treats either disease.

The clinical question remains open

Exercise can reduce depression symptoms, but dose and intensity matter. A 2025 meta-analysis of randomized trials found benefits from high-intensity exercise in people with depression. Those interventions generally involved much more exercise than roughly 600 extra, likely light-intensity steps per day. Evidence from walking programs also cannot be transferred without caution: a meta-analysis in postpartum depression found that walking interventions reduced symptoms, but postpartum depression differs substantially from depression occurring alongside type 2 diabetes.

DIAMANTE's reported step result therefore has statistical importance without a settled clinical interpretation. The available trial reporting does not establish whether the step gain changed HbA1c or PHQ-8 scores at 24 weeks. A statistically significant increase in walking can be worthwhile, especially in a population facing two conditions that make sustained activity difficult. It is still not proof of a clinically meaningful change in glycemic control or depression severity.

Reinforcement learning may not be the essential ingredient

The trial also cannot tell us whether reinforcement learning beats simpler forms of personalization. In the IMPACT randomized trial, young adult cancer survivors used an adaptive, theory-based mHealth program without formal reinforcement learning. The intervention improved physical activity and maintained gains at 12 months, according to the primary trial report and its 12-month follow-up.

IMPACT does not overturn DIAMANTE. It studied a different population, used a different intervention, and followed different outcomes over a different schedule. But it makes the right comparison harder to avoid: perhaps the benefit comes from thoughtful, consistent adaptation rather than reinforcement learning specifically. DIAMANTE compared adaptive AI with random messages, not with a well-designed rule-based adaptive program.

Limits of a difficult real-world trial

Behavioral messaging cannot be blinded. Participants could tell whether texts became tailored, and that awareness could affect expectations, engagement, and walking independent of the algorithm's selections. The same problem applies to study staff. This is common in digital behavior trials, but it limits how confidently anyone can attribute the adaptive arm's advantage to reinforcement learning itself.

The January 2020 to June 2022 trial also ran through major COVID-19 disruptions, when population walking patterns changed sharply. Midway through recruitment, the study shifted from clinic recruitment with electronic-record eligibility confirmation to web recruitment through platforms such as Facebook, Google Ads, and Craigslist, where diabetes status was self-reported. Those recruitment streams likely enrolled people with different levels of health engagement and digital comfort. The published results available here do not report subgroup outcomes by recruitment source.

Participants received $40 at baseline and $70 at 24-week follow-up. Compensation can help retain people in demanding studies, but it can also favor the retention of participants most willing to complete outcomes. The available report does not provide arm-specific attrition details. Nor does it report whether effects differed by English or Spanish preference, despite language-stratified randomization.

What to do with this result

DIAMANTE offers a credible proof of concept for AI-guided health coaching. Its strongest result is modest but useful: in 245 adults managing diabetes and depression symptoms, adaptive reinforcement-learning texts increased daily steps by about 608 relative to passive app use over 24 weeks, with random messages performing in between.

For individuals, the practical lesson is less futuristic. A daily step goal, a reminder system, and some form of accountability can help; waiting for a perfectly personalized AI coach is unnecessary. For clinicians and product builders, the lesson is stricter: test adaptive systems against strong non-AI adaptive comparators, report whether algorithms truly learn individual preferences, and measure hard outcomes such as HbA1c and depression symptoms. Until then, the evidence supports AI messaging as a way to nudge walking—not as a proven treatment for diabetes or depression.