
Imagine your doctor tells you your blood sugar is in the prediabetes range and offers you a place on a lifestyle programme. Then adds: the coach will be an app. No human on the other end. Would you take it?
That is exactly what a trial published in JAMA last December set out to test — and the answer is more interesting than "the robot won" or "the robot lost".
What the trial found
Researchers at Johns Hopkins and a hospital in Pennsylvania recruited 368 adults with prediabetes and overweight or obesity (Mathioudakis et al., JAMA, 2025). The median age was 58, about seven in ten were women, and the group was fairly mixed: 61% white, 27% Black, 6% Hispanic. Each person was randomly referred to one of two versions of the Diabetes Prevention Programme — a structured lifestyle course built around weight loss and physical activity. England runs its own version.
Half were referred to a human coach-led programme, delivered remotely. The other half were referred to a version run entirely by AI, through a mobile app and a Bluetooth scale — no human coach at all. Both ran for 12 months, independently of the research team.
The researchers set a demanding target. To count as a success at 12 months, a person had to keep their HbA1c — the three-month average blood-sugar test — below the diabetes threshold and achieve one of the following: lose at least 5% of their body weight; lose at least 4% and clock at least 150 minutes of physical activity a week, measured with a wearable rather than self-reported; or lower their HbA1c by at least 0.2 percentage points.
The result: 31.7% of the AI group hit the target, against 31.9% of the human-coached group. A difference of 0.2 percentage points — well inside the margin the researchers had set in advance for calling the AI programme "non-inferior". The pattern held across each component of the target, and in the sensitivity analyses.
One detail caught my eye. More people actually started the AI programme after being referred — 93% versus 83% for the human-led one. When the coach lives in your pocket and doesn't need a diary slot, the first step is apparently easier.
What's still uncertain
Two-thirds didn't get there — either way. The headline is that the app kept pace with the humans. The quieter headline is that roughly one person in three reached the target in both groups. That is not a failure of AI; the programme itself is hard, and the target was strict. But it is a useful corrective to any pitch that makes an app sound like a guarantee.
"Non-inferior" is a specific claim. The trial was designed to show the AI version was not meaningfully worse, with a margin of 15 percentage points agreed beforehand. It did that convincingly — the lower bound of the confidence interval was about −8%. It was not designed to show the AI was better, and it didn't.
One year, two US sites. Prediabetes plays out over decades. Whether people who lost weight with an app keep it off at year three, and whether fewer of them go on to develop diabetes, is a different question. And a trial in Baltimore and Pennsylvania doesn't tell us how the same app would do with a 70-year-old who doesn't own a smartphone.
Who paid, and which app. The abstract names neither the funder nor the product. I've flagged that in the research notes rather than guess.
What you'll hear online
"AI is replacing health coaches." Not quite. What this trial shows is that, for a structured programme with clear targets, an automated version can match a remote human version over a year. It says nothing about the conversation where a coach notices you've gone quiet for two weeks and asks why — nobody measured that.
What you can do
If you've been told you have prediabetes, the single most useful move is to ask your doctor whether you're eligible for a structured prevention programme. In England that is the NHS Diabetes Prevention Programme, a nine-month course you can take face-to-face in a group or through a digital service; the NHS says completing it cuts the risk of type 2 diabetes by more than a third. The format matters less than the starting.
And if you're offered the app version and you'd honestly rather not discuss your weight with a stranger, this trial is quietly reassuring: it is a real option, not a second-class one.
The most human thing about this result may be that it doesn't matter much who — or what — is coaching. What matters is whether you show up for twelve months. Apps are good at making that easier to begin. The rest is still us.
Curiosity first.
— Dr. Brugal
Related on Dr. Brugal
If prediabetes is new to you, start here: The quiet signs you might be on the road to type 2 diabetes. and Prediabetic? What fasting can and can't do for you. For more on machines in the clinic: The artificial pancreas is on the NHS. Here's what that actually means. and The AI reading your scans before the radiologist does.
References
All studies retrieved via PubMed.
- Mathioudakis N, Lalani B, Abusamaan MS, et al. An AI-Powered Lifestyle Intervention vs Human Coaching in the Diabetes Prevention Program: A Randomized Clinical Trial. JAMA. 2025;334(23):2079-2089. Pragmatic non-inferiority randomised controlled trial, 368 adults with prediabetes and overweight or obesity, 12 months. doi:10.1001/jama.2025.19563 (PMID 41144242)
Awareness content — not medical advice. Always talk to your doctor or pharmacist before making changes.