Digital twins have moved from factory floors and Formula 1 garages into people’s kitchens. In a new installment of her Optimizer newsletter, The Verge AI senior reviewer Victoria Song describes roughly 100 days with Twin Health. The service builds an AI model of your metabolism and then tells you, meal by meal, how your blood sugar is likely to react. Her verdict points to a problem the whole personalized health industry will have to solve: the twin works, but it takes a lot of data to keep it going.
🧬 What a Health Digital Twin Actually Is
In medicine, a digital twin is a virtual model of a person. It’s built from lab results, wearable data, imaging and sometimes genetic profiles, and it keeps updating over time. The long-term idea is that clinicians could “test” a drug or treatment on the model before they try it on the patient. That could help predict how a disease will progress, flag injury risk and cut down on side effects.
Song’s twin is a far more modest version. According to The Verge AI, it’s “an AI-powered, color-coded food database” that predicts glucose responses:
- Green means ideal
- Orange means so-so
- Red means not so good
Protein Cheerios with milk come up orange if she eats them on their own and green if she takes a walk afterward. A turkey hot dog with ketchup and relish is green. A cup of pineapple is orange.
The setup is heavy. Twin Health sent 90 days’ worth of Dexcom G7 continuous glucose monitors, a smart scale, a blood pressure cuff and supplements, and it connected to her Apple Watch. Then came a blood draw at Quest Diagnostics and Zoom reviews with a doctor and a human coach. The service isn’t available to download yet. You can only get it through corporate or private health plans, and it’s mostly aimed at people with chronic metabolic conditions.
⚖️ The Upside and the Catch
This matters because metabolic coaching is, as Song puts it, “perhaps the lowest-stakes, highest-reward version” of digital twin technology you can test in the real world today. A wrong food prediction won’t hurt anyone. If the predictions are right, though, they could add up over time for millions of people with insulin resistance, fatty liver disease or type 2 diabetes.
The catch is the feeding. Song describes the twin as constantly hungry, and keeping it fed left her feeling “drained.” Her coach told her that tracking fatigue is common feedback from Twin Health users. That’s notable coming from someone who already tracks and reviews health data for a living.
Two things stand out to me:
- Behavior change came partly from being watched. Song says she skipped treats because she didn’t want “to get scolded by apps.” That’s an effect of logging, not of AI prediction. It raises a fair question about how much of the benefit comes from the model and how much from simply paying attention.
- The human layer is still holding things up. The doctor review, the coach’s goal-setting and the follow-up calls every two weeks are what make the program hang together. The AI personalizes the advice, and people keep users engaged.
🔮 Where This Goes in the Next 1-3 Years
This is my forecast, not Twin Health’s. Expect three shifts:
- Passive data collection will win. The companies that cut down on manual logging will keep users. That likely means models that estimate meals from photos, read glucose automatically and pull more signals from smartwatches.
- Employers and insurers will act as gatekeepers. Twin Health already sells through health plans. Getting paid will depend on clinical results, such as fewer medications or lower HbA1c, more than on how well the app works.
- Data scrutiny will grow. A model built from bloodwork, body composition and minute-by-minute glucose is some of the most sensitive data a company can hold. Regulators and enterprise buyers will ask who owns the twin and where it ends up.
🛠️ What Practitioners and Businesses Should Do
- If you’re building health AI: Count logging burden as a core product metric. A more accurate model doesn’t help if users stop feeding it by day 40.
- If you’re an employer weighing these benefits: Ask vendors about engagement after 90 days and about measured clinical outcomes, not just how many people signed up.
- If you’re a user: Treat the colored labels as guidance, not orders. The data is useful, but the decisions are still yours.
Digital twins for health are moving from an idea to a subscription. Whether they become standard care depends less on smarter AI and more on whether people can keep feeding them. Song’s full account, including what her 90-day bloodwork showed, is available at The Verge AI.