The metabolic AI twin that made better habits feel like another job

Victoria Song’s months-long experiment with a metabolic digital twin shows how personalised health coaching can improve habits while demanding an exhausting stream of intimate data.

Victoria Song’s long experiment with a metabolic digital twin offers a rare first-hand look at where this still-emerging health technology is useful, and where it begins to feel intrusive. In the Verge newsletter Optimizer, Song describes spending months feeding an AI system a steady stream of personal data, from continuous glucose readings and blood pressure measurements to lab results, activity tracking and body composition figures, in order to help the Twin Health platform learn how her body responds to food and exercise.

The process, as she recounts it, was deliberately mundane and repetitive: logging meals, taking readings, waiting for the app to colour-code foods as favourable or less favourable, then checking whether a walk could turn an orange meal into a green one. Twin Health’s own approach relies on a wide net of connected devices and recurring measurements, and Axios recently reported that the company says it gathers more than 3,000 data points a day in some programmes, using that information to build personalised metabolic profiles that can even take account of factors such as menstrual-cycle stage and hormone shifts.

That kind of granular tracking is part of the appeal of digital twins in medicine. A review published in ScienceDirect says the technology is particularly well suited to diabetes and other metabolic conditions because wearables and glucose monitors can produce continuous real-world data, allowing machine-learning models and more traditional mechanistic systems to forecast how a patient might respond to treatment. Researchers argue that these virtual models could one day help clinicians test therapies before using them on patients, improving glucose control and reducing unwanted side effects.

But the same studies also underline the limits. The broader literature, including a 2026 review in Frontiers, says digital twins will only become truly useful in personalised medicine if they can combine many layers of information rather than relying on isolated data streams. That means blending physiological measurements with other signals, using more advanced tools such as causal inference and hybrid mechanistic-AI models, and creating governance frameworks that address bias, privacy and regulation. Another review, published in 2025 by Springer, describes digital twins as potentially valuable learning systems for preventive and personalised care, but says their clinical promise depends on solving the problem of translating research into everyday practice.

Song’s account highlights those tensions in unusually human terms. She says the system nudged her towards better habits, increased accountability and helped motivate a return to running and strength training. She also reports measurable changes after 100 days, including improved liver markers, lower triglycerides and a better metabolic health score, although she notes that some improvements had already begun before she joined the programme. That matters because it suggests the digital twin may have amplified progress rather than creating it from scratch.

At the same time, the experiment exposed the emotional cost of being continuously measured. Song writes that the system felt hungry for information and that tracking fatigue became part of the experience, echoing concerns raised in recent reporting and academic reviews that these tools depend on persistent, sensitive data to remain accurate. Frontiers and other reviews warn that the more ambitious digital twins become, the harder questions of trust, consent and data ownership will be to avoid, especially if the technology moves beyond lower-stakes metabolic coaching into areas such as cancer care or cardiovascular disease.

For now, the article’s larger lesson is that digital twins may be most persuasive when they behave less like science fiction and more like structured, data-heavy coaching. They can already help some people understand how food, movement and medication interact with their bodies, but the evidence also suggests that their value rises only when the user is willing to keep supplying information and when clinicians are cautious about what those virtual models can really predict. Song’s experience captures both sides: the promise of personalised feedback and the burden of maintaining it.