WHOOP, AI and Health Data: What the All-In Interview Reveals
Will Ahmed’s All-In interview raises bigger questions about wearable health data, AI training, privacy and the evidence behind personalised insights.
By My Bio Age
What happens when a fitness tracker becomes a health data company? In the All-In interview published on 29 October 2026, WHOOP founder and CEO Will Ahmed described a business built around continuous measurement, personal context and increasingly ambitious AI. It is a useful starting point for a question that extends well beyond one wristband: what should we expect in return for sharing more information about our bodies?
Ahmed reported more than 50 billion hours of physiological data and discussed WHOOP developing its own models using that information. These are company statements, rather than independently audited measurements. The interview’s headline also mentions four million members; in the conversation, Ahmed presents that as an expectation for the end of 2026. The distinction matters when turning a compelling business story into a health technology claim.
What WHOOP collects beyond a heart-rate trace
WHOOP’s privacy policy lists heart rate, HRV, breathing, temperature, movement and sleep information, alongside optional details about habits, diet and medication. Depending on the services you use and information you supply, it also covers lab results, clinical notes and Coach conversations. The dataset can extend well beyond sensor readings.
The appeal of combining these sources is context. A sleep graph records one part of a night; a journal entry about travel may help someone interpret it. Repeated observations could help a system distinguish a person’s usual pattern from an unusual week. Linking signals to reliable outcomes could also support research into which patterns are informative. Those are useful possibilities, provided the resulting features are tested against the questions they claim to answer.
Using your data can mean several different things
An app can calculate a score from your recent readings, give selected information to an AI model to answer your question, or use examples from many people to develop a model. All three involve data, but they are different activities with different implications for privacy and consent.
In machine learning, training changes a model using examples. Inference applies a trained model to new input. For a chatbot, that means generating a response to a prompt. Providing your recent sleep summary as context can personalise an answer without that particular interaction retraining the model. Google’s machine learning glossary explains this distinction. Whether information is also retained or used for later training is a separate question for the service’s policies.
That distinction is essential to understanding WHOOP. Its privacy policy, updated 3 August 2026, restricts its third-party Coach language model partner from retaining personal data or training on WHOOP metrics. Separately, it permits WHOOP research and development using aggregated or de-identified information and describes retained Coach history with deletion controls.
These provisions concern different organisations and purposes. A partner’s no-training commitment does not rule out WHOOP’s own model development, and WHOOP’s development work does not establish that the external provider trains on member records.
There is a concrete example beyond chat: WHOOP’s own explanation of workout detection describes models trained on member data to recognise activity patterns. An algorithm identifying a bike ride and a conversational model explaining your week need not be the same system. Asking which system uses which information is more useful than treating every feature labelled AI as one process.
A large dataset still needs evidence of useful results
Continuous readings could help researchers ask questions that occasional measurements leave unanswered. They also create a demanding evaluation problem. Fifty billion hours describes volume. It does not tell a reader how accurately a specific model performs, who it works for, or whether following its suggestions improves anyone’s health.
Consider a hypothetical app that associates late dinners with a lower recovery score. Before changing your routine, you might ask whether those were also late work nights, travel days or evenings with less time in bed. A pattern can be useful for forming a question even when it cannot isolate a cause. Adding an AI explanation does not resolve those missing details by itself.
The same care applies to ambitious predictions. In the interview, Ahmed discussed early indications that physiological signals might help identify abnormal blood biomarkers. That is a research possibility, not evidence that a wristband can replace a laboratory test. The useful follow-up questions concern the precise outcome, the comparison test, missed cases, false alarms and performance in new groups of people.
This emphasis on evaluation has wider support. The World Health Organization’s guidance on generative AI in health calls for well-defined tasks, appropriate accuracy and reliability, and independent assessment of large-scale deployments. It recommends examining outcomes across different groups. Those are more informative standards than dataset size or how fluent an answer sounds.
For a reader comparing health apps, the practical implication is to look for evidence that matches the feature. A study of heart-rate measurement does not automatically validate a recovery recommendation. A convincing explanation does not establish a diagnosis. An app should make it easy to distinguish a recorded value, a calculated estimate and an AI interpretation.
What to ask before connecting more health information
Start with the job you want the app to do. Reviewing sleep patterns may call for different information from organising existing lab results. Before connecting another source, check whether the benefit is clear and whether you can use the features you need with fewer permissions.
- Collection: Which readings, records and written answers will the app receive, and which are optional?
- Processing: Does the information stay on your device, go to the app company, or reach another AI provider?
- Development: Is it used only to deliver your experience, or also for research and model training? What choices apply?
- Control: What do export, permission changes and deletion each do, including for saved conversations?
My Bio Age takes a local approach: its health calculations and optional AI Coach run on your iPhone, with a model download for generated answers. Our guide to local health processing explains the boundaries, including sharing you choose and Apple’s separate storage settings. Local processing addresses where analysis happens; it does not guarantee that an AI answer is correct.
The most useful health technology should help you understand what changed, see the information behind an explanation and decide what deserves attention. That is a reasonable expectation whether your tool is a wristband, a phone app or an AI assistant. To explore the Apple Health history you already have, download My Bio Age free on the App Store.