Controversy already surrounds wearable fitness trackers, such as those made by Fitbit, Apple, Garmin, Samsung, or Oura, for their poor privacy practices. However, there is little public concern around potential biases in many of these trackers’ features. In this report, I will discuss just two features with evident bias: PPG-based heart rate monitoring, and accelerometer-based activity tracking. Minority groups miss out on wearables’ espoused benefits due to their poorer accuracy for such groups. They may also suffer physical and psychological harms, as these technologies contradict users’ own senses, push them to over-exercise, or fail to flag serious cardiovascular problems.
Before the widespread adoption of smartwatches, there was evidence that the technology underlying wrist-based heart rate (HR) monitors was less accurate for darker skin tones, as melanin absorbs more of the light emitted from the sensor (Fallow et al., 2013). Despite this inherent limitation, photoplethysmographic (PPG) signalling is still the dominant heart rate technology for wearables, and such devices are marketed without note of this limitation (Apple, 2025; Garmin, n.d.). Accuracy has also been found to be worse for those with a higher body mass index (BMI) (Shcherbina et al., 2017). Recent research has shown that even the most up-to-date devices show racial bias, and that accuracy is worse for more intense exercise (Hung et al., 2025).
There are many risks from wearables using biased PPG-based HR monitoring. When underestimated, affected users may be pushed beyond their limits in a guided workout. For those with darker skin, underestimation may exceed 20% (Asif et al., 2025). Automation bias, where a user places greatest authority in technology, enables these harms, and occurs across all demographics (Cummings, 2017). However, those who are more likely to defer to biased HR monitoring, rather than their own senses, may include those who are least confident exercising, and so most susceptible to harm as a result of exercise. Additionally, smartwatches now include features which may be critical for vulnerable groups, such as arrhythmia detection, hypertension monitoring, and telehealth reliability (Asif et al., 2025). These features, if relied on but inaccurate, may lead to more devastating impacts, disproportionately affecting those with darker skin, higher body mass index, and/or greater automation bias. In clinical settings, PPG is also used, and may lead to long-term underdiagnosis of cardiovascular problems for those with darker skin (Kerry Setchfield et al., 2024). Additionally, public health research, often used to justify wearables’ privacy harms, will fail to benefit affected groups when using biased wearable data (Brodie et al., 2018).
Another basic feature of modern smartwatches is activity tracking. Accelerometer readings from wearables are interpreted by machine learning (ML) algorithms, which have been trained on datasets such as HHAR (Blunck et al., 2015) or PAMAP2 (Reiss, 2012). Some of these datasets specifically collect data for minority groups, such as for those with Parkinson’s disease (Roggen et al., 2010), or for those in Ambient Assisted Living (Davis & Owusu, 2016). However, these datasets tend to be small, compromising accuracy in general, as well as bias (Barocas et al., 2023, p. 6). Indeed, even typical users may be underrepresented in wearables’ training data, if that data is sourced from athletes with a high level of experience practising certain moves (Radanliev, 2025, p. 6). This feature leads to the same harms mentioned above. Wearable users who received deflated step counts ‘perceived their activity as more inadequate; ate more unhealthily; and experienced more negative affect, reduced self-esteem and mental health, and increased blood pressure and heart rate’. (Zahrt et al., 2023)
Both HR and activity tracking are used to compute derivative statistics, such as sleep score, calories burned, and training readiness (Colvonen et al., 2020). There is evidence that inaccurate scores cause anxiety, as users receive scores which do not reflect how they really feel. Users who received sham negative feedback evidenced impaired daytime function and increased evening fatigue (Gavriloff et al., 2018). Some users may experience increased anxiety from the constant surveillance of such devices: anxiety which is only compounded by biased readings (British Heart Foundation, 2025).
Those impacted by bias in PPG-based HR monitoring and activity tracking include those with darker skin, higher BMIs and non-typical movement patterns. Crucially, users may belong to all affected groups, and so experience multidimensional harms (Crenshaw et al., 1991).
Those impacted may share many of the same interests as well-represented wearable users. These may include greater fitness, longevity, and overall wellbeing. These are all embodied as virtues to be cultivated, rights to be respected, or simply sources of utility in normative ethical theories. Wearables, even when biased, may further these interests for all users. On the other hand, we should consider that fitness trackers, biased or otherwise, reflect our often impersonal, individualistic, and competitive society. Undesirably, exercise is increasingly quantified and broadcast in exchange for “kudos” on fitness-centric platforms such as Strava (Couture, 2020). So, fitness wearables may not represent affected groups’ interests in social interaction, wellbeing, and care for others.
Additionally, data-driven markets are forcing consumer adoption of these biased trackers. For example, “interactive life insurance” offers cashback or cheaper premiums for those with higher “fitness scores”, as calculated by the insurance company (Spender et al., 2019). In addition to bias from HR and activity tracking, the research underpinning such insurance practices is not reflective of all demographics (Heaton, 2025). One insurer says its assessments encourage you to “train to live longer” (Athlete Life Insurance, 2026), but, in reality, without supervised training, consumers may engage in dangerous or unhealthy short-term training to improve their score. These markets’ adoption of wearables will only perpetuate the biases present within them.
On the other hand, unbiased wearable technology will only encourage these markets’ growth, leading to the undesirable expansion of surveillance capitalism and exploitation of consumers’ information capital (Gidaris, 2019). Before subjecting impacted groups to these practices through debiasing technology, regulators should investigate insurers’ recent adoption of wearables for surveillance-powered price differentiation, and establish its fairness.
In response to the evident bias concerns, I recommend the following interventions:
- 1.
Advertising regulators should require manufacturers of fitness wearables to disclose their potential for bias, as well as inaccuracy. Relatedly, “seamful” design should inform wearables’ design, exposing limitations to users rather than encouraging automation bias (Chalmers et al., 2003).
- 2.
To counter techno-determinist narratives, manufacturers should offer devices excluding biased features, at an appropriately reduced price.
- 3.
Legislators should enshrine the right to opt out of surveillance technology, including wearables, particularly in the domains of private health insurance. Crucially, regulators, government, and wider society should critically evaluate the fairness of wearable-based surveillance practices.
- 4.
Regarding PPG-based HR monitoring, manufacturers should research alternative technologies, more inclusive of those with darker skin and higher BMIs.
- 5.
Regarding activity monitoring, more representative datasets should be constructed and made publicly available. Manufacturers should allow users to pick datasets which best represent them, and allow accuracy validation for the particular user, to expose user-level inaccuracies.
- 6.
Civil society should reflect on the desirability of health and fitness quantification, self-tracking, and social-network-based sharing, especially when it relies on technology which is inherently biased.
- 7.
Finally, manufacturers should collaborate with academia to research the harms of fitness trackers, whether psychological, social, or physical, and investigate their distribution across minority groups.
While this report has considered potential harms on groups subject to wearables’ bias, it has failed to hear directly from those groups. Instead, I have used normative ethical theory to suggest the possible interests of these groups. While I have challenged the impersonal, individualistic, and competitive narratives underpinning fitness wearables, greater effort should be made to account for the real interests of affected groups, and to consider alternative ethics, such as feminist ethics of care.
In addition to these limitations, the literature on which this report is based is sometimes of poor quality. For example, a researcher at the University of Oxford, regarding wearables, states: ‘it is essential to ensure that diverse participants are avoided’ (Radanliev, 2025, Section “How to address bias”). This is more likely the output of a Large Language Model, than a typo, as elsewhere the report includes several sentences apparently taken from an LLM prompt (see Appendix). There is also some disagreement between studies, such as on the impact of subjective sleep quality on cognitive performance, discussed above (c.f. Gavriloff et al., 2018; Zavecz et al., 2020). More broadly, this report cannot anticipate all harms arising from wearable fitness trackers, which remain under-researched (Folkvord et al., 2021).
In conclusion, the technologies underlying wearable fitness trackers are biased against multiple groups, including those with darker skin, higher BMIs, non-typical movement patterns, and disabilities. These biases risk exacerbating existing health inequalities, and restricting access to the espoused benefits of fitness wearables. Market forces not only lead to an additional, financial, dimension of harm against those for whom wearables are biased, but such surveillance practices also force adoption of these biased technologies. I presented 7 avenues for addressing this bias, as well as for addressing the ethical harms which are compounded by such bias.
Adapted from coursework submitted for the Algorithmic Bias, Fairness, and Justice MSc course at the Edinburgh Futures Institute, led by Zee Zalat.