In the traditional narrative of AI fairness, bias is a software problem. We imagine that if we could just "fix the labels" or "balance the dataset," our models would treat everyone equally. However, a growing body of research suggests that for many real-world applications—from medical diagnostics to wearable fitness trackers—bias is baked into the system long before a single line of code is written. It is embedded in the physical materiality of the sensors themselves.
The Physics of Signal Interference
The fundamental challenge is that sensors do not capture an objective reality; they capture a physical interaction. When a smartwatch measures heart rate or stress, it often relies on photoplethysmography (PPG), which shines light through the skin to detect blood flow. As Sofia Yfantidou and colleagues argue in (Un)fair devices: Moving beyond AI accuracy in personal sensing, this physical interaction is inherently demographic-dependent.
Higher levels of skin melanin can attenuate light waves, while variations in muscle mass and body composition can change how a device sits on the wrist, introducing motion artifacts. Adela C. Timmons and her team demonstrated this empirically in Biased Bots, finding that physical differences in body composition lead to unequal signal-to-noise ratios. For a machine learning model, this means the "ground truth" data for one demographic is crisp and clear, while for another, it is noisy and degraded. No amount of software-based label correction can recover information that was never captured by the hardware in the first place.
The Sensor Blind Spot in AI Safety
Current AI safety frameworks, such as the NIST AI RMF or the EU AI Act, tend to be "sensor-blind." They focus on data governance and algorithmic transparency but ignore the silicon. Matthew Stewart and Emanuel Moss point out in Materiality and Risk in the Age of Pervasive AI Sensors that social biases are often embedded directly into the calibration of imaging hardware. If a camera sensor is optimized for specific skin tones at the circuit level, the resulting bias persists throughout the entire AI pipeline.
This hardware-level discrepancy creates a "garbage in, gospel out" problem in critical care. Alex K. Pearce and Shamim Nemati highlight how pulse oximeters—the ubiquitous clips used to measure blood oxygen—are less accurate for patients with darker skin tones. When deep learning models are trained on this physiological data to predict respiratory failure, they inadvertently incorporate and scale these hardware inaccuracies, potentially leading to inequitable life-saving interventions (Can we predict the future of respiratory failure prediction?).
Hardware as a Hidden Variable
Beyond biological traits, the technical specifications of the hardware itself can introduce "technical noise" that models mistake for signal. In a study of smartphone sensors, Ramzi Halabi found that the noise profiles of Android and iOS sensors are so distinct that models can identify the operating system with 98% accuracy (Comparative assessment of multimodal sensor data quality). In clinical trials where participants use their own devices, a model might inadvertently learn to predict health outcomes based on the quality of the participant's phone rather than their actual physiology.
Similarly, in the field of eye-tracking, Bilal Taha showed that physical traits like iris color or the presence of eyewear introduce systematic measurement errors (Investigating Bias in Affective State Detection Using Eye Biometrics). These errors propagate downward, causing a 36% difference in performance for emotion recognition models across different ethnicities.
Toward Hardware-Centric Fairness
If the bias is physical, the solution must be physical as well. Researchers are now advocating for a "Fairness by Design" approach that moves the site of intervention from the model to the sensor. Alessandra Angelucci and Stefano Canali argue that device selection must become a primary determinant of systemic equity (The paradox of the artificial intelligence system development process).
This shift involves:
- Hardware Calibration as Mitigation: Treating the way sensors are tuned and calibrated as a primary tool for bias reduction (Stewart et al.).
- Weighted Variables: Adjusting clinical scores to account for known hardware limitations, such as transforming oxygen delivery from a binary "yes/no" to a weighted scale based on actual concentration ($FiO_2$), as demonstrated by Jonathan Clarke in Predictive value of the National Early Warning Score 2.
- Inclusive Hardware Standards: Moving away from "normative body" assumptions in hardware design—such as standard standing heights or limb configurations in XR systems—to ensure that the foundational data capture is inclusive (Daniel Killough et al., XR for All).
By recognizing that AI is a physical system, not just a mathematical one, researchers can begin to address the root causes of algorithmic unfairness where they often begin: at the point of contact between the sensor and the world.