New contextual auditing method targets flaws in AI motion-capture

AI systems increasingly observe and make inferences about humans—from workplace safety cameras to fitness trackers—but auditing whether these systems actually work remains tricky. Traditional AI audits require knowing what inputs to test and what "correct" outputs should look like, which often isn't clear. In a new arXiv paper (2608.10194), a team led by Emma Harvey, Emanuel Moss, Hauke Sandhaus, Abigail Z. Jacobs, and Mona Sloane proposes a better approach: contextual auditing. This method evaluates measurements within the context of the practices that produce them, forcing auditors to explicitly define ground truth as verifiable real-world measurements. It also reveals hidden assumptions baked into the audit process itself.

To show how this works in practice, the researchers present a case study of skeleton inference in motion capture—a technology that estimates human body positions from sensor data. They borrow the concept of symmetry from science and technology studies, which allows audits to proceed even when ground truth is unknown, unknowable, or contested. By treating all perspectives symmetrically rather than privileging one "reference" output, auditors can assess system behavior more fairly. The paper, accepted at the AAAI/ACM Conference on AI, Ethics, and Society (AIES 2026), also highlights promising future audit targets for motion capture systems, including applications in animation, robotics, and clinical assessment.