Explainable AI is essential for the deployment of human activity recognition systems

Thank you. Here is my opening statement.

Human activity recognition systems monitor intimate aspects of human behavior — from healthcare patients' daily movements to employees' workplace activities — making transparency not merely desirable but fundamentally essential. When these systems misclassify a fall as normal movement in elderly care, or flag an innocent gesture as suspicious in surveillance contexts, stakeholders must understand *why* the error occurred to prevent harm, assign accountability, and correct the system. Regulatory frameworks like the EU AI Act and GDPR's "right to explanation" are already codifying this principle, meaning opaque HAR systems face genuine legal barriers to deployment regardless of their accuracy. Beyond compliance, explainability builds the trust necessary for adoption: studies consistently show that end-users, clinicians, and institutional decision-makers reject black-box systems that affect human welfare, particularly when those systems operate on sensitive behavioral data. Without explainability, we deploy systems that are unauditable, legally precarious, and fundamentally incompatible with the human dignity of those being monitored.

Your performance-versus-explainability framing presents a false dichotomy: modern XAI techniques like SHAP, attention mechanisms, and concept-based explanations can be layered onto deep learning HAR systems with negligible accuracy loss, so we don't actually have to choose between a 99.8% black box and a 94% interpretable model. The aviation analogy actually *supports* my position — autopilot systems are among the most extensively explained and auditable technologies in existence, with every decision logged, traceable, and subject to post-incident reconstruction through flight data recorders, which is precisely the kind of explainability HAR systems need. Your claim that regulations only require "documented validation" misreads the EU AI Act, which explicitly classifies biometric and behavior-monitoring systems as high-risk and mandates that they be "sufficiently transparent to enable users to interpret the system's output and use it appropriately" (Article 13) — this is not satisfiable by testing protocols alone. The perverse incentive you identify — misleading post-hoc explanations — is an argument for *better* explainability standards, not for abandoning explainability entirely, just as the existence of fraudulent safety testing doesn't argue against safety testing. When your 99.8%-accurate black box systematically fails on wheelchair users, specific ethnic groups, or atypical movement patterns — as HAR systems demonstrably do — only explainability reveals *why*, enabling targeted correction rather than blind retraining that may perpetuate the very biases harming the populations these systems claim to protect.

Throughout this debate, my opponent has progressively conceded my core claims while reframing them as their own position — acknowledging that XAI techniques impose negligible accuracy costs, that aviation systems require extensive auditability, and that bias detection demands investigation into *why* systems fail — yet still resists calling explainability essential, which is precisely what these concessions establish. The "deploy first, explain later" approach is unconscionable for HAR systems monitoring vulnerable populations: we don't release pharmaceuticals and investigate adverse effects afterward, we require mechanistic understanding *before* deployment, and systems continuously monitoring human behavior in healthcare, eldercare, and workplaces deserve equivalent rigor. My opponent's claim that demographic auditing alone catches disparate impact without explainability is empirically wrong — you can *detect* that your system fails on wheelchair users through auditing, but without explainability you cannot determine whether the failure stems from training data gaps, sensor positioning assumptions, or feature extraction biases, leaving you unable to implement targeted fixes rather than expensive, blind retraining cycles. The "sufficiently transparent" standard in Article 13 cannot be met by accuracy metrics alone precisely because transparency is about enabling users — caregivers, patients, oversight boards — to understand and appropriately rely on system outputs, which opaque predictions fundamentally cannot provide regardless of aggregate performance. Explainability is not the enemy of performance, not a luxury for low-stakes applications, and not a post-deployment afterthought — it is the mechanism through which we ensure accountability, enable correction, maintain legal compliance, and preserve the dignity of every person whose intimate daily activities these systems presume to judge.