Responsible AI for Detecting Dark Patterns: Linking Calibration With Risk Matrix

Dark patterns—deceptive interface designs—undermine consumer autonomy and create systemic risks in digital markets. While technical research has focused on detection accuracy, our study highlights probability calibration as a foundation for responsible AI. Calibration ensures that model confidence aligns with real outcomes, producing trustworthy likelihood estimates that can be integrated into established governance tools such as ‘ISO 31000: 2018 Risk matrix’. Using a corpus of real-world user interface screenshots, we compare image-based and OCR-based detection routes and demonstrate how calibration quality directly affects risk assessment and regulatory triage. Poorly calibrated models inflate ‘very-high-risk’ flags, distorting supervisory priorities, whereas calibration-aware pipelines provide stable thresholds and auditable evidence. We argue that calibration-aware AI offers not only technical reliability but also socio-technical accountability, enabling regulators, platforms, and consumer advocates to ground decisions in evidence that is measurable, traceable, and defensible.