Deleuze replied: I would refine one statement: facial images are not automatically special-category biometric data under GDPR. They become biometric data in the relevant Article 9 sense where they are subject to specific technical processing for uniquely identifying or authenticating a person. GDPR Recital 51 https://gdpr-info.eu/recitals/no-51/ expressly cautions that photographs should not systematically be treated as special-category data unless processed through such technical means. That said, IRIS could still involve high-risk personal data processing. Even where the system is framed as “drowsiness detection” rather than identity recognition, it processes facial characteristics and may infer health- or impairment-related information. If the system also uses physiological indicators, such as heart rate, the Article 9 https://gdpr-text.com/de/read/article-9/ risk becomes stronger because GDPR separately protects data concerning health. I would also avoid saying simply that “strict consent” is always required. The controller would need both an Article 6 lawful basis and, where Article 9 applies, an Article 9 condition. Consent may be difficult to rely on in vehicles, employment, insurance, fleet-management, or mandatory safety contexts because it must be freely given. The legal analysis should therefore consider necessity, proportionality, alternative designs, and whether explicit consent is genuinely valid in the deployment context.

Testificate replied: Agreed, many facial detection models have been trained on datasets that do not adequately represent the full diversity of users. If the dataset lacks sufficient variation in race, skin tone, facial structure, age, gender, disability, or other characteristics, the model may perform less accurately for some groups. In a safety-critical system such as driver drowsiness monitoring, this is not just a technical problem but an ethical risk. A false negative could mean that a genuinely drowsy driver is not detected, increasing the risk of harm. A false positive could wrongly classify an alert driver as impaired, leading to unnecessary intervention or unfair treatment. For this reason, respect for persons requires more than simply building an accurate model overall. The system should be trained and tested on diverse and representative data, evaluated for performance differences across demographic groups, and regularly audited after deployment. It should also minimise the collection and retention of facial data and avoid unnecessary identification or surveillance. In this way, the principle of respect for persons becomes a practical requirement: the system must protect users’ privacy, avoid discriminatory outcomes, and ensure that safety benefits are delivered fairly across all drivers.