A/HRC/44/24 profile, communicate only via secure messaging apps, delete protest-related conversations and only use prepaid SIM cards.67 30. Another development that is particularly problematic is the practice of routinely making audiovisual recordings of assembly participants, often in combination with the deployment of facial recognition technology. Such automated technology relies on the comparison of the digital representation of a face captured in a digital image – known as a “template” – with other templates in a database. From the comparison, a higher or lower probability is deduced that the person is indeed the person to be authenticated or identified. The probability threshold to consider two or more templates a match is determined by the users of the system. When footage obtained from closed circuit television video cameras deployed in public spaces is compared in near real-time with images in databases, this is referred to as live facial recognition technology. 31. The use of facial recognition technology brings about significant risks for the enjoyment of human rights, including the right of peaceful assembly. Despite remarkable accuracy gains in recent years, this technology is still prone to errors. For example, an image may be falsely considered a match (known as a “false positive”), with significant consequences to a person���s rights, including in cases where a person is wrongly flagged as a suspect of a crime and may be detained and prosecuted. When facial recognition technology is used on a large number of people, even low rates of error may result in the inaccurate flagging of hundreds of individuals. 32. Moreover, facial recognition technology may perpetuate and amplify discrimination, including against Afrodescendants and other minorities, women or persons with disabilities, 68 because it can be used to profile individuals on the basis of their ethnicity, race, national origin, gender and other characteristics. This technology may also lead to unintended discrimination in light of the fact that its accuracy depends on factors such as skin colour or gender, and experience has shown lower accuracy rates for the recognition of dark-skinned persons and women.69 33. The use of facial recognition technology to identify persons in the context of assemblies has considerable adverse effects on the rights to privacy, freedom of expression and peaceful assembly, if effective safeguards are not in place. A person’s image constitutes one of the key attributes of her or his personality as it reveals unique characteristics distinguishing her or him from other persons. 70 Recording, analysing and retaining someone’s facial images without her or his consent constitute interferences with a person’s right to privacy. By deploying facial recognition technology at assemblies, these interferences occur on a mass and indiscriminate scale, as this requires the collection and processing of facial images of all persons captured by the camera equipped with or connected to a facial recognition technology system. 34. Assemblies traditionally have allowed participants a certain level of protection against being singled out or identified. 71 This protection was already considerably weakened by many States that routinely made audiovisual recordings of assembly participants. 72 The rise of facial recognition technology has led to a paradigm shift in comparison with practices of audiovisual recordings, as it dramatically increases the capacity to identify all or many participants in an assembly in an automated fashion. This is particularly problematic if live facial recognition technology is deployed, permitting real-time identification as well as 67 68 69 70 71 72 Contributions of Access Now and Amnesty International. Contributions of Amnesty International and the Digital Rights Foundation. See Joy Buolamwini and Timnit Gebru, “Gender shades: intersectional accuracy disparities in commercial gender classification”, Proceedings of Machine Learning Research, vol. 81 (2018), pp. 1–15; and Inioluwa Deborah Raji and Joy Buolamwini, “Actionable auditing: investigating the impact of publicly naming biased performance results of commercial AI products”, Conference on Artificial Intelligence, Ethics and Society (2019). European Court of Human Rights, Reklos and Davourlis v. Greece (application No. 1234/05), judgment of 15 April 2009, para. 40. OSCE/ODIHR, Guidelines on Freedom of Peaceful Assembly, 3rd ed. (2019), para. 71. Ibid., para. 172. 9