A/HRC/56/68 6,000 languages in the world have high-quality data resources that can be used to train artificial intelligence models. To address that gap, companies have begun to develop multilingual language models. However, multilingual models do not perform as well as English-language models. The use of large language models in educational settings could disadvantage students from linguistic backgrounds that are not represented in the underlying data resources, which may have racially disproportionate impacts.71 48. There are debates about whether generative artificial intelligence tools based on large language models should be banned among students rather than integrated into curricula. There are also steps in some educational settings to try to restrict the use of generative artificial intelligence tools that rely on large language models among students. Some educational institutions are using artificial intelligence tools to detect the use of artificial intelligence by students. The use of such tools, which may contain algorithmic bias, to patrol cheating may introduce further biases that harm students from marginalized racial and ethnic groups. Such harm is bound to be exacerbated in cases in which institutions have not set up equitable appeals processes.72 (d) Facial recognition in educational institutions 49. Facial recognition technologies have been introduced in many educational settings around the world, despite evidence of racial bias in their operation, as described above. Facial recognition systems are being used to automate attendance-taking, to enhance school security, to perform examination proctoring functions and even to record the emotions of children in schools to monitor how much they are learning. This is often without adequate human rights due diligence or regulatory oversight. For example, in Brazil, an increasing number of schools are adopting facial recognition tools to streamline operations, track attendance and enhance security.73 However, it has been reported that neither municipalities nor states conducted human rights impact assessment studies or analysed the risks of discrimination associated with facial recognition software before implementing these projects.74 50. The use of facial recognition software in educational settings is having racially discriminatory impacts. There have been cases, including one reported in the Kingdom of the Netherlands, in which students of African descent have had to shine lights in their faces to be recognized by the artificial intelligence systems used to mediate access to important examinations. Such experiences impact students’ equal right to education but also create friction and exclusion when students from marginalized racial and ethnic groups are given the impression that the system was not designed for them. The recording and monitoring of children’s emotions in schools has significant privacy implications for all students and can perpetuate racial bias. These systems have been found to interpret the facial expressions of individuals of African descent and white individuals differently, attributing negative feelings, such as contempt and anger, more frequently to those of African descent. 75 71 72 73 74 75 GE.24-08849 Felix Richter, “The most spoken languages: on the Internet and in real life”, Statista, 21 February 2024; Emily M. Bender, “The #BenderRule: on naming the languages we study and why it matters”, The Gradient, 14 September 2019; Gabriel Nicholas and Aliya Bhatia, “Lost in translation: large language models in non-English content analysis”, Center for Democracy and Technology, 23 May 2023; A. Bergman and Mona Diab, “Towards responsible natural language annotation for the varieties of Arabic”, in The 60th Annual Meeting of the Association for Computational Linguistics: Findings of ACL 2022 (Association for Computational Linguistics, 2022); and BigScience Workshop, “A 176B-parameter open-access multilingual language model” (ArXiv, 2022). See Regina Ta and Darrell M. West, “Should schools ban or integrate generative AI in the classroom?”, Brookings Institution, 7 August 2023; and Robert Topinka, “The software says my student cheated using AI. They say they’re innocent. Who do I believe?”, The Guardian, 13 February 2024. InternetLab submission. Ibid. Ibid. 13