Berkeley study maps LLM human-rights risks across law, media and schools

UC Berkeley’s Human Rights Center has warned that large language models can expand access to legal services, information and personalised education while threatening freedom of thought, equality, privacy and access to reliable information. Its new report examines the effects of LLM use by lawyers, journalists and educators on clients, audiences and students. The research, conducted from 2023 to 2025, drew on more than 15 months of work and interviews with 56 people in 24 countries, including representatives of companies developing LLMs. Researchers spoke to at least one expert and one practitioner in each of the three fields on every inhabited continent. The study combined a literature review, global AI regulations and guidance, stakeholder interviews, sector-specific model evaluations and analysis of human-rights impact assessments. It used the due-diligence framework associated with the UN Guiding Principles on Business and Human Rights and sought to link model-performance measures to downstream rights impacts. The project was prompted by ChatGPT’s public launch in late 2022, as researchers sought to examine potential gains alongside risks. Across all three fields, the report found that LLMs can speed up work and lower costs. In law, the technology can help people understand or translate technical documents and assist unrepresented litigants, including in small-claims cases. The study cited a Singapore service that helped people prepare small-claims-court documents free of charge, and legal professionals in Qatar who used LLMs to translate documents more quickly. But the report cautioned that incomplete, inaccurate or insufficiently nuanced output in high-risk legal settings could endanger due-process, fair-trial and remedy rights. It also identified privacy as a particular concern when lawyers, litigants or judges enter confidential case material and personal information into LLM systems. For journalists working with limited resources, LLM-assisted writing and research can improve productivity and working conditions. The researchers pointed to journalists in South Africa and India using the tools to produce information on sporting results more quickly. Yet hallucinated material can fuel misinformation and impair the public’s access to information. The findings give a human-rights dimension to the cautious, piecemeal adoption of AI in Asian public-interest newsrooms that we reported in August. More broadly, the Berkeley study found that professionals in the Global North were generally more risk-averse, while participants in parts of Africa and Asia more often stressed lower costs and wider access to information. In education, personalised systems could help teachers respond to individual students’ needs. An AI tutor in the United Arab Emirates was used by hundreds of students, including non-native English speakers and learners with disabilities, while educators in Singapore used LLMs to reduce the time needed to assess student work. The report nevertheless found that LLMs are not generally aligned with varied educational standards and may struggle to support students’ recall. Across the three sectors, excessive reliance on the systems could weaken higher-level critical thinking and threaten freedom of thought. The researchers also warned that training data drawn mainly from dominant, readily available sources can marginalise other worldviews through cultural and ideological homogenisation, creating equality and nondiscrimination risks. They said differences in training, resources, geography, language and access shape both the opportunities and the harms of the technology. UNESCO has separately warned that unequal access can deepen those divides: about 2.6 billion people, nearly one-third of the world’s population, lacked internet access in 2024. It has identified increased access and personalised learning as potential gains from digital technology and generative AI, while cautioning that privacy, safety, equity and governance problems can threaten the right to education and other rights. Berkeley recommended comprehensive human-rights risk assessments throughout the AI lifecycle for both LLM developers and organisations deploying them. It also called for clear processes to identify harm, assign responsibility and provide remedies. Rather than banning LLMs outright, the report advocated regulation focused on specific high-risk uses. Its other recommendations include more globally representative datasets, risk-calibrated refusal mechanisms and response styles, profession-specific models, and stronger AI literacy among users, developers, deployers, industry groups and policymakers. The report follows the center’s 2025 paper, “Assessing Human Rights Risks in AI: A Framework for Model Evaluation,” which translated human-rights concepts into measures including scope, scale and likelihood.