Berkeley report maps human rights risks of LLMs in three fields

Researchers at UC Berkeley's Human Rights Center have published "An International Analysis of the Human Rights Impacts of Large Language Models," a study of how generative AI is reshaping law, journalism and education, the authors said in an episode of the Berkeley Law Voices Carry podcast published September 28, 2026.

What the study did

The report was written by Human Rights Center Executive Director Betsy Popken and two former center researchers, Vyoma Raman, now pursuing a PhD in information science at Cornell, and Camille Chabot, now a JD student at Harvard Law School.

According to host Gwyneth Shaw, the team spent 15 months interviewing 56 people from 24 countries, including at least one expert and one practitioner in each of the three sectors on each inhabited continent.

Popken said the project deliberately combined human rights and computer science methods, and that the report includes evaluations of models tied to specific risks. Raman said a companion paper on operationalizing human rights language into evaluation metrics was published through the AI Ethics and Society Conference run by the Association for the Advancement of Artificial Intelligence.

Risks and opportunities the authors identified

Chabot said the team assessed risks and opportunities together, measuring the scope, scale and remediability of each. For law, she said, the most salient risk was privacy, arising when lawyers, litigants or judges entered confidential case information, including personal information about parties, into the tools.

For education and journalism, the authors identified freedom of thought as the most salient risk, citing student over-reliance on models, educators using them to generate feedback, and time-pressed journalists leaning on them to write, source and fact-check articles.

Popken said the same rights appeared as both risk and opportunity. She cited an AI tutor used in the UAE that reached hundreds of students, including non-native English speakers and students with learning disabilities, but said the tool could fail the right to education if it is not aligned with educational or testing goals. Other examples the authors gave included journalists in South Africa and India writing sports results, educators in Singapore assessing student work, legal professionals in Qatar translating legal documents, and a Singapore system helping complainants file small claims court documents at no cost.

A global north and global south split

Chabot said the team observed more risk aversion among professionals interviewed in the global north, who emphasized human rights risks, and more enthusiasm in the global south, where interviewees in Africa and Asia focused on cost savings and increased access to information.

Raman said the team debated grouping people by those labels, acknowledging that the terms are not universally used and have drawn valid critiques, but said the split reflected differences in available resources and alternatives to language models.

Nine recommendations, six stakeholder groups

The report makes nine recommendations addressed to six stakeholder groups: resource providers and model engineers under LLM developers, and distributors, operators, industry groups and government under LLM deployers.

Cross-cutting recommendations include assessing the human rights impacts of LLMs, defining remediation stages and cultivating critical AI literacy. Others are assigned to specific actors: regulating general purpose AI to government policymakers, setting guidelines for professional use to operators such as law firms and industry groups such as the International Bar Association, creating user feedback mechanisms to distributors and operators, and designing LLM interfaces responsibly to distributors.

Raman said the team is asking companies that already run large-scale capability evaluations to extend them to human rights risks and opportunities, and recommends international assessments that account for differing levels of resource access. Popken said the team plans webinars with the Global Network Initiative and one hosted by Dropbox, and intends to expand engagement to industry groups and government.

What to do

• Assess the human rights impacts of large language models, whether at large or small scale, across both development and deployment.
• Extend existing model evaluations beyond capability benchmarks to cover human rights risks and opportunities, with developers examining training data and labeling practices upstream and deployers evaluating real-world task contexts.
• Run international assessments that account for how risks and opportunities differ by geography, language and resource access.
• Define remediation stages so problems that emerge can be fixed and risks mitigated.
• Cultivate critical AI literacy so students and professionals learn to perform and judge tasks themselves rather than replacing that skill development with AI.
• Have governments regulate general purpose AI, while operators such as law firms and industry groups such as the International Bar Association set guidelines for professional use.
• Build user feedback mechanisms (distributors and operators) and design LLM interfaces responsibly (distributors).
• Avoid entering confidential case or personal information into LLMs in legal work, the privacy risk interviewees raised most often.

Read the original from UC Berkeley →