The Human Rights Center at UC Berkeley published An International Analysis of the Human Rights Impacts of Large Language Models: In Law, Journalism, and Education (August 2026), drawing on 56 interviews with legal professionals, journalists and educators across 24 countries, algorithmic audits evaluating LLM performance in these three fields, and a human rights impact assessment framework grounded in the UN Guiding Principles on Business and Human Rights.
Human Level’s Take:
- When AI enters the workplace, two human rights conversations begin to run in parallel. One is about the impacts that follow from introducing AI into how people work. The other is about the risks that AI systems themselves introduce or amplify as they are deployed. Together, they raise a foundational question for businesses integrating AI: which risks need to be on our radar, and what actual and potential human rights impacts will we need to identify and address as these systems become embedded in the way we work?
- One thing the Human Rights Center’s analysis brings into view is just how broad that lens may need to be. The social impacts of large language models stretch across their lifecycle: from the data, labor and natural resources that support their development, to the biases and harmful outputs that can surface in their use, as well as the ways these technologies can support communities.
- There is, however, already a familiar framework for working through many of these questions: human rights due diligence. Applying an HRIA to LLMs illustrates how established human rights considerations, like privacy, non-discrimination, freedom of thought and access to remedy, can take on new dimensions when AI is introduced into professional settings. It also highlights the importance of context: risks can look quite different depending on where, how and by whom a system is used.
- For companies, this suggests that some of the most useful questions may be familiar ones: where could impacts arise, who could be affected, and what mechanisms are available to prevent or address harm? Asking those questions shifts the focus from AI in the abstract to how particular systems and uses affect people in practice.
- That can widen the set of issues companies need to consider as they integrate AI into work: looking beyond whether AI replaces jobs to how it changes hiring, entry-level opportunities, workloads and expectations; considering whether workers and users have meaningful feedback and grievance channels; and examining whether general-purpose tools are appropriate for specialised or high-stakes uses.
Some key takeaways:
- Human rights impacts emerge across the LLM lifecycle: The Human Rights Center finds that the human rights risks of large language models begin at their technical foundations. LLMs are trained on vast, web-scraped datasets and remain largely "black boxes," making it difficult to fully control or explain their outputs. This structural limitation underlies risks such as hallucination, bias and the generation of harmful content. The report also traces human rights impacts to how LLMs are built, including the labour and environmental resources on which they depend. It notes that LLM development can reproduce existing patterns of unequal power, citing 2021 reporting that OpenAI outsourced content-labeling work to Kenyan workers paid between $1.32 and $2 per hour. Training and running LLMs also carries a significant environmental cost, with the Human Rights Center noting that training a single, comparatively small BERT model consumes energy on the scale of a trans-American flight. These risks can carry through into model outputs and real-world applications: because LLMs are trained on datasets containing existing societal biases, the report finds that they can reproduce discriminatory associations, including caste bias toward Dalit job applicants and harsher treatment of speakers of African American Vernacular English than speakers of Standard American English in criminal decision-making contexts. At the same time, the report emphasizes that these technologies can also support social empowerment when communities have meaningful control over their development and use, pointing to Māori-run media organization Te Hiku Media’s use of language technologies to help preserve and revitalize the Māori language as an example of communities exercising data sovereignty in how technology is applied to their own context.
- Applying an Human Rights Impact Assessment to AI systems reveals context-specific and cross-cutting risks: The Human Rights Center’s assessment of LLM use in law, journalism, and education shows how evaluating scope, scale, remediability, and likelihood can help identify where human rights risks are most salient and where prevention and remedy may be needed. Applying this framework, the Human Rights Center finds that the right to privacy is most salient in law due to its high rating on both scope and scale because lawyers and litigants may input confidential case information into LLMs, with low remediability once that information is exposed. Across all three professions, the report identifies freedom of thought as the most salient cross-cutting risk, noting that overreliance on LLMs can introduce inaccuracies and bias that are difficult to detect, and that this risk may compound over time as reliance shifts from a choice to an operational obligation. Importantly for practitioners assessing AI use, the Human Rights Center also finds that risks and opportunities can stem from the same underlying application: the efficiency that can enhance access to information, for example, becomes a risk when hallucinations occur, making impacts difficult to predict in advance. This, according to the study, reinforces the importance of ongoing assessment rather than treating an HRIA as a one-time exercise. For remediation, the report recommends a three-stage process across actors in the AI lifecycle: identify harms as early as possible through development-stage evaluation, deployment-stage monitoring, and user feedback channels; determine responsibility according to whether an actor caused, contributed to, or is directly linked to a harm, in line with the UN Guiding Principles; and implement remedies, including restitution, compensation, rehabilitation, satisfaction, or guarantees of non-repetition, that are timely, proportionate, and delivered by the actor best positioned to act.
- Assessing LLMs through a human rights lens brings risks into existing due diligence processes: For companies and practitioners integrating LLMs into their work, the report's central recommendation is to conduct comprehensive human rights risk assessments across the AI lifecycle, scaled to the size and role of the organization, and to treat this as an extension of existing corporate due diligence obligations under the UN Guiding Principles. On labour practices specifically, the report finds that while direct job losses in law, journalism and education have so far been limited in scope, LLMs are changing how work gets done and who gets hired: employers are sometimes forgoing new hires or not replacing retiring staff, and reduced availability of routine tasks is narrowing the entry-level opportunities junior employees have traditionally used to build foundational skills. Where LLMs augment rather than replace workers, the report notes mixed effects on well-being: some practitioners report reduced workloads and more time for higher-value work, while others describe rising productivity expectations. Looking ahead, the report calls on companies to establish clear user feedback and grievance mechanisms, design refusal mechanisms and response tone calibrated to the severity of the human rights risk at hand, and, where feasible, develop or adopt profession-specific tools rather than relying on general-purpose LLMs for specialized, high-stakes tasks such as legal research or election reporting.
