An International Analysis of the Human Rights Impacts of Large Language Models

Summary

This report examines how the use of LLMs by legal professionals, journalists, and educators affects people’s human rights, including the rights of clients, audiences, and students. To guide this assessment, we ask three central questions:

• How might the use of LLMs in law, journalism, and education negatively and positively impact human rights?
• How do the human rights risks and opportunities associated with the use of LLMs vary based on the user’s location in the world and the social and economic conditions in which people in their communities live and work?
• What measures can be taken, and by which entities, to effectively mitigate the human rights risks and leverage the opportunities presented by LLMs?

Methods: The assessment follows a mixed-methods approach. We conducted a literature review of the impacts LLMs have on human rights, and their specific impacts in the fields of law, journalism, and education. We analyzed regulations and guidance on artificial intelligence by states across the world. Drawing on the human rights impact assessment framework used as a due diligence tool under the UN Guiding Principles for Business and Human Rights (UNGPs), we blended stakeholder interviews with a literature review and our own analysis to identify rights-impacting trends, allocating scores to each human rights risk based on its saliency. Although not required in human rights impact assessments, we also included human rights opportunities we identified that resulted from peoples’ use of LLMs in their work. We interviewed 56 people from 24 countries, including at least one expert and one practitioner from each sector on each continent, and representatives from corporations developing LLMs. We also analyzed existing sector-specific model evaluations to connect performance metrics to human rights impacts.

Findings: In the legal field, LLMs can improve access to justice by assisting litigants who do not have counsel, such as in small claims settings. However, reliance on LLMs by legal practitioners in high-risk settings can risk violating clients’ due process and fair trial rights, as well as the right to remedy, if the responses generated by LLMs are incomplete or wrong. The use of LLMs by journalists can affect peoples’ right of access to information. While providing assistance with writing and research can improve the productivity and efficiency of journalists operating in resource-constrained settings, LLM hallucinations can also lead to the spread of misinformation. In education, personalized learning systems make it easier for educators to adequately assist students with individual challenges, but LLMs are not typically aligned with various educational standards. Across sectors, freedom of thought is at risk because peoples’ overreliance on LLMs can stifle higher-level critical thinking. The rights to equality and nondiscrimination are also at risk. Because the data that models are trained on mostly originates from dominant, widely available sources, LLMs can promote cultural and ideological homogenization, privileging certain worldviews while marginalizing others.

Recommendations:To effectively mitigate the human rights risks posed by LLMs, different actors must jointly implement measures. Key recommendations include establishing clear remediation processes to identify, assign responsibility for, and remedy harms; requiring comprehensive human rights risk assessments across the AI lifecycle; and creating regulations that target specific high-risk use cases rather than banning LLMs altogether. We also recommend diversifying datasets to better represent global communities, designing refusal mechanisms and response tones based on the severity of human rights risks, and developing profession-specific LLMs to address field-specific challenges. Finally, LLM developers and deployers as well as industry groups and policymakers should collectively support the development of users’ AI literacy.