LLM use may threaten freedom of thought and equality across law, journalism and education

The Human Rights Center at UC Berkeley School of Law has conducted a comprehensive assessment of the potential human rights implications of large language models (LLMs) in the realms of law, journalism, and education. The study, titled "International Analysis of the Human Rights Impacts of Large Language Models," examines how the use of LLMs can create both risks and opportunities in these fields, taking into account varying geographic and socioeconomic contexts.

One of the most striking findings was the significant differences in LLM usage across regions, with practitioners in developed countries tending to be more cautious about potential risks. The researchers interviewed experts and practitioners from 24 countries, as well as representatives from companies developing LLMs, to gain a holistic view of the issues at hand.

The assessment identified several cross-cutting risks, including the risk of stifling critical thinking by overreliance on LLMs, as well as cultural and ideological homogenization challenges that could undermine efforts to promote equality and nondiscrimination. While LLMs hold promise for improving access to justice, especially in resource-poor settings, their use can also pose risks to due process and fair trial rights through the provision of inaccurate or unbalanced legal advice.

In journalism, LLMs can enhance productivity and working conditions, but the potential for misinformation due to hallucinations remains a concern. In education, personalized learning systems can help educators address individual student needs, but aligning LLMs with educational standards and facilitating student recall pose ongoing challenges.

To mitigate these risks and capitalize on opportunities, the researchers recommend several measures, such as diversifying training data to better reflect global communities, incorporating refusal mechanisms and response tones that reflect the severity of identified human rights risks, and developing profession-specific LLMs tailored to the unique challenges of each field.

Additionally, they advocate for establishing clear remediation processes to address harms caused by LLMs, conducting comprehensive human rights risk assessments throughout the AI development lifecycle, and implementing specific regulations targeting high-risk use cases, such as judges using LLMs to make legal decisions.