A new human rights assessment reveals diverging patterns in how large language models are used globally, with surprising implications for freedom of thought and equality. Researchers at the UC Berkeley Human Rights Center interviewed 56 experts and practitioners across 24 countries, blending those insights with a review of existing LLM evaluations.
“What surprised us the most was how people in different parts of the world use LLMs so differently in their work, with practitioners in wealthier nations predominantly more risk averse,” noted Betsy Popken, executive director of the Human Rights Center and report author alongside Vyoma Raman and Camille Chabot. The assessment identifies risks to critical thinking and the spread of misinformation, alongside opportunities to improve access to justice and personalized education.
LLM Reliance Risks Stifling Critical Thinking & Higher-Level Thought
Stakeholder interviews and a literature review underpinned the UC Berkeley Human Rights Center’s assessment, applying the UN Guiding Principles for Business and Human Rights framework to evaluate LLM impacts across multiple sectors. Researchers identified a concerning trend of diminished critical thinking skills linked to overreliance on these models, a risk extending beyond specific professional fields into broader cognitive abilities. This finding emerged from interviews conducted with 56 individuals between 2023 and 2025, representing expertise and practical application across all continents and within law, journalism, and education.
Analysis of existing model evaluations further connected performance metrics to potential infringements on fundamental human rights. Practitioners demonstrated a markedly more cautious approach to LLM implementation compared to their counterparts, resulting in divergent patterns of risk and unexpected opportunities.
While this risk aversion mitigated certain harms, researchers also observed novel applications and benefits arising from more expansive LLM use in different socioeconomic contexts. The assessment revealed that LLMs, while potentially improving access to justice by aiding unrepresented litigants in small claims courts, also pose risks to due process and fair trial rights when relied upon in high-stakes legal settings. Providing inaccurate or incomplete advice could jeopardize a defendant’s right to remedy, highlighting the need for careful calibration of LLM assistance within the legal system.
Similarly, in journalism, LLMs offer productivity gains through writing and research support, but their propensity for generating false information contributes to the spread of misinformation, a critical concern for maintaining public trust. Researchers also found a complex interplay of positive and negative impacts on the right to education.
Personalized learning systems powered by LLMs can better address individual student needs, but challenges remain in aligning these models with established educational standards and ensuring they effectively support student recall. The report, titled details these findings and proposes a series of recommendations for LLM developers. These include diversifying datasets to reflect global communities more accurately, designing refusal mechanisms sensitive to the severity of potential human rights violations, and creating profession-specific LLMs tailored to address unique field challenges.
The team’s work builds on earlier research, including the publication of “Assessing Human Rights Risks in AI: A Framework for Model Evaluation” in arXiv and its presentation at the AAAI/ACM Conference on AI, Ethics, and Society in 2025. Co-author Vyoma Raman, in collaboration with Popken and Chabot, developed an approach for evaluating models to establish grounded claims about the level of risk they pose to specific human rights.
Popken and Raman co-authored “Building LLMs? Think Beyond Borders” with Tech Policy Press, emphasizing the importance of considering global implications during LLM development. Researchers advocate for establishing clear remediation processes to address harms resulting from LLMs, requiring comprehensive human rights risk assessments throughout the AI lifecycle, and implementing regulations targeting high-risk applications, such as judicial decision-making.
What surprised us the most was how people in different parts of the world use LLMs so differently in their work, with practitioners in Global North states predominantly much more risk adverse.
Betsy Popken, executive director of the Human Rights Center and author of the report alongside researche
See today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals.
