When ChatGPT launched in early 2023, professionals across law, education, and journalism quickly began experimenting with large language models. Yet little practical guidance existed on how to use these tools responsibly. Researchers set out to understand the human rights implications by speaking with 56 practitioners across 24 countries spanning seven regions. A striking pattern emerged: those working in the Global South deployed LLMs for complex, high-stakes tasks at far greater scale than their counterparts in wealthy nations, who tended toward caution.
The difference reflects structural realities on the ground. In regions where institutional capacity is already strained, LLMs have begun filling critical gaps—substituting for lawyers, teachers, and journalists who simply are not available. This necessity drives optimism among practitioners in South America, Sub-Saharan Africa, Asia, the Middle East, and North Africa. Yet that optimism is tempered by harsh constraints: uneven access to devices, training, and reliable connectivity means that only some people can use these tools at all, and performance varies widely depending on local conditions.
Law
In Singapore, LLMs power chatbots in small claims tribunals, offering legal information to people without lawyers. This expands access to justice for those who cannot afford counsel. Yet if the system provides wrong guidance or misses details a human lawyer would catch, it can undermine the very right it aims to protect. Across Singapore and Qatar, legal professionals use LLMs to translate documents and interviews between English and other languages, speeding up case processing. But translation errors can slip through if practitioners do not verify the output, potentially damaging clients' interests.
Journalism
In South Africa, newsrooms deploy LLMs to handle formulaic stories drawn from press releases and sports coverage, freeing journalists to pursue investigative work that demands critical thinking. Rather than writing routine match reports, a journalist might spend time profiling a women's cricket team. Similar patterns appear in India, where LLMs handle sports reporting, weather forecasts, and business stories on market movements. A journalist might use the system to draft a standard weather forecast, then invest energy in analyzing an approaching monsoon season and advising readers on safety. This approach lets news organizations deliver both routine and analytical content. The danger lies in publishing unverified LLM output—readers may act on inaccurate information without realizing it came from an automated source.
Education
The United Arab Emirates deployed an AI tutor reaching thousands of students, translating lessons between English and Arabic for language learners and converting content into audio for students with disabilities. In Mexico, innovative teachers have asked students to use ChatGPT to tackle math and physics problems, then evaluate the answers for correctness—an exercise designed to build critical thinking about AI itself. Yet Mexico illustrates the access problem starkly: some schools have only one computer for the entire building, and many educators have received no training on LLM use. Without access and guidance, students miss opportunities, and those who do gain access may lean on LLMs instead of developing their own reasoning.
The Assessment Gap
When LLM companies evaluate human rights risks at all, they typically do so only when deploying models into applications. Even then, no company interviewed reported assessing harms across multiple countries simultaneously. Model developers described engineering practices aligned with responsible AI policies, though these policies did not always explicitly reference human rights frameworks.
Most model-level testing focuses on general performance metrics and sometimes specific risks like social bias, without explicitly connecting those measures to human rights outcomes. Evaluation can be redesigned around rights impacts. A complementary study on LLMs in political news journalism proposes a rights-aligned framework that anchors assessment in human rights instruments from the start and traces model behavior to real-world consequences. The approach calls for aligning evaluation tasks with specific human rights risks, developing metrics that capture both the risks and their frequency, and interpreting results based on seriousness, likelihood, scope, and the possibility of remediation.
Yet engineering alone cannot solve the problem. Benchmarks measure what models can do in theory, not how they perform in practice across different contexts. Meaningful human rights engagement requires understanding how geopolitics, resource disparities, and access gaps shape the way people actually use LLMs. LLM developers and deployers must listen to practitioners worldwide. Only people working in the field can describe what these systems do when deployed in their communities.
Source: Tech Policy Press
