Building LLMs? Think Beyond Borders

In early 2023, just following the public launch of ChatGPT, discussion surrounding the opportunities and risks of using LLMs across a variety of disciplines exploded. Legal professionals, educators, and journalists were among those beginning to explore how they could use LLMs in their work. With little practical research and few recommendations available across the fields of law, education, and journalism at the time, we began an analysis of the human rights risks and opportunities posed by the use of LLMs in these three fields. One of the biggest lessons we learned was the need to engage with users all around the world, because the way practitioners use LLMs in their work varies significantly based on where they live and work. While there were some exceptions, we found that practitioners in Global South states tended to use LLMs more ambitiously in their work, while those in Global North states were typically more risk-averse.

To learn more, we interviewed 56 experts and professionals across 24 countries in seven regions of the world to find out how they use LLMs in their work. We found that using LLMs for more complex, high-value tasks, as many did in the Global South, can expand human rights opportunities for people but can also heighten the likelihood and severity of risks people may endure. Where institutional capacity is already stretched thin, LLMs have started to fill gaps, substituting for lawyers, teachers, and journalists.

The critical gaps that LLMs are used to fill contribute to the generally positive attitudes toward LLMs expressed by the people we interviewed in South America, Sub-Saharan Africa, Asia, the Middle East, and North Africa, who describe LLMs as having positive effects in their communities. This contrasts with the experiences shared with us by many of their more risk-averse counterparts in Europe, North America, and Oceania. The optimism of experts and practitioners based in the Global South, however, is tempered: interviewees describe how uneven access to devices, training, and connectivity in their communities forcibly limits those who can use LLMs at all and how well the tools perform for required tasks. In this piece, we share stories told to us anonymously by people we interviewed in the Global South to illustrate how LLMs create opportunities, even as they introduce risks.

Law

In Singapore, LLMs are used in small claims tribunals to provide legal information to litigants. Because no lawyers are involved in such tribunals, the idea is to help civilians who are unclear about how to file legal documents. This means that a conversational chatbot provides access to this information without legal fees. This has helped democratize access to justice, providing more assistance to litigants than they may otherwise have been able to afford. On the flip side, if the chatbot provides incorrect information, or an attorney could have provided additional information to further help a claimant’s chances for success, this could risk impacting their right to access justice. Interviewees in both Singapore and Qatar described how LLMs are being used by legal professionals to translate interviews and other documents to or from English or between other languages. While this helps advance the speediness of accessing justice, it also risks harming it if gaps exist in the translation of the document because the LLM lacks certain language capabilities. It also requires that legal professionals double check the translation or risk inaccuracy.

Journalism

In South Africa, journalists use LLMs for research and writing stories that tend to be more formulaic, such as those based on press releases or sporting events. This means that journalists there can focus on tasks that require critical thinking and emotional intelligence, leaving repetitive tasks to machines. For instance, rather than spending time writing about the outcome of a cricket match, a journalist would have the time to research and write an in-depth feature of a female cricket team. Journalists also rely on LLMs for sports reporting in India, as well as weather reports and business stories on market shares. For example, instead of writing a weather report for the forthcoming week, a journalist could focus on the upcoming monsoon season, detailing its likely severity and steps people can take to stay safe. This makes it easier for journalistic organizations to provide access to both routine and analytical information to readers and listeners. However, it also means that if journalists or editors do not double check the content, inaccurate information could be provided to readers and listeners who then rely upon it for making decisions.

Education

In the United Arab Emirates, an interviewee described an AI tutor providing personalized learning to thousands of students. The tool supports the right to education for students who struggle with the language of instruction by translating lessons between English and Arabic, and for students with disabilities by transforming educational content to modalities like audio. In Mexico, educators with access and exposure to LLMs have used them innovatively in classrooms. In one instance, a teacher asked students to use ChatGPT to answer math and physics questions relevant to the curriculum and then required students to assess its responses for accuracy. Such efforts have sought to strengthen critical thinking in AI use, which is essential to protecting freedom of thought.

However, Mexico faces stark disparities in access and training, with schools in some regions having access to only a single computer for the entire building and some educators reporting that they have received no guidance about how to use LLMs in their work. While LLMs may provide opportunities to students, guidance needs to be provided to teachers and students about how to use LLMs effectively. When access is not available, such opportunities fail to reach some students, which risks harming those students’ right to education. Without clear guidance to students, they may rely more heavily on LLMs than on their own critical thinking, potentially impacting their freedom of thought.

- - -

If it occurs at all, most assessment of human rights risks by LLM companies happens at the deployment stage, where models are integrated into user-facing applications. And while these assessments sometimes focus on outputs in countries outside of the United States, no company we spoke with told us they had evaluated harms in numerous countries at once. In all of our interviews with LLM companies, representatives described engineering practices to evaluate and align models in accordance with values specified in their responsible AI policies—which aligned explicitly with human rights norms and instruments only in some cases. But model developers and deployers must do more.

Despite growing awareness of human rights risks posed by LLMs, most model-level evaluations remain focused on metrics for general performance, and sometimes specific risks like social biases, without explicitly linking those metrics to human rights outcomes. Model evaluations can be intentionally designed around human rights impacts. In a complementary piece focused on the use of LLMs in political news journalism, we propose and demonstrate a rights-aligned evaluation framework that draws on human rights instruments to center rights from the outset and link model behavior to real-world consequences. We recommend aligning evaluation tasks with specific human rights risks; developing metrics that capture the risks and their frequency; and interpreting results based on the seriousness, likelihood, and scope of the impact as well as the degree to which it is possible to remediate the impact.

Nonetheless, engineering is not enough. Model benchmarks do not capture the full context. Human rights stakeholder engagement is precisely designed to consider how geopolitics, resource and access disparities, and other factors shape how people use LLMs, and therefore the risks and opportunities they pose. LLM developers and deployers need to explore how people around the world use their products. Benchmarks measure what models can do in theory, but only people can describe their impacts in practice.