Researchers at Berkeley’s Human Rights Center completed an investigation lasting over 15 months into the effects of large language models on human rights, interviewing 56 people from 24 countries to gather input from legal, journalistic and educational sectors on every inhabited continent. The team, led by Human Rights Center Executive Director Betsy Popken and including Vyoma Raman, now at Cornell pursuing a PhD in information science, and Camille Chabot, beginning her JD at Harvard Law School, examined both the benefits and drawbacks of generative AI.
Popken explains the study arose from considering potential risks alongside opportunities when ChatGPT launched in late 2022, asking, The resulting report offers a robust set of recommendations to identify and address potential harms in these critical fields.
Berkeley Law Center Examines LLM Impacts Across Global Sectors
The Berkeley Law report focuses on how the use of large language models are affecting the fields of law, journalism and education. The results provide insight into the benefits and drawbacks of this technology in these sectors, and the authors conclude with a robust set of recommendations aimed at identifying and remediating potential harms in these important fields.
Vyoma Raman highlighted a common thread observed across journalism, law and education sectors. “One was the use of large language models to expedite tasks,” she explained, citing examples of journalists in South Africa and India using LLMs to rapidly generate reports on sporting event outcomes.
While increasing efficiency, this reliance on automation raises concerns about the erosion of critical thinking and thoroughness, particularly in fields demanding accuracy and detailed understanding. The report emphasizes that LLMs enable shortcuts, bypassing essential steps in research, organization and critical analysis.
This circumvention of established processes poses a significant risk, particularly for developing minds. “I think this was the biggest one that really stood out to us as one that would be significantly at risk across all three sectors,” stated Betsy Popken, Executive Director of the Human Rights Center.
“And that’s because LLMs enable you to do things without having to put all of the forethought into it, without having to organize your thoughts, without having to research; it cuts back on so many steps.” Popken specifically noted the impact on young students learning fundamental skills like reading, research, outlining and revision, arguing these steps are important for supporting independent thought and comprehension.
This duality necessitates a subtle approach to regulation and oversight, acknowledging the benefits of these technologies while proactively addressing their inherent risks. Researchers observed that communities demonstrating excitement about LLMs often possess greater access to resources and opportunities, while those expressing risk aversion may lack the means to fully use the technology’s potential. This disparity emphasises the need for equitable access and support to ensure LLMs do not exacerbate existing inequalities.
The team’s recommendations extend beyond simply identifying risks. They advocate for proactive assessment of human rights impacts. “We think all developers, all deployers should be doing that, and they could do it at a large scale or small scale level,” Popken clarified. This assessment, the report suggests, should be tailored to each group’s specific role and responsibilities.
Collaboration between these groups is also deemed essential, as many of the recommendations require a coordinated effort to effectively address the complex challenges posed by LLMs. To disseminate their findings, the Human Rights Center is hosting two webinars. The first will target a broad audience in October, and a second is planned for November to extend outreach to the broader industry space. “Our hope is to share what we have learned to as broad of an audience as possible,” Popken stated.
The intention is to move beyond academic discussion and support practical implementation of the report’s recommendations. The podcast “Berkeley Law Voices Carry,” hosted by Gwyneth Shaw, features a discussion with Popken, Raman, and Camille Chabot, a former center researcher now pursuing a JD at Harvard Law School.
The episode examines the report’s methodology and key findings, offering a comprehensive overview of the human rights implications of large language models. The report’s authors emphasize the importance of a comprehensive approach, considering both the potential benefits and risks of LLMs across diverse sectors and stakeholder groups.
2022 ChatGPT Launch Prompted Human Rights Risk Assessment
Their involvement stemmed from a growing recognition that while ChatGPT and similar technologies offered significant potential, they also presented considerable risks, particularly within professional sectors. The impetus for the study arose at the close of 2022, as the Human Rights Center sought to proactively address these emerging challenges and guide stakeholders toward responsible implementation. The research team deliberately focused on three key fields, law, journalism and education, recognizing these areas as particularly susceptible to both the benefits and harms of large language models.
This selection was also driven by the center’s existing network of global contacts within these professions, allowing for a broad and subtle understanding of the potential impacts. Researchers intentionally sought input from individuals across the globe, aiming to capture regional variations in risk perception and application. This global scope extended to linguistic diversity, facilitating direct engagement with experts and practitioners in diverse cultural contexts.
A key component of the investigation involved translating established human rights principles, such as those outlined in the UN Guiding Principles, into quantifiable metrics suitable for computer scientists. The goal was to create a framework for assessing risk saliency, considering factors like scope, scale and likelihood, allowing for a more systematic evaluation of human rights impacts during the development and deployment of language models. This approach aimed to bridge the gap between abstract ethical considerations and concrete technical implementation, providing developers with tools to proactively identify and mitigate potential harms.
Observations revealed a notable difference in risk aversion between professionals interviewed in the Global North and those in other regions. Those in the Global North tended to emphasize the potential human rights risks associated with large language models, while others focused more on the opportunities these technologies presented.
This disparity highlighted the importance of considering contextual factors and diverse perspectives when evaluating the impacts of AI. Researchers noted a particular concern regarding the potential effects on children, whose developing minds may be especially vulnerable to the influence of these models, and the importance of supporting critical thinking skills to counteract potential manipulation.
Developers, for example, were urged to evaluate potential risks “upstream,” during the data collection and labeling phases, while distributors were tasked with ensuring responsible deployment and monitoring. “We believe each of these stakeholder groups may find different human rights impacts based on their assessment or how in-depth it was,” the researchers found. The study highlighted the potential for both positive and negative impacts across various rights, suggesting that a comprehensive approach is essential.
The team observed that the same factors could simultaneously pose a risk to one right while offering an opportunity to advance another, necessitating careful consideration of trade-offs and unintended consequences. This complexity reinforces the need for ongoing dialogue and collaboration between diverse stakeholders to ensure that the benefits of large language models are realized while minimizing potential harms.
The researchers emphasize that developers should evaluate human rights risks at every stage of the process, regardless of the scale of their operation. The findings are not intended to remain within academic circles; the Human Rights Center is actively disseminating its recommendations through multiple channels, including hosting webinars with organizations like the Global Network Initiative and Dropbox.
Human Rights Center Blends Legal & Computer Science Expertise
The Center’s investigation, initiated in late 2022, paired legal expertise with the technical insights of computer scientists like Vyoma Raman, now a PhD candidate in information science at Cornell, to address a gap Popken identified in existing research.
“I think this has been studied quite a lot in sort of the algorithmic fairness literature at the level of the technical systems,” Raman explained, “And I saw this project as an opportunity to go beyond that, to connect to the downstream human rights impact in a way that human computer interaction scholars end up doing.” The interdisciplinary approach was important, stemming from Popken’s experience co-founding and leading the business and human rights practice at the law firm Orrick, Herrington Sutcliffe. Observing the rapid deployment of ChatGPT, she recognized both the potential benefits and inherent risks of the technology, particularly within professional sectors.
This prompted a search for collaborators who could bridge the gap between legal frameworks and computational capabilities, ultimately leading to the inclusion of Raman and Camille Chabot, who began her JD studies at Harvard Law School following her work on the report. “We also have a number of journalists on our team at the Human Rights Center, so those were three fields where we felt like we could really begin to understand what those risks might be, and we’d have kind of connections to people all around the world,” Popken noted.
This focus extended to considering how LLM use varied across different regions, acknowledging that risks and opportunities would not be uniformly distributed. Raman further detailed the team’s commitment to a thorough evaluation.
Raman articulated a frequent sentiment among computer scientists, noting, “I think it’s important to care about rights and ethics and those sorts of things, but what I have the power and the expertise to do is to tune this parameter in the model that I end up training, and I don’t really understand how that relates to rights.” The over 15-month investigation aimed to bridge that understanding, providing a framework for translating technical adjustments into measurable human rights outcomes.
Algorithmic Fairness Concerns Extend to Downstream Human Rights Impacts
Fifty-six individuals from twenty-four countries contributed to the report’s findings through interviews, ensuring a broad and globally representative understanding of large language model impacts across multiple sectors. This extensive outreach included at least one expert and one practitioner in law, journalism and education on every inhabited continent, demonstrating a deliberate effort to capture diverse perspectives. The research team prioritized input from those directly experiencing the effects of these technologies, rather than relying solely on theoretical analysis.
The report’s methodology extended beyond typical algorithmic fairness evaluations, which often focus on technical system performance, to specifically excavate downstream human rights consequences. “And I think working with Betsy and Camille was especially valuable because of their expertise in law and international relations that helped to ground exactly what the problems could be and design our methodology around excavating those,” Raman said, emphasizing the importance of interdisciplinary collaboration in identifying and addressing potential harms.
Popken described a need to move beyond simply identifying opportunities and instead proactively assess and mitigate potential risks for various stakeholders. This involved assessing how the technology might affect journalistic practices and, consequently, the public’s access to accurate information. Raman detailed that the team’s approach was comprehensive, pairing human rights principles with computer science expertise to provide a thorough analysis.
“So since many of the evaluations that they already do focus on the capabilities of these models in terms of doing math or writing or understanding things, that sort of thing, like we ask what it would take for them to extend those evaluations to focus concretely on human rights risks and opportunities,” Raman explained. These groups are assigned a robust set of recommendations, some of which require cross-sector collaboration to implement effectively. Popken emphasized the importance of proactive assessment of human rights impacts by all stakeholders, regardless of scale.
This highlights the need for tailored recommendations and collaborative efforts to address the complex interplay between technology and fundamental rights. Camille Chabot, now a JD candidate at Harvard Law School, noted the importance of considering geographical and linguistic variations in risk assessment, recognizing that the impact of LLMs may differ significantly across regions. The report’s findings are not limited to identifying risks. It also proposes actionable steps for each stakeholder group to promote responsible AI development and deployment.
These recommendations range from conducting thorough human rights impact assessments to supporting collaboration across sectors and prioritizing transparency in algorithmic design. The team hopes to share their learnings with a broad audience, empowering stakeholders to navigate the ethical challenges posed by LLMs and harness their potential for positive social impact.
See today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals.
