Large language models are often described as either intellectual prostheses or intellectual solvents. A new assessment from the Human Rights Center at UC Berkeley School of Law suggests that the more useful distinction is less dramatic: what matters is whether the model extends a person's ability to act, or quietly replaces the parts of the task in which judgement is formed.

The report, An International Analysis of the Human Rights Impacts of Large Language Models, examines LLM use in law, journalism and education. Its authors — Betsy Popken, Vyoma Raman and Camille Chabot — interviewed 56 experts and practitioners across 24 countries, reviewed existing research and regulation, and analysed model evaluations through a human-rights framework derived from the UN Guiding Principles on Business and Human Rights.

Its central finding is deliberately two-sided. LLMs can expand access to information and services, particularly where professional capacity is scarce. They can also introduce error, homogenise the information people encounter and encourage users to delegate the cognitive work required to detect those failures.

That tension is clearest in places where the alternative to imperfect AI assistance is not expert human assistance, but little assistance at all.

In law, interviewees described uses ranging from translation to help for self-represented litigants. The Berkeley report argues that conversational systems may improve practical access to justice by making procedures and legal material easier to understand. But the value of cheap advice depends rather heavily on the advice being right.

That remains a serious constraint. A Stanford-led preregistered evaluation of commercial AI legal-research systems found that products from LexisNexis and Thomson Reuters still hallucinated in more than 17 per cent of tested queries, despite performing substantially better than general-purpose chatbots. An earlier study of public-facing models found much higher error rates on questions about US case law. Those benchmarks do not establish how every current model performs in every legal task, but they make one point difficult to evade: lowering the cost of legal information and lowering the cost of legal error can happen at the same time.

Education shows a similar reversal. The Berkeley researchers identify personalised explanation, translation and adaptation for students with disabilities or different learning needs as potentially important gains. Yet they rank overreliance as a threat to freedom of thought because students can avoid the effort through which reasoning skills are developed.

There is empirical evidence for the concern, but not for a simple conclusion that AI makes people think less. A 2025 randomised controlled trial of 120 undergraduates found lower long-term knowledge retention among students who had used ChatGPT as an unrestricted study aid. A Microsoft Research study of 319 knowledge workers found that greater confidence in generative AI was associated with less reported critical-thinking effort, while greater confidence in one's own ability was associated with more.

But a 2026 systematic review of 67 empirical studies reached a more conditional conclusion. ChatGPT use was associated with stronger critical and creative thinking when embedded in inquiry-based, scaffolded teaching, while unstructured use was more often associated with cognitive offloading and weaker critical engagement. The technology appears capable of supporting reasoning and bypassing it. Instructional design decides a great deal about which one occurs.

Journalism produces the same pattern in institutional form. The Berkeley interviews found examples of LLMs taking on routine reporting and research tasks in resource-constrained newsrooms, potentially leaving journalists more time for investigation and analysis. The report also notes the less attractive possibility: management can take the productivity gain as a staffing reduction instead. Time saved by automation is not automatically reinvested in journalism.

This is where the human-rights framing becomes both useful and contestable. Freedom of thought is a protected right under Article 18 of the International Covenant on Civil and Political Rights; the UN Human Rights Committee has described it as exceptionally broad and fundamental. Extending that legal concept to gradual skill loss, cognitive dependence or narrower exposure to viewpoints is an interpretive move, however, not a measured finding that LLM use itself constitutes a rights violation.

The Berkeley authors are reasonably careful about that distinction. Their study is not a population survey and does not estimate the prevalence of particular harms. Interviews were conducted in English and participants were recruited digitally, which the report acknowledges biases the sample toward people who are already connected and English-speaking. Its risk ratings combine interview evidence, literature and normative judgement rather than supplying experimental estimates of how much LLM use changes cognition.

One of the study's more revealing observations is geographic. Interviewees in parts of the Global South were generally more willing to use LLMs for ambitious, high-value tasks than counterparts in Europe, North America and Oceania. The researchers interpret some of that difference as a response to scarcity: where lawyers, teachers, journalists, specialist information or translation are harder to obtain, an unreliable tool may still be preferable to no practical access.

That complicates the easiest policy response. A rule designed solely to prevent dependence can also preserve existing inequalities in access. Conversely, calling AI "democratising" because it supplies a cheap approximation of scarce expertise can turn unequal provision into a feature rather than a problem.

The report therefore recommends controls aimed at uses rather than a blanket prohibition: stronger human-rights assessment, profession-specific systems, better remediation when harm occurs, more representative training data, AI literacy and regulation concentrated on high-risk applications. Its example of judges using LLMs to make rulings is an obvious case where convenience should not determine the boundary.

The more general lesson is less about whether LLMs are good or bad for thought than about where the thinking remains. Using a model to translate a document, test an argument, expose an unfamiliar source or generate alternatives can enlarge the field in which judgement operates. Using it to decide what is relevant, what is true and what conclusion to accept can compress that field instead.

The distinction is not stable. A tool introduced as assistance can become infrastructure; infrastructure can become expectation. The human-rights question begins there, when an optional shortcut starts determining who receives knowledge, whose judgement counts, and which forms of thought institutions still require humans to practise.

Sources