The author would like to thank prof. dr. Natali Helberger for her helpful commentary on this blog post.
Introduction
The legal system can be a fortress. While anyone can freely read their country’s laws, much more is needed to grasp the complexity of the legal system. Lawyers train for years to gain the skills to engage with the law. Yet, LLM-based chatbots provide billions of people now with access to this, often almost esoteric, type of knowledge. Though far from perfect, LLMs have nevertheless produced a societal revolution in the provision of legal services and access to justice for years to come.
This rapid expansion of access, raises questions about the reliability and nature of the legal guidance LLMs provide. This is because LLMs do not truly ‘reason’ like humans do. They find and apply patterns in the datasets that they have been trained on, and are further adjusted in this regard by human reviewers through reinforcement learning. This method has not produced a system equal to a human legal advisor. As Wachter, Mittelstadt, and Russell point out, ‘LLMs produce responses that are plausible, helpful and confident, but that contain factual inaccuracies, misleading references and biased information.’ Consequently, LLMs can engage in ‘bullshitting’, as they are ambivalent about whether or not they produce accurate or truthful information. LLMs can be ‘sycophants’ that aim to manipulate their user to believe that their output is true, simply to satiate their users’ requests.
Considering these issues, this blog post explores the ability of LLMs to enhance access to justice, as well as the risk that hallucinations pose for misinforming citizens about their rights. I will also discuss one potential method to reduce such ‘hallucinations’, through the introduction of Retrieval Augmented Generation (RAG). Nonetheless, I argue that the benefits of RAG-enhanced LLMs have thus far only been available to legal publishers, who require their users to pay high fees to access these systems. This unfairly limits access to more reliable RAG-enhanced LLMs to professional lawyers, while lay citizens can only access general-purpose LLM-based chatbots that are prone to legal hallucinations.
LLMs’ role in access to justice
Access to justice can be limited in a number of different ways. Individuals might not be able to engage with the law because they cannot afford legal services, they might not understand difficult legal jargon, legal service providers might not be available in their region, or litigants might not speak the right language. The OECD recommendation on access to justice and people-centered justice, underlines the potential of technology to address such issues. To that end, LLMs could increase access to justice by offering informative, immediate, easy-to-use, and free legal guidance to those unable to afford, understand or otherwise receive, traditional human-based legal services.
LLMs could assist in identifying legal issues, explain what an individual’s rights are, and guide users through legal procedures. More concretely, LLMs can assess one’s eligibility for certain legal rights and benefits and gather relevant information to that end. Scholarship has also identified the usefulness of LLMs in making legal texts more intelligible for those whose access to justice is contingent on the use of more plain language. To illustrate, Pałka et al. argue that LLMs can be used to summarize privacy policies and make them more understandable to non-lawyers. LLMs can also be accessed by anyone with an internet connection, reaching even citizens who live in ‘rural legal desserts’.
Moreover, besides these financial, legibility, and geographical benefits, LLMs can also address cultural and psychological barriers to access to justice. LLMs can play the role of a neutral legal advisor who will not judge or shame a person looking for aid. Such issues are, notably, a very real obstacle for victims of sexual violence, who can be limited in both their willingness and ability to seek help because of fears for negative repercussions.
In short, LLMs can help provide legal services to people who, because of financial, language, geographical, or shame-based issues, otherwise cannot engage with the law. Nevertheless, LLMs are not an equal replacement for free legal service providers that currently provide a vital lifeline for access to justice. In this regard, we can think of organizations like the ‘Juridisch Loket’, in the Netherlands, the LawWorks charity with its various ‘law clinics’ in the U.K., or the government-sponsored ‘Beratungshilfe’ initiative in Germany. Stronger still, according to art. 6(3)(c) of the European Convention of Human Rights, legal advice must be given free of charge to those who cannot afford aid in criminal cases.
In an ideal world, we should promote the provision of such free (or affordable) legal advice through additional funding. But, the world is a place with scarce resources, and it only contains a limited amount of (financial) means to support such initiatives. Logically, LLMs can provide these services faster, cheaper, and more efficiently. These benefits must therefore also be taken into consideration in academic and policy debates on access to justice. But, in this regard, we must be careful, not to argue that LLMs are in any way shape or form equal to actual skilled legal advice offered by humans. Nor does the existence of LLMs justify defunding such service providers.
Potentially, however, social workers and social lawyers, can benefit too from the help and advice offered by LLMs. Ideally, this would lead to a synergy of strengths; efficiency and mass-scale access to justice from LLMs, combined with high quality legal reasoning from real humans. LLMs are especially apt in this context, as it is often social lawyers that are overworked and understaffed, or not available at all. To illustrate, in the Netherlands, a severe shortage of social lawyers is imminent, and in the U.S.A. 92% of low-income citizens are not able to receive help for civil legal problems. To therefore promote both the accessibility and availability of free or affordable legal help, social legal service providers could very much benefit from the use of LLMs to help streamline their workflow.
But, as Ryan and Hardie also state, social legal service providers need to be wary of providing inaccurate LLM-hallucinated legal advice. The same holds especially true for LLM-based chatbots that directly engage with lay citizens, who have never had any legal education. It is this group that would especially be unable to discover any legal mistakes in the outputs produced by the system in question.
LLMs in legal practice
Although LLMs can serve the wider public, the advent of legal LLMs have so far mostly benefited professional lawyers. Law students, legal practitioners, and even judges, have been quick to adopt free-to-use systems such as ChatGPT, Gemini, or Claude. We use them to structure our texts and to answer questions both legal and factual. It’s no surprise then to also see that legal publishers have found a new lucrative opportunity in this technology. A flurry of new LLM-based chatbots has arisen in recent years from these companies. Examples include Lexis+AI. Professional users can ask this system legal questions in natural language and receive quick answers based on verified legal sources. The tool can also draft legal documents, summarize cases, and help with creating a legal analysis, making research faster and more efficient (at least according to its developer, LexisNexis). Similarly, in the Netherlands, the legal publisher SDU, has developed their own LLM-based system too, called GenIA-L, for the Dutch legal field.
However, the issue of hallucinations has also been a persistent topic of concern. Famously, a New York judge even sanctioned a lawyer for citing non-existent case law in a legal brief in 2023. Recently, the law firm Morgan & Morgan send out an urgent email to its 1000+ lawyers, warning for LLM-made hallucinations. Research shows that such hallucinations are far from rare. One study on legal prompting shows that in 58% of cases, legal hallucinations occur with ChatGPT 4 and in 88% of cases with Llama 2.
Reducing hallucinations with retrieval augmented generation
One way to limit the occurrence of hallucinations is through the use of retrieval augmented generation (RAG). With RAG, LLM-based systems rely on a retriever that first fetches relevant documents from an external knowledge base. Then, the LLM uses these documents to generate a response for the user. This approach helps reduce hallucinations by grounding answers in real, verifiable data instead of relying solely on an LLM’s vast (and sometimes incorrect) internal memory. Because RAG uses actual verified sources to inform its responses, it also enhances transparency: users can trace answers back to the original documents, which supports fact-checking and builds trust. By connecting retrieval with generation, RAG provides more reliable outputs, especially for knowledge-intensive tasks like legal argumentation. This has furthermore been shown in empirical research as well, which illustrates that RAG can improve the quality of LLM-generated legal texts, and lower the amount of hallucinations. RAG has, subsequently, been adopted in legal tech design as well. A good example of this in practice, can be seen with the INACIA system, which relies on RAG-enhanced LLMs to support Brazilian courts in triaging cases.
It is no wonder then, that the aforementioned legal publishers, LexisNexis and SDU, rely on RAG-enhanced systems too for their LLMs. It is especially these actors, after all, that have access to a treasure trove of legal commentary, articles, case notes, and case law. This provides these companies with quite a strategic benefit over any other competitors wishing to develop reliable and accurate legal LLMs.
Currently, the customer base of these publishers has so far has been leaning toward large (commercial) legal service providers, like law firms. Or in other words: companies that have the financial capacity to pay high subscription fees. Pro bono legal advice providers, like law clinics, social law firms, or charity organizations, do not always have enough money to pay these fees, and therefore do not get to have access to these RAG-enhanced LLMs. In turn, those without the required financial means will turn toward multi-purpose and free-to-use LLM-based chatbots, like Gemini or ChatGPT, which have a very high risk for generating legal hallucinations.
High-quality, fast, and easy-to-access LLM-based legal advice will consequently only be accessible for legal professionals who already have the knowledge and skills to analyze and research the law themselves. Meanwhile, those ordinary citizens who would benefit most from reliable and non-hallucinated LLM-based advice, are excluded from the use of such systems.
Open data and RAG-enhanced LLMs
One could argue that this is not necessarily unfair, as the aforementioned publishers do often own the copyright to these works. Nevertheless, these systems rely on the legal commentary and articles that are in the hands of these publishers, which are often written free of charge by legal scholars. Stronger still, these scholars are themselves often paid with public funding. Moreover, the case law that these publishers rely upon, is also provided, edited, and anonymized, by judiciaries, for the benefit of the general public. These same judiciaries are then, however, sold back their own data in the form of assistive AI systems, as Stern et al. point out in the case of China’s highly digitized court system.
Currently, there is already a growing movement toward open access publishing in academia. One example of this is the diamond open access project, supported by various Dutch universities. This initiative strengthens and promotes non-profit scholarly- and library-led publishing. Arguably, this makes it possible for other parties, besides big publishers, to develop RAG-enhanced LLMs based on this open data.
Nevertheless, open access publishing does not mitigate the need for paid labor to actually develop and finetune these chatbots. How then, should universities and judiciaries ensure public access to RAG-enhanced LLMs? Arguably, there are two potential avenues to ensure that chatbots are accessible for the broader public. First, judiciaries can develop their own open access chatbots with additional public funding. To this end, they might also leverage their (unpublished) case law to enhance these systems with RAG. Alternatively, both judiciaries and universities can engage in public-private partnerships with publishers. Remember, without the data from the former two, publishers are not able to properly enhance their chatbots using RAG. In exchange for continued supply of such data, case law and legal commentary, publishers could be required to provide their chatbots freely (or for a discount) to social law firms, or other organizations that provide free or subsidized legal advice to the public.
Conclusion
My argument in this blogpost has not been that LLMs have the skills or ability to replace real, knowledgeable, and qualified lawyers. An overly laudable description of LLMs can risk legitimizing the lack of human-based legal services. Still, the argument stands that the material reality of the world, in particular the lack of economic and epistemic capital available to the vast majority of people, must be addressed. In such a world, LLMs can be a much-welcome support for non-lawyers.
Nevertheless, the benefits of RAG-enhanced LLMs are disproportionately only available to those already embedded in the legal ecosystem, such as law firms and legal publishers. To ensure equitable access, it is crucial to integrate open access initiatives and public-private collaborations that allow free or affordable legal service providers and ordinary citizens to benefit from these technologies. RAG-enhanced LLMs, supported by open data policies, can help thereby minimize hallucinations and improve access to justice for all.
Ljubiša Metikoš
Ljubiša Metikoš is a postdoctoral researcher at the Institute for Information Law (IViR), and the Research Priority Area Human(e)AI. His research combines insights from information law and legal-philosophy to understand the impact of AI on legal decision-making.
