From Hallucinations to Trust: Can Legal AI Reform Justice?
Imagine asking an AI chatbot for legal advice and receiving a response that seems confident, detailed, and authoritative. You act on it, only to discover later that the advice was entirely fabricated — a phenomenon known as “hallucination” in the world of generative AI. This isn’t just a hypothetical scenario; stories abound of AI tools confidently producing misinformation, sometimes with severe consequences. One of the more well-known examples involved a New York attorney that cited fake cases generated by ChatGPT in the brief he filed with the federal court.
As we’ve seen, while the potential of generative AI in law is immense, offering a vision of cost-effective, accessible legal services, the specter of inaccuracy raises critical questions: Can legal chatbots truly serve the public without compromising reliability? Are ChatGPT and other free chatbots really enhancing access to legal services, or do their flaws outweigh the benefits to justice? Think about it, if an experienced New York attorney didn’t spot the mistakes in his brief, how is an ordinary person that is seeking some form of legal advice or assistance from ChatGPT or one of its counterparts, going to extrapolate the accurate or relevant information for their case?
Does this mean that generative AI tools should be completely dismissed? Of course not. But one thing is certain – the industry can only really claim that the result of generative AI products is enhanced access to justice once it can demonstrate more accurate output.
This article explores the challenges posed by hallucinations in AI and the potential solutions that could transform these tools into dependable allies for justice.
Understanding AI Hallucinations
At the heart of the debate regarding whether tools such as ChatGPT are truly advancing access to justice, is hallucination — the tendency of AI models to generate incorrect or fabricated information while sounding convincingly accurate. For instance, chatbots might create fictitious legal precedents, misinterpret statutes, or oversimplify complex legal doctrines.
In routine applications like drafting a brief response email or brainstorming ideas, these errors may be fixable, or even inconsequential. In the context of fully fledged legal proceedings, comprehensive and high-risk drafting or advice, however, the stakes are far higher. In these circumstances, erroneous advice could lead to financial loss, damage to reputation, or even wrongful convictions.
Legal Chatbots: Promise and Peril
Generative AI tools like ChatGPT are celebrated for their ability to process vast amounts of information and respond quickly. In theory, they could democratize access to legal resources, providing guidance to individuals who cannot afford professional counsel. But in practice, their current limitations make them risky for high-stakes decision-making.
The Potential Benefits
- Accessibility: AI chatbots could help bridge the justice gap by offering legal information to lower income individuals or communities.
- Efficiency: These tools could streamline legal research, identify precedents, and generate preliminary drafts for contracts or motions.
- Scalability and availability: Unlike human lawyers, AI can handle an unlimited number of queries simultaneously, making legal support available 24/7.
The Risks
- False Confidence: Users may not realize AI’s limitations and may take its advice as infallible.
- Lack of Nuance: AI models lack the contextual understanding necessary for nuanced legal reasoning.
- Accountability Issues: If the advice leads to harm, who is responsible—the user, the developer, or the AI itself?
Challenges in Using AI for Legal Decision-Making
1. Accuracy and Reliability
The legal domain requires precision, as even minor inaccuracies can have significant ramifications. Current AI models are trained using generalized data and are not equipped to deliver jurisdiction-specific advice or interpret nuanced legal language accurately.
2. Ethical and Regulatory Concerns
- Confidentiality: AI tools must ensure that sensitive legal information remains secure.
- Bias: Training data often reflects systemic biases, which can result in unfair or discriminatory outcomes.
- Oversight: Without clear regulations, there is a risk that these tools could be misused or over-relied upon.
3. Public Understanding
Many users mistakenly assume that chatbots are as reliable as human experts. Educating the public about the limitations of AI tools is essential to prevent misuse or overconfidence in their outputs.
Exploring Solutions
So, now that we know the information we receive from AI, especially if not properly trained, could be a hallucination, is all hope lost? And if not, how can legal AI tools move from unreliable assistants to trusted advisors? Luckily for those seeking access to justice, the public faults of ChatGPT have resulted in improvements on several fronts, including the introduction of several more accurate tools.
One emerging solution is EPO-Bot, a specialized tool designed to navigate legal challenges with greater accuracy and domain-specific expertise. While the broader answers lie in the integration of ethical frameworks, technological advancements, and human oversight, tools like EPO-Bot represent a promising step forward.
The Legal Wire recently interviewed Dr. Tomer Libal, who heads up the legAI team at the University of Luxembourg, regarding his work in the generative AI space, and particularly to find out how the legal challenges surrounding hallucinations can be tackled. Dr. Libal noted that “there are various approaches to reducing hallucinations. For example, the RAG approach decreases the set of documents which are used for generation and in this way, both decreases the chance for hallucination and make it easier to verify the source. The solution taken by the legAI team and within the ExAILe project with Dr. Aleksander Smywiński-Pohl from the AGH University in Krakow is to eliminate hallucinations almost completely by reducing the role of generative AI to a minimum and replacing the main work which is done by generative AI in tools like ChatGPT with other forms of AI which do not exhibit hallucination, such as symbolic reasoning.”
Even with the increased accuracy, however, Dr. Libal warned that users should remain cognizant of the need for safeguards to be in place. Due to the limitations of AI tools, “the user needs to double check the result (as is being done by all other chatbots and assistants)” he warned. What’s improved in the context of EPO-Bot, however, is that “the final legal advice contains detailed explanations with references to all legal articles in question and how they have been applied in the specific case.”
The EPO-Bot is funded by the FNR agency in Luxembourg and NCBR in Poland and has been created via a collaboration between the University of Luxembourg and AGH in Krakow, Poland, within the ExAILe Project, headed by Dr. Libal and Dr. Smywiński-Pohl. Drs. Libal and Smywiński-Pohl have also co-founded the startup Enidia AI, which develops tools for legal professionals with the same emphasis on accuracy, transparency and accountability.
The challenges in the legal domain, identified by Dr. Libal as “accuracy, transparency but mainly accountability” can be addressed “by using a combination of AI tools and putting the user in the loop.”We understand from Dr. Libal that “[in] this way, accuracy and transparency are provided by other AI techs which are highly accurate and transparent. Accountability is provided by the user who is involved in the process in a way that gives her/him a full understanding of the role they play in the decision making. Finally, generative AI and other machine learning tools are used to improve the interaction with human users.”
Last week, The Legal Wire wrote about the risks of AI partnerships involving big tech companies for effective competition. When asked about the role of partnerships between AI developers and legal professionals in creating accurate and ethical tools, Dr. Libal responded “this is an excellent question! Unfortunately, there is not always a seamless interaction between AI researchers, Legal researchers and actual legal practitioners. For creating tools that provide a value for users, whether the public or legal practitioners, such interaction is highly important. The EPO-bot is based on such a collaboration between the CS and Law departments, as well as with extensive discussions with lawyers.”
A Hybrid Future: Humans and AI in Law
The future of legal AI likely lies in hybrid systems where increasingly refined AI tools complement, rather than replace, human expertise. For instance, chatbots could handle routine queries or draft documents, while humans focus on strategy and complex problem-solving. Such a model would combine the efficiency of AI with the critical thinking and ethical judgment of humans that are using AI responsibly and ethically.
In this regard, whether legal AI tools could eventually work independently of human supervision, according to Dr. Libal, “depends on the purpose of the tools. For legal advice which depends on legal interpretation, which in turn depends on case law and a cognitive process to match facts between cases and regulations, a human in the loop is required for the proper work. Nevertheless, this human can be a lay person, if empowered to be able to play this role by understanding, for example, how to interpret the law properly, as is done in the EPO-Bot.”
So, Is It All Doom and Gloom for AI in the Fight for Access to Justice?
The potential of legal AI is undeniable, but so too are its risks. Chatbots have the potential to revolutionize access to justice, but only if they are built and deployed responsibly. Reducing hallucinations, ensuring accountability, accuracy and transparency are essential to this journey.
AI may never fully replace the nuanced reasoning of a human lawyer, but it can still play a transformative role in democratizing access to legal resources — if we get it right. Fortunately, those creating tools such as EPO-Bot are working tirelessly on ensuring the risks and challenges associated with traditional generative AI are tackled.
So, what’s next? For Dr. Libal and his team at least, they are continuing, through the academic project, to improve their legal assistants for better access to justice and are planning to improve this work on two axes: increasing the legal problems these assistants can provide legal advice for, such as GDPR, human rights and more; and improving the quality of the algorithms behind all these assistants, mainly by allowing the user to respond better to the assistant questions via a better understanding of the relevant case law.
