At the forefront of a rapidly evolving field of artificial intelligence (AI) is a suite of tools that purport to save lawyers valuable time and effort on standard research chores. A recent Stanford study – the first of its kind – has put these tools under the microscope, finding that they routinely ‘hallucinate’ or deliver false or erroneous information long after the testing period. What’s the story behind the story? How can quickly evolving technology promise to transform the age-old relationship between practitioners and law, and what challenges does modernisation pose?
RAG technologies are leading the charge among advances that might mitigate hallucination risk It’s an ominous finding, and it reveals a fundamental problem that could plague future AI researchers. Even if every research-based AI is equipped with, say, a retrieval-augmented generation (RAG) technique – which delivers answers to queries by drawing information from large numbers of legal documents in a way that helps contextualise the AI’s response – a troubling number of hallucinations will still be created. In fact, that is exactly what the Stanford study found: for 17-33 per cent of the legal queries they tested, hallucinations were still generated.
Legal questions are often indeterminate – there is no single ‘right’ answer. As a result, a Herculean task for AI is figuring out which documents are ‘relevant’ to a given query. The Stanford study brings home a major point: legal AI tools, when they go wrong, lead to interpretative failures, even hallucinations, which might misguide legal decisions.
The quest for the perfect legal AI assistant may be a fool’s errand, but it also points to a real need: more transparency and benchmarking in legal tech today. This study calls for an open conversation and shared standards that we hope will herald a new wave of accountability and improvement in legal research technologies.
Despite the long shadow that such hallucinations cast over what AI might do, we shouldn’t lose sight of the SHARP new benefits that these tools bring to legal practice. The enhanced speed and efficiency of legal research, the insights into new lines of legal argument, the identification of relevant legal precedents – each of these capabilities (and more) can potentially be opened up by AI-assisting tools. But, just as the Stanford study vividly illustrates, they should be seen as just that – as assistants, not replacements, for the kind of careful scrutiny that must be undertaken by legal practitioners.
The path of integrating AI into legal research is beset by challenges but brings with it added potential for creativity and imagination. At the heart of it lies a culture of openness, whereby legal tech providers talk openly about what works and what doesn’t work in the context of sharing-algorithms research. They help to shape the development of AI tools that are not just sharp, but steadfast.
To understand the study’s results and their implications, however, it’s important to unpack what ‘sharp’ means in this context. ‘Sharp’ denotes a commitment to accuracy, clarity and fidelity as an aim of AI-enabled legal tools – an aspiration to the point at which AI legal research tools don’t do research ‘more quickly’ as an added benefit, but do research ‘correctly’ full-stop.
Legal professionals at the threshold of a technological transformation have an obvious path ahead of them. By using AI wisely, promoting transparency and engaging in constant improvement, they have the opportunity to harness these powerful tools to the fullest, giving clients the insightful and unerring advice the law requires.
More Info:
© 2024 UC Technology Inc . All Rights Reserved.