As Amer Mouawad sits in his office surrounded by AI interns, he is not at all surprised by the rapid adoption of artificial intelligence, particularly within the legal profession. Heading AI at Siren Analytics, he can empathise with the impact of the new wave on those who are unprepared, grappling with stacks of paper and tackling tedious tasks that can span days.

Recent advances in AI such as the public release of ChatGPT in late 2022 are revolutionising the way information is processed and disseminated, driving huge efficiency gains. As AI assistants continue to evolve, the potential for transformative impact across industries has become increasingly apparent. It’s clear anyone who does not embrace change will be left behind.

Amer has spent the past few months analysing the impact on the field of law of Large Language Models (LLMs) - a type of AI model designed to understand and generate human-like text. The findings point to an urgency for those within the judicial system to leverage available resources to their benefit.

Easing the workload

LLMs are trained on vast amounts of text data and use advanced machine learning techniques, particularly deep learning, to grasp the intricacies of language. In law, LLMs provide a robust, singular approach with impressive summarisation and translation capabilities. Moreover, they can be trained and adapted to perform a wide range of complex functions, including language inference and sentiment analysis.

“The integration of LLMs is modernising the legal landscape at a fundamental level as these models offer unprecedented capabilities in legal research, document parsing and decision making. By effectively processing immense volumes of legal data, LLMs streamline workflows, enhance access to justice and can potentially democratise legal expertise,” Amer stated.

Their adeptness at comprehending and generating natural language allows LLMs to provide precise and accurate legal analyses, empowering lawyers to offer more thorough counsel. They can also interpret the meaning of legal jargon across languages, summarise and synthesise case files and legislation, and draft legal documents and memos. All this saves legal professionals time, supporting quicker decision-making and enhancing accessibility.

Obstacles remain

Despite the hype, LLMs have their flaws. They can occasionally misinterpret negations or struggle with context-dependent meanings, especially in intricate sentences or ambiguous contexts. Training them on low-quality or flawed data also leads to poor outputs, including bias and discrimination. Then there’s the risk that models generate content not grounded in reality, lacking authenticity, or that deviates significantly from the input provided.

“Determining the optimal length of generated content can also be a complicated assignment. If too short, the generated text may lack detail to fully address the prompt, while if the length is too long, the text may be repetitive or lose focus. If the user provides incorrect or biased input prompts to the model, it can produce outputs that reflect those shortcomings,” Amer said.

Finally, it can be costly and time consuming to train an LLM. It involves processing massive amounts of text data through neural network architectures, which demands extensive computational power, specialised hardware and substantial time investments. Differential access to these resources can further embed existing global inequalities, with the power to determine how AI is developed and deployed concentrated among a few large companies.

Workarounds

According to Amer, we can address the limitations of LLMs through a combination of techniques and approaches.

Their performance and reliability can be improved by connecting them to a wealth of external knowledge and information. For example, retrieval-augmented generation systems enable LLMs to search through and retrieve information from external sources. Grounding - which involves giving the LLM access to pre-defined databases or text corpora – similarly ensures outputs are supported by evidence and relevant context.

To improve an LLM's ability to generate logical explanations, the LLM can be trained to generate and provide a step-by-step explanation of its reasoning process. This technique, known as chain-of-thought processing, helps improve the transparency and interpretability of outputs. Engineers can also train LLMs to generate summaries of varying levels of detail using multi-level summarisation techniques.

To overcome problems around limited training data, few-shot learning - a machine learning paradigm where a model is trained to learn and generalise from a limited number of examples or "shots” – can help LLMs generate contextually appropriate responses. Lastly, using indexed and normalised data sources ensures that the information fed to LLMs is consistent, well-structured, and free of biases, further improving the accuracy and reliability of their outputs.

So, while LLMs may have their drawbacks, AI assistance has the potential to simplify, elevate, and optimise the legal discourse for all stakeholders. As LLMs advance further, they are expected to remain a positive force, reshaping the study, execution and accessibility of law for both legal professionals and the general public.