Practical implications of ChatGPT for arbitration practitioners
What is ChatGPT?
ChatGPT has captivated public imagination in recent weeks and months, a supposed ‘quantum leap’ in artificial intelligence (AI) for the everyday user. A chatbot that can not only answer complex and nuanced questions on a myriad of topics but also exhibit impressive creativity like writing songs and poetry. ChatGPT is the newest service from the Microsoft-funded OpenAI. But what is this technology beneath all of these eye-catching features? We can ask the model itself and set a limit of 50 words.
“ChatGPT is a large language model developed by OpenAI that can generate human-like text. It can be used for a variety of tasks such as conversation, language translation, and content creation.”
Large language models incorporate deep neural networks with a massive amount of parameters (ChatGPT’s underlying model has 175 billion parameters), to feed inputs through a multi-layered set of semantic associations to estimate the ‘right’ answer. ChatGPT builds on this structure by adding natural language processing (NLP) techniques such as contextual learning whereby the model retains the conversation and previous information it has received from the user to generate more accurate responses. Learn more about ChatGPT here.
Using ChatGPT in international arbitration
Now being somewhat familiar with what ChatGPT is ‘under the hood’, the most important question for international arbitration practitioners is how might it benefit our work?
Let’s look at a number of generalised use cases and then discuss their potential utility for practitioners:
Language translation: ChatGPT is well-suited to providing translated text at an accuracy that matches or exceeds other free tools like Google Translate, with the benefit of being able to fine-tune the model on where it has made errors.
Text summarisation: The model can be fed articles, essays and other longer documents and provide tailored summaries based on constraints set by the user (word count, inclusion/exclusion of certain aspects of the content etc.)
Text generation: The model can draft text of variable length and style on an infinite number of topics, and can be ‘taught’ about topics it is not familiar with. As alluded to above, some of the most eye-catching examples of this in the media have been poetry, songwriting and mimicry of famous individuals’ styles.
Question answering: As a chatbot, ChatGPT is specifically tailored towards this use case, but when answering on technical subjects it may not have been trained on, it may use its ‘imagination’ to estimate the right answer, and therefore describe things that have not happened or are not correct.
In the context of international arbitration, diverse language skills are always in high demand and thus the ability to gain access to high accuracy, and tuneable translations for free or for a low subscription amount (as part of any future paid ChatGPT plan) is valuable versus the cost of traditional services. The same value could further be derived from text summarisation processes, which would serve as an efficiency multiplier for junior and mid-level practitioners in legal research and the production of relevant case notes, allowing them to focus on more critical tasks within a dispute.
In terms of assisting in legal drafting itself, ChatGPT is not trained to know legal principles or to apply them but it will estimate what it sees as a logical answer. Thus, ChatGPT could save time for a practitioner through its responses being used as a rough skeleton to draft within, but its utility would be limited beyond that.
Is using the model as a ‘legal’ Google possible? As mentioned above, ChatGPT is designed to react like an advanced chatbot and therefore seeks to provide highly accurate estimations of correct responses to users’ questions. However, because by its construction, the model is only using probabilities to produce the ‘right’ answer and does not cite its sources, ChatGPT should not be leveraged to assist practitioners in answering questions that the user themselves does not know the answer to.
In conversation with Tim Harrison, Founder & CEO of Arkus Consulting, an independent eDiscovery service provider, I asked what he viewed as its most promising features and biggest drawbacks. Specifically for eDiscovery and document review tasks, Tim sees NLP models like the one within ChatGPT as the next step in improving efficiency and minimising human review in the years to come. However, due to the fact that ChatGPT produces responses on a heavily generalised (and unknown) dataset, it cannot yet be trusted to work in an unsupervised way. For the global disputes industry, the highest levels of certainty are usually required and therefore it may be quite a few years before we see this technology tailored for legal nuances.
For Tim, the promise of the technology does remain though, even in use cases beyond document search and review. Two hypotheses come to mind, the first being internal information governance for companies, whereby a NLP model is searching not only for rudimentary keywords and conceptual clusters in employees’ communications, but even analysing their sentiment and flagging issues in real-time. The second hypothesis relates to legal predictive analytics, and leveraging how these models might be able to quickly find and weight commonalities with prior disputes to forecast legal outcomes. This kind of use case would obviously represent a boon for third party funders and large international firms in how they would allocate funding and resources to individual disputes.
Therefore, as it stands, ChatGPT may not yet be the revolutionary tool that gives arbitration practitioners superhuman productivity, but that point may be closer than we think. Many developers have already started to iterate on top of some of OpenAI’s many products, including ChatGPT, and we are already seeing useful API-based browser extensions and web applications. One of the most interesting of these is a web-enabled ChatGPT, which can use search engines to inform its responses and supplement search results. Versions of OpenAI’s GPT-3 model can already be accessed and trained on very specific datasets to create specialised use cases. We may very soon see models trained on case law of various jurisdictions that can produce relevant legal research almost instantly. The same may become possible for drafting, whereby models are trained using templates from a law firm’s previous disputes, and produce memos and pleadings in that style. As such, ChatGPT may just be the tip of the iceberg for AI-assisted working and the disruption of traditional legal practice.
Interested in learning more about the effects of new technologies for international dispute resolution? Watch our discussion on the current state of crypto disputes here!
