Can artificial intelligence be a justice superhero?

It is an open secret that the civil justice system in New Zealand is prohibitive to most of us due to high costs and long delays. Many trials last months and even years, with legal fees running into tens or hundreds of thousand dollars. For claims of less than $100,000, it is not worthwhile to go to court. Even if you win in the end, you may still end up losing money because it is unlikely that the court will order the losing party to pay for your full legal costs.

As a result, more and more people come before the courts unrepresented, but they face significant hurdles due to a lack of legal knowledge. From time to time, judges have to lean in to ensure that the unrepresented party has a good and fair chance to present their case, resulting in longer trials and increased costs to the other party and the government.

One way to help self-represented litigants is to use chatbots and other AI applications to offer basic legal assistance. In this regard, there have been a number of interesting initiatives overseas. However, so far they have proved to be of limited success, mostly because law presents unique challenges to AI. Generally speaking, AI needs large amounts of labelled examples to learn from. In law, such training datasets simply do not exist. Now, with the arrival of GPT-3, created by research lab OpenAI, robot lawyers may become reality in the not too distant future.

GPT-3 stands for “generative pre-trained transformer version 3”, a language model that uses machine learning to produce human-like text. This model is trained on hundreds of billions of words, and has 173 billion parameters. Unlike other language models, GPT-3 can do a variety of tasks with a non-coding user providing only a few illustrative examples and simple instructions in plain English. For example, Michael Tefula, an associate at a London-based technology investment company, asked GPT-3 to translate legalise into plain English. Here is the result:

Original text: “Upon liquidation of the Company, the Series A Shareholders will receive in preference to all other shareholders an amount in respect of each Series A Share equal to one times the Original Issue Price (the “Liquidation Preference”), plus all accrued but unpaid dividends. To the extent that the Company has assets remaining after the distribution of that amount, the Series A Shareholders will participate with the holders of Ordinary Shares pro rata to the number of shares held on an as converted basis.”

GPT-3: “If the startup is wound up, the Series A investors will be paid back at least what they invested and they will also share any leftover assets with ordinary shareholders.”
Conversely, GPT-3 can also turn plain English into legal language (Credit: Francis Jervis at Augrented).

Original text: “My apartment had mold and it made me sick.”

GPT-3: “Plaintiff’s dwelling was infested with toxic and allergenic mold spores, and Plaintiff was rendered physically incapable of pursuing his or her usual and customary vocation, occupation, and/or recreation.”

Another way in which GPT-3 can help self-represented litigants is to summarise previous relevant judgments.

To test its capabilities, we provided two learning examples and then asked GPT-3 to summarise the following paragraph taken from a US judgment:

Original text: “The next question is whether the evidence demonstrates that the claimants' method of reducing the data collected has been sensible and proportionate. In my view, the evidence demonstrates that it has been. Only two further folders/file paths, beyond those originally identified by the claimants' custodians, have been subsequently disclosed. That is a statistically negligible amount of further disclosure, thereby demonstrating the reasonableness of the claimants' original approach.”

GPT-3: “The claimants' method of reducing the discoverable data has been sensible and proportionate.”

Currently, the Court of New Zealand’s Rules Committee is consulting publicly on how the existing court procedural rules can be amended to address the access to civil justice problem. In addition to tweaking the current procedures, it seems worthwhile to consider how to incorporate AI and other modern technologies into our judicial system.

Of course, more work needs to be done to make AI applications useful in practice, and the advice given by a robot lawyer may never be completely accurate. But that should hardly be a reason not to adopt AI. After all, human lawyers make mistakes too. Given the rapid development of AI, it seems possible, even likely, that GPT-3 (or its later, more powerful successors) could be a solution to the access to civil justice problem.