The Singapore Law Gazette

On a Knife’s Edge: Generative AI and Using It in Legal Practice

There’s a new kid on the legal block: GPT, crafted by OpenAI. Since ChatGPT 3.5 first launched on 30 November 2022, the new kid has received quite the range of responses: confusion, elation, ambivalence, fore-telling of the end-times, and my personal favourite: bemusement. Like all new kids, it is taking a while for everyone to warm up.

With this new kid’s digital prowess, and uncanny likeness to human communications, GPT has certainly stirred controversy everywhere, from New York courtrooms to hawker tables. While some herald it as the messiah of a new age of legal practice, some as an erstwhile saviour of late-night drafts, others see it as just another tool that does not perform very well anyway. So, where does the truth lie?

This article takes the view that such discussion is probably best undertaken elsewhere, by brighter and sharper minds than the author’s. Instead, this article takes the view that quite like a chef and their knife, a chef would not refuse a knife that keeps a keen blade, and that allows them to perform their craft with more refinement and efficiency. Most discussions about GPT’s use in practice veer into another territory: whether such a tool takes practice out of the domain of client-centric, bespoke service, into that of the mass-manufactured – ultimately insensitive to the needs and desires of individual clients.

It is how the blade is used rather than the blade itself that changes the complexion of practice. It is perhaps more telling that generative AI and large language models have engendered such a huge reaction in lawyers. Other professions have undergone sea-changes with the advent of new tools. It is now our turn. So, without further ado, let us dive right into how to wrangle the best out of GPT and to avoid its pitfalls. Much of the following would also apply to the use of other large language models.

The Cloze Passage and Human Feedback

As we dive right in, one principle to bear in mind is that GPT was not trained to understand veracity, truth, and accuracy. Imagine a primary school-going child practising cloze passages (remember filling in the blanks?) day in and day out, and finally becoming a master at those cloze passages.

Do we expect the child to understand veracity, truth, accuracy, and more importantly how to cite, and verify provenance?

I’m sure most of us would not. To extend the analogy, while GPT has been trained on comparatively gargantuan volumes of cloze passages, it is also at a similar point in terms of understanding those concepts. After being trained on cloze passages, GPT is further trained by providing it with feedback, encouraging some behaviours and answers, whilst disincentivising others. We should interact with it accordingly: as a text generator that has learnt the patterns and shapes of human writing, syntax, semantics, grammar, and information that we deem to be knowledge.

DOs for the Tech-Savvy Solicitor

1. Confidentiality and sensitivity: Treat GPT like you would any other Software-as-a-Service (SaaS). Use GPT only for non-sensitive, and non-confidential tasks.
2. Initial drafts: Turn to GPT for first-cut drafts, much like you would of the most junior member on your team.
3. Brainstorming: Turn to GPT for brainstorming sessions when your neurons play hide-and-seek, or simply go MIA on you.
4. Give clear and precise orders: The clearer and more precise your instructions, the better GPT behaves.
5. Give context: Lawyers and other professionals love the phrase “it depends” – and it really does depend sometimes! It is the same with GPT: how well GPT performs depends on how much context you provide it.
6. Research assistant and buddy: Use GPT for research. When you need to conduct targeted research, GPT can be your familiar. You can ask it to provide information about legal concepts and principles, and it can also give you a preliminary overview. Be wary. Citations can be correct, but GPT can also entirely misunderstand the ratio of a case, just like a budding lawyer or law student might.
7. Learning buddy: GPT can give a rough idea of the lay of a land of a topic, or even come up with suggestions for subtopics that one should cover to understand a particular topic. As always, ensure that the output is verified by an expert within your team, or use it to supplement your own learning. In verifying and coming across information repeatedly, one is likely to gain more fluency in the topic.
8. Efficiency expert: Delegate non-sensitive tasks, like summarisation and information extraction to GPT, and watch your efficiency soar.
9. Talking global: GPT is useful for both translation, and also editing. From Mandarin contracts to Spanish injunctions, GPT can assist. You may also ask GPT to provide editorial suggestions, and point it to specific style guides. Some of my personal favourites: Ernest Hemingway’s writing style (similar functions already present in some erstwhile browser plugins like Grammarly and the Hemingway Editor), and Tim Peters’ The Zen of Python.
10. Validate, validate, validate: Like checking your attire before court, always validate GPT’s output. Twice.

The GPT Red Flags: DON’Ts for the Digital Advocate

1. GPT is not your secret keeper: Would you shout confidential, client, or personal data in a crowded elevator? No? Then do not put it into GPT. The same would apply for any SaaS.
2. Do not delegate finality and veracity: Relying solely on GPT for tasks that require finality and veracity is like blindly delegating decision-making to an assistant. GPT is a tool to assist and generate information, not to make ultimate legal determinations or verified citations.
3. Do not abdicate your advisory responsibilities: GPT may be able to pass Bar examinations, but does not hold a practising certificate. Nor does it actually comprehend the human condition. Avoid using it to provide direct legal advice.
4. Research buddy, not boss and intellectual soulmate: GPT is good, but it should not be your only source of research. GPT should supplement and enhance, not replace traditional legal research.
5. Do not rely on GPT to understand ethics: Assuming GPT understands legal ethics is like assuming children understand legal ethics after filling out many legal-related cloze passages. You may get ethics-shaped sentences, but that does not equate to an understanding of ethics.
6. Do not forget the human touch: Before using any GPT output, all output should be reviewed by a human for accuracy and relevance.
7. Do not overlook the potential for bias: GPT can have biases, just like the training material provided from the internet by its human masters. While OpenAI has done a commendable job in filtering biases, GPT is understandably not going to be perfect.

Prime Use Cases

1. Document wizard: Let GPT craft the initial drafts of your contracts, pleadings, and documents.
2. Research assistant: Need an overview or quick summary on a legal topic? GPT’s got your back. Remember to prompt it for citations. While they are not guaranteed to be correct, they will probably give you good indicators of what to read, and what pitfalls to avoid.
3. Communication wingman: Drafting client communication becomes easier, subject to review and modification by a human lawyer.
4. Idea factory: Use GPT to brainstorm ideas, or to bounce ideas off when no one else is available. Prompt it specifically to challenge you, or to be more creative. You can ask it to be a devil’s advocate, or even throw you off with out-of-the box ideas, e.g. the famous musician Brian Eno’s Oblique Strategies.

Conclusion

GPT, in the vast legal kitchen, is like a reliable mechanical crew member and appliance. It supports the main crew, can play many roles, but is not the heart, soul and orchestrating mind of the crew or the operation. Use it wisely, and you could serve many more dishes to many more satisfied guests. Just ensure it doesn’t accidentally end up as a dangerous tool, flailed about on a whim.