ChatGPT is a language model developed by OpenAI. It is designed to understand and generate human language, and can answer questions, generate text, and engage in conversations. It is an AI model that can generate human-like text. (Yes, it wrote this for me). It uses OpenAI's GPT-3 model and has a vast knowledge base meaning it can understand and respond to a wide range of topics.

In brief, it's a tool that you can prompt with a variety of inputs to get specific outputs, whether that is drafting a LinkedIn post, writing an adventure story with your dog as the main character, summarising text so a 5-year-old can understand it, or negotiating your phone bill down. It is still in beta and is open to all to test at https://chat.openai.com/chat with plans to release GPT-4 in the future, which will be a much more powerful model. It's not the first application of machine learning to generate content but it is the best implementation and has gathered the most attention across lots of different industries.

There has been a lot written about ChatGPT over the past few weeks, with people heralding an end to the age of trusted information on the internet and others saying it will be forgotten about in a few months as another gimmick marketing ploy. The truth, as usual, sits between the two in my opinion.  

This is a substantial step forwards in AI development, and it is also within a "human-like" chat interface, which means not only is it impressive, but it is easy to use and built in a way that allows people to create funny / insightful conversations with a robot that can then go viral on social media. It can process responses much quicker and use the context you give it to tweak those responses, one of the main skills with ChatGPT is prompting it in the correct ways to give you a reliable or useful answer.

Have a play around with it and you will soon come to the conclusion that this thing is pretty good, you might even chuck in some of your usual tasks like "draft an email out to a client that states that we've looked through the documents sent over and the main risks are X, Y and Z" or "write me a LinkedIn post about how it is crucial to organise and structure your precedent contracts before purchasing a CLM tool", you might then add "make it less corporate-y" when you get the response. You aren't going to immediately copy this and use it, but it does give you a decent starting point; it can be helpful getting rid of writer's block or just summarising a lot of information down into bullet points.

Since December, my actual main use has been copying blogs or articles into it and asking for a 5-line summary, not only do I now 'read' more in less time, but this is a pretty low risk task. If it misses some key points the worst that happens is I look daft when I make conversation with someone more well-read, and less lazy, than me.

We have also been testing it for specific legal use cases, testing how well it can answer a query about a contract or re-word a clause to be more supplier friendly. The results are impressive, this type of machine learning model is good at understanding the initial query and finding the relevant information. When re-drafting or drafting from scratch, it isn't always 100% accurate and doesn't always include every point needed, but it is a good first draft to work from. This is just in a test phase at the moment, but the GPT-3 model is finding its way into more developed product offerings aimed at legal, which we are looking into.

The reason for the fuss over ChatGPT is mainly because of its potential, but there is also a lot of shouting about the dangers. There is huge potential for large language models, put over the right dataset and monitored closely to ensure consistent and usable results. But there are real dangers to be conscious of, around ethical use of generative AI, over-reliance on unverifiable information and then the over-saturation of this unreliable information becoming further training data for other AI models. There is also attention being drawn to the fact that ChatGPT will make up plausible sounding information to fill a prompt; the model is interested in following a pattern of language, not being factually accurate, basically it lies.

One of the issues is that there is no predictability in what these models will come back with, no real consistency, you can ask the same question several times and get different answers. This is fine if you're on the spot trying to put your kid to sleep by reading them a story with them as the protagonist tackling a mighty dragon, but not so fine if you're trying to find out what steps you need to do as a director to file your company accounts. There are other issues as well of course. GPT-3 has over 300 billion words fed into it, but how many of those words were verified, true information, how much of the verified, 'gated', often paid for content on the internet has found its way into this open-source model?

Why does this matter for a law firm? Well, for one, maybe reading this piece means you can answer when someone says, "Have you seen this new AI thingy, apparently it is going to take your job?!". More seriously, it is worth understanding this because there will be issues caused by people using it, possibly instead of legal advice, either due to cost or time pressures. Also, because it is an interesting development in AI that is focused on text and language, which is the majority of a lawyer's work and it may be able to help further reduce some of the lower value work in a lawyer's day.

This is a very early example of another tool that could support lawyers, but it is also may enable us to take the tools we have, a step further. We can already do a lot to reduce the admin burden that a lot of legal work requires, but this may help add a friendly and intelligent veneer to some of our other tools. We can extract information out of contracts, but ChatGPT might make that a more conversational review stream between lawyer and machine. Legal research through PLC, or summarising regulations from the FCA, could be streamlined to take in specific questions and provide answers backed up by verifiable domain knowledge.

Whilst there is potential here, there is a lot of work to do to make this suitable for practice or to allow us to recommend its usage to our clients. As a leading firm within the use of technology in law, we can have a say in how this may be used across the industry, support start-ups with our domain knowledge and begin to test it within a safe environment and start applying it to problems where it is best suited. To do this successfully, we need lawyers within our firm to be interested and optimistic about the future of AI within law.

In playing around with ChatGPT a lot of use cases have come to mind. I asked my wider R&D team what their thoughts were and have put a few potential use cases below. 

MIKE KENNEDY – SENIOR MANAGER

Contract drafting 'buddy'.

In the future we will be able to specify what content the language models are trained on, or what content to give more credence to. This will enable us to plug something like GPT-3 into our precedent bank and within a document editor to support lawyers when looking for specific wording or clauses. Having the ability to ask a contract "Can you give me an indemnity policy that protects the seller more than this current one?" and being presented in seconds with a range of options that are not only in firm style but are collated from the firm's own pre-approved precedents, even if the specific clause you want does not yet exist. An effective implementation of these models could remove the need for things like clause or precedent banks if we can just target the model into verified information.

See below for a first attempt at ChatGPT drafting a Privacy Notice. The chat style interface means you can tweak this alongside the tool, for example adding in specific exceptions or further definitions.

ROBERT MALISZEWSKI - SENIOR LEGAL TECHNOLOGIST

One of the use cases that I can think of is using ChatGPT to rephrase paragraphs. Many non-native speakers may want to 'upgrade' their sentences with the AI support, to sound more sophisticated and formal. It is great for checking grammar errors and spelling mistakes, allowing users to get better at learning new vocabulary. I have tested ChatGPT as well as a translator and it gives much better results than any other tools I have used before.

Here, I have tested it for explaining grammar rules, which can really enhance process of learning foreign languages:

LEO TARN - LEGAL TECHNOLOGIST

ChatGPT is a great tool to utilise when updating and revising wording and there are many ways we can integrate this into our legal processes.

Gender Neutral Wording.

Despite the emphasis placed on Diversity and Inclusion initiatives, many law firms are falling short by using gender-specific language in their contracts.  Legal Cheek reported that nearly two-thirds of legal contracts between 2017 and 2022 contained gender-biased language, meaning that this is a prevalent issue.  We can use ChatGPT to address this issue quickly and easily by inputting contract wording and instructing it to make the language gender neutral.

Repapering Process.

Repapering is a time consuming, labour-intensive process and contracts can be complex and difficult to understand, making it challenging to identify and revise specific clauses or terms.  ChatGPT is a useful tool to utilise in this process.  All you need to do is input your contract wording and instruct it to replace specific words or phrases with updated language. This is an efficient solution that will save you time and reduces the chances of human error.

YEE XIAN KONG - LEGAL TECHNOLOGIST

Another useful utilisation of ChatGPT is analysing and comparing large amounts of legal documents, such as contracts and case law. I think that this skill is what commercial firms seek in lawyers, evidenced by the level of attention to detail and information assimilation required to be a 'good' lawyer. This smart chatbot tool can simplify this task by providing analysis and a quick summarised comparison between documents. Essentially, ChatGPT identifies relevant information (e.g. key terms and phrases, legal citations), patterns and trends (common themes, or arguments), similarities and differences, which are crucial for efficient legal research and decision-making. Legal professionals can then save time and effort, having an overview of the documents and subsequently elect appropriate next steps, be it further analysis or making a business case. See below for a simple example:

A commercial contract on the sale of a tech start-up:

A commercial contract on the sale of a mature tech company:

Analysis and comparison between the two commercial contracts:

HENRIETTA REDDING - LEGAL TECHNOLOGIST

ChatGPT also has the potential to provide a far more advanced model for AI contract review. Current contract review tools aim at extracting information, answering pre-built questions, or comparing versions; trained with machine learning they can be asked to extract specific clauses and values from the agreements. Whether this training is done by the creators of the tool or by the user, it is time-consuming and highly dependent on training data.

But ChatGPT uses a large language model, GPT-3, that removes the need to train. Using an unsupervised learning technique on a large dataset, that is later fine-tuned, the only user input will be exposing the tool to the dataset. While the model does away with the need to be trained, it also removes the requirement to correctly phrase queries so that the tool responds in the desired way. Instead, this model has the capacity to interrogate the dataset from the user's natural language questions and produce natural language responses. This expands the tool's response capabilities from just clause and value extraction to explanation and evaluation of the data.

As you can see on the example below, ChatGPT's beta version requires the user to provide an extract of the data they want to query. However, if developed to answer questions from a back-end data bank, the use of this tool in contract review could be far more efficient, specific, and accessible than any of the current tools on the market. It's worth noting that GPT-3 isn't the only AI on the market with a large language model that doesn't require training, but it does allow for a more conversational interrogation of the data. The use and development of this model in AI contract review has clear use cases in junior training and access to legal contexts.