All Aboard The Next LegalTech AI Hype Train, The GPT-3 Express!

Brad Newman, Associate Director of Practice Innovation Services at Cooley LLP writes an interesting take on GPT: "Unlike, say, the fever dreams caused by IBM's Watson (RIP ROSS Intelligence) or your lawyer's latest journey into the Metaverse, we may see real advancements in intellectual efficiency from legal applications powered by GPT tools, in particular those built on Open AI's GPT-3 series."

All Aboard The Next LegalTech AI Hype Train, The GPT-3 Express!

Why You Should Read This

Simply put, the holidays are coming up and you will want some convo subject matter to impress your co-workers or lull your family into a near-slumber, avoiding political discussions.

Also, this technology may change key facets of the economy of legal information. If that means anything to you, whether as a lawyer, legal business professional, investor, etc. - you should probably read on.

The Hype May Be Real This Time

If you're reading this on LinkedIn you've probably already seen folks fawning over DALL-E 2 (creates images from text, for example the image above) or, more recently, ChatGPT (conversational dialogue generator, including the ability to "answer follow-up questions, admit its mistakes, challenge incorrect premises, and reject inappropriate requests.")

Below, I argue that unlike, say, the fever dreams caused by IBM's Watson (RIP ROSS Intelligence) or your lawyer's latest journey into the Metaverse, we may see real advancements in intellectual efficiency from legal applications powered by GPT tools, in particular those built on Open AI's GPT-3 series models. However, there are likely obstacles that developers will likely have to overcome before they can expect lawyers to adopt, or for reasonably minded Associate Directors of Practice Innovation Services to recommend the trialing of, such applications.

The Tech

Both DALL-E and ChatGPT are finely tuned versions of Open AI's Generative Pre-trained Transformer (GPT) series of neural network machine learning model.

Open AI's GPT-3 has 175 billion parameters and was trained on 45 terabytes (45,000,000,000,000 bytes) of text (source) (composed of text from crawling innumerable websites, the text of web pages from all outbound Reddit links from posts with 3+ upvotes every Wikipedia page, and public domain online books) to "learn the probabilities of a sequence of words that occur in a commonly spoken language (say, English) and predict the next possible word in that sequence" (source). For context, the complete works of William Shakespeare is about five megabytes, or 0.000001% of the 45 TB used to train GPT-3.

A basic example of GPT-3's predictive ability:

Basically, Open AI has built tools that come as close as we've seen to software that can understand language in a way that allows it to generate new content in response to writing prompts, instructions or existing prior text.

The GPT-3 demo that first caught my attention regarding potential in the legal space is GitHub's Copilot, which utilizes Open AI Codex, a version of GPT-3 trained on gigabytes of source code in a dozen programming languages. When provided with a programming problem in natural language, Codex can generate solution code. It can also describe input code in English and translate code between programming languages (source).

Just to be clear: what you see above is a plain language description by a programmer of what they want a function to do - and GitHub's Copilot just... creates it! Truly amazing and one can easily envision how this might apply to drafting contracts, pleadings, etc.

Here Comes ChatGPT

ChatGPT adds a new layer of interactivity. You can ask it for ideas for planning parties, to help you understand how you can best help your mother deal with a late-stage cancer diagnosis (a personal example), to write code, to explain concepts, and more.

Consider this example:

Now how about a simple drafting instruction:
 

 

But what about reviewing existing provisions?

Below, ChatGPT is asked to debug a piece of code and it proceeds to explain the bug and how to fix it:

Pretty rad.

One can easily see how GPT-3 could be adapted to, say, identify and explain contractual drafting errors from a given text. Some say Code Is Law, but does that mean Law Is Code - at least to the extent that GPT-3 can be fine-tuned to similarly explain inappropriately placed or drafted Change of Control or Non-Solicitation provisions? (I tried something similar but with a poorly drafted Entire Agreements clause and didn't get anywhere - yet.)

The Hype is Real - But Can It Be Utilized?

We already have a few companies that have made their GPT-3 (or similar) powered projects public such as Spellbook by Rally (contract drafting) and Harvey (research memo / outlines / letters).

I haven't had a chance to get a demo of Spellbook but I have kicked Harvey's tires. It is no doubt impressive. I asked it to draft a demand letter to a contractor who botched the installation of a garage door and it did a pretty darn good first draft, case citations and all. Whether those case citations are appropriate or still current - that would require more investigation (one of the obstacles I discuss below).

That said, it was cumbersome having to type out all the facts of my case - and I had to make sure that I conveyed all relevant facts. I'm curious to see if ChatGPT will allow for a more conversational - and complete - intake process that will help ensure all relevant facts have been supplied.

I particularly enjoyed this envisioned use case for ChatGPT by Joshua Browder:

If not negotiating directly with each other, a neat implementation for law firms might be content- and context-aware negotiation simulators ("Simulate a negotiation where you're representing an investor trying to include a high liquidation preference multiplier...")

Maybe Just Leave Document Assembly to the Web Forms

GPT-3-enabled tools may be great at explaining concepts, finding relevant concepts, and perhaps analyzing and generating discrete contractual provisions or briefings. However, I think web form-based document assembly tools like ContractExpress will continue to have their place in a lawyer's automation toolkit.

The efficiency of filling in a few text fields and flipping a few boolean switches and generating precise results based on a carefully maintained form ... I just don't see GPT-3 making that process better.

On the other hand, I can see how a finely tuned ChatGPT tool may help attorneys quickly find the best resources available (e.g. "Tell me where I can generate a mutual non-disclosure agreement?", "Who at the firm is fluent in English and Japanese?", or "How many private financing deals have we done in the biomedical space where a pay-to-play provision was included?" - to be clear, technology to field these natural language queries exists but not (easily) under one hood nor in the same universe of potential comprehensiveness and usability).

Features Necessary to Encourage Adoption (IMHO)

(Self-Service) Fine Tuning on Custom Corpus

Lawyers understandably don't like using forms or precedents without knowing who wrote them and in what context. So, any legaltech tool utilizing GPT-3 modeling needs to ensure that it can be fine-tuned on a firm's own corpus, such as precedents, intranet pages, and any other data deemed relevant. Providing the ability for the firm to do this itself is a bonus.

Reference to / Integration with Specific Training Elements

Much like the current iteration of "AI-enhanced" due diligence software, work product generated by GPT-3 will still require review by a real human lawyer. Accordingly, every effort should be made to enable users to easily check the relevancy and currency of case citations. If possible (and I'm not confident this is actually possible), the document(s) - or at least a specific library or subset of documents - used by the tool to generate, e.g., a contractual provision should be explicitly referenced/linked.

Without this configurability and transparency, vendors may find it difficult to make inroads with firm decision makers (though this is not an unfamiliar disposition).

---

Brad Newman

Share post:
Legal.io Logo
Welcome to Legal.io

Connect with peers, level up skills, and find jobs at the world's best in-house legal departments

Legal.io Logo
Welcome to Legal.io

Connect with peers, level up your skills, and find jobs at the world's best in-house legal departments