How long before machines can write your contracts?

Let us begin with an anecdote.

A few years ago, the self-proclaimed gonzo data scientist and writer of writers Ross Goodwin went on a road trip. He outfitted a Cadillac with a surveillance camera, a GPS unit and a microphone. Then he drove the way from New York to New Orleans. While he drove, all the attached sensors transferred input data to an AI writing machine that previously had been trained on a large body of classic world literature.

Based on what the camera filmed, the microphone recorded and the GPS reported, the machine would print sentences on a long scroll of receipt paper - just as any gonzo writer would scribble down what they saw along the way. Line by line the manuscript for a novel emerged - the first novel to have been written by a computer.

1 the road, as the novel was called in the spirit of Jack Kerouac, is a barely consistent narrative with surrealistic impressions like:

‍

"Blue and white background stood in the middle of the street, but the stars were already still and hard, and the sky had been floating on the walls. Sun shining on the sunny stove seemed to be hanging on the corners of his eyes."

‍
Poetic, brilliant, but also down-right bizarre.

The novel is just one of Goodwin's grotesque AI-art projects. He also masterminded a computer that masterminded a touching science fiction short film, and he programmed a poetry project on Trafalgar Square. Elsewhere, we have seen computers produce catchy music, contribute to the Harry Potter-universe, and generate art that goes for thousands of dollars on auctions.

For long, art has been considered the crown jewel of human creation. So now what? What does it mean that computers can be used to produce award-winning novels and films shown at the Sundance Film Festival? Can they make all kinds of texts? Can they make this article you are reading right now? (Am I on the road to unemployment?!). And, more related to Contractbook, what about all your legal documents? Does it mean that we are about to fulfil a decade-old dream of fully-automated contract creation?

That is what we are going to take a closer look at in this article: Can machines draft our contracts for us? And is that even desirable?

‍Natural language processing

‍Just to finish the anecdote about Goodwin. His writings have an uncanny artistic attraction that is very hard to comprehend but also impossible to dismiss. That said, the works are incoherent, monstrous, and very long from the average Booker Prize winner (or even just your plain beach read).

However, his technology is already used by most people and to some extent, to write contracts. Goodwin used something known as statistical language modelling, where the machine predicts the next word in a given sentence based on what precedes it. Feed the machine a headline, a word or in this case, data from a GPS/camera/microphone, and it will be able to complete the sentence and generate text based on a probabilistic model. This technology is already widely used in spelling correction programs, to make machine translations, text summarization, word suggestions and much more.

Statistic language models are crucial tools in the field known as natural language processing which is the branch of artificial intelligence that helps computers to read, understand and interpret human language. It can roughly be divided into national language generation which is the process of turning data into natural language. And natural language understanding which is the process of making computers understand natural language.

Natural language processing has been around for a long time, but it took a significant leap forward in the late 1980s with the steady increase of computational power and the introduction of machine learning algorithms instead of handwritten rules. In the past years, the method known as the neural language model has become more popular because it is able to make more generalized predictions. We will get back to that.

AI text generation - augmented and enhanced

So, agreements are (most likely as you are reading this) being drafted with such advanced technologies because they can be used to augment humans and enhance their writing skills to produce texts both faster and more grammatically correct. One example is Google's AI-powered Smart Compose that suggest words and phrases while you are writing. The complexity and smartness of such a technology can vary a lot depending on how many data points it includes and, of course, what data it is trained on.

Another great example of artificial intelligence technologies used in the contract creation process is those used to screen and review documents. For example, as a research programme involving the Israeli startup Law Geex showed, AI's are both faster and more precise contract reviewers than their human counterparts. Such technologies can be used to identify problematic clauses and find contradictory language in contracts, produce a contract analysis and even suggest various outcomes. That is useful both in the initial drafting process and the following negotiation process.

So machines are, of course, highly useful and already used widely to augment contract creators and enhance the legal writing. Just as this article is written by a human but has also been enhanced by Grammarly's AI.

But can machines make contracts autonomously and totally from scratch? And are we really nearing the "dawn of fully automated contract drafting" as this paper suggest?

The future of autonomous writing tools

In the past years, we have seen machines compose opinions for newspapers, do content marketing, create digital media content and much more. The quality has always been inconsistent.

In one end, we have experienced useless, inappropriate and awkward chatbots, Holocaust-denying robot tweeters and sexist slogan generators. We have also seen flawless robotic sports journalists on Washington Post and machine-powered Reddit-users are going unnoticed for days. As stated on the involuntarily comic disclaimer for the AI content generation tool Kafkai: “You will receive articles that can be used almost immediately (...) and you will receive articles that are unusable. Most articles will be between those two.”

That statement sums up the status quo pretty well, but it also explains the problem: That contracts, of course, require a lot more certainty than other texts. Such as this mediocre blog post. At least we hope that your contracts are a bit better than somewhere in between perfect and unusable. With contracts, you need the best of the best.

The avantgarde in AI writing tools is delivered by the San Francisco-based and Elon Musk-funded search laboratory, OpenAI and their Generative Pre-trained Transformer GPT-3. It uses deep learning to produce human-like text and has impressed with some very coherent and engaging writings. No wonder. The machine is trained on an enormous data set and has a capacity of 175 billion machine learning parameters (the second largest model has 17 billion!). It was released this summer and has been hyped ever since. So if there is a game-changer on the road towards fully automated contract creation, then this is it.

GPT-3 and contracts

It is assumed that even though the GPT-3 is not specifically trained on legal language, it can still be useful in legal. Because of its capacity, it would also be able to learn the skills of contract generation quite quickly anyway. In fact, it is already being used by the legal tech company DoNotPay or in Hackathons. The technology could then be used to not only flag risks in a contract but also suggest alternative content, structure metadata and generate contracts all together.

However, the challenges right now are multiple. First of all, GPT-3 costs between $4m-$12m to train, so it is still very hard for such a technology to return the investment. Secondly, the material it is trained in should be very controlled because of the nature of contracts.

These machines learn with data, so they are dependent on correct, high-quality data. Since the machines learn from both good and bad legal writing and is often unable to distinguish between the two, they are just as likely to reproduce good contracts as well as bad contracts.

That leads some to believe that full automation will never materialize, while others believe in the sophistication of the technology. However, as a paper from Duke University suggests, good automated contract writing might be within reach. Still, it requires that “certain non-technological obstacles must be overcome: 1) the collection of contract performance data, 2) publication of private contracts and their corresponding performance data, and 3) changes in the ethical restraints on computer usage in legal practice. This will be a lengthy process, but our suggested policy initiatives may.”

Where does it leave the dream of fully automated contract drafting?

The dream of a fully automated contract drafting process is decades old, but we are still a bit from being there. While the most advanced technologies are being implemented, they are mostly used for augmentation and skill enhancement. That is awesome. But the fully-automated contract generator is not only too expensive at the moment, it is also too imprecise for contract generation. Bear in mind that it is still not widely applied in content marketing and journalism. I am still human! (I guess a robot could also write that, but trust me, I am).

That said, you can prepare yourself for it, and you can leverage existing technologies to make contracts smarter and more autonomous.

Automated contact drafting technologies became popular in the 1990s where most law firms used drafting software based on pre-coded questionnaires to auto-generate standardised contracts. These tools were hard-coded if-then logic trees used to increase efficiency and prevent users from making mistakes. Such technologies are still in use today as they are both easy, cheap and efficient. While they started as simple solutions for lawyers, they are now offered in more sophisticated versions that are so well-designed, intuitive and client-centric that they can be used by anyone. Like our contract drafter solution, which automates the contract creation process for modern businesses.

Compared to the dream of fully-automated document creation powered by AI, such tools are inflexible and difficult to customize. They are made for very standardized documents and not for the more idiosyncratic projects or areas where the law changes often. To meet those criteria, you need machine learning software that is able to draft contracts and chose the right clauses from across multiple contracts to approximate a single standard document.

To prepare yourself for this technology, you need to harness your data and apply a modern contract lifecycle management software. As concluded in the research paper mentioned above, such software must be trained on good legal writing. And to identify the good legal writing, you must have a clear idea of the performance of your contracts.

Contract automation: What you can do here and now!

Okay, and now comes our advertisement. Come on, you did not think we made this long piece of content marketing without trying to sell you something, did you?

We believe that the best way you can prepare for a future of fully-automated contract creation is to organize your contracts, use a data-rich format and establish a great overview where you can leverage the data to track the performance of your contracts.

All of that is something, Contractbook’s contract lifecycle format facilitates. And while you wait for more advanced technologies, we have several contract automation solutions that are able to automate the entire contract flow in your business. Like we have done for Dreivers.

To get started in your journey towards AI, sign up today and use our smart new Gmail Importer to gather your contract legacy in one place. Meanwhile, automate your entire contract lifecycle with deterministic logic. It is cheaper - and it works!