Waymo Shows Legal AI The Way – Artificial Lawyer

Waymo took over 20 years to finally deliver safe, accurate, self-driving vehicles. In San Francisco they’re now 100% for real and Artificial Lawyer travelled in one. Legal AI can learn a lot from this journey. Here’s how.

Show Us The Way, Waymo

Waymo can trace its roots back to around 2005 when a team of Stanford University researchers centred on the idea of an autonomous vehicle came together in California. In 2016 the group officially became Waymo, an offshoot of Google and Alphabet.

At around the same time, back in 2016 Artificial Lawyer started as a blog about the arrival of ‘legal AI’, based on the belief that this technology would transform the legal sector.

Both Waymo (and other companies), as well as the pioneers of legal AI, have since spent many years and plenty of money trying to deliver on their original promises.

For Waymo the promise was to provide a safe and reliable autonomous vehicle that could navigate even the complex and sometimes congested streets of a city such as San Francisco – completely on its own.

For legal AI….well, it depends on who you speak to as to what the goal has been, but for AL the goal has always been crystal clear: to reach a point where AI is able to absorb entire legal workflows, where a human lawyer is primarily a quality checker at the end of the process. Naturally, this is for the most routine tasks, albeit processes that can be quite complex and involve plenty of dimensions and unstructured data. We still need lawyers to handle a wide range of anthropocentric needs – and likely always will.

In short, both problems are very hard. But, as AL can tell you, Waymo has solved its challenge and did so with aplomb. The Jaguar – which Waymo works with – drove safely, assertively, and accurately. I felt totally safe.

Moreover, in about 10 seconds it already felt entirely natural to travel this way. Perhaps unconsciously this site could sense the deliberateness of each manoeuvre of the autonomous vehicle as it wove smoothly through traffic in that very American way of using the road.

By the time it had pulled up outside my hotel – (which incidentally was accessed by a key card and then ‘operated’ via an iPad in the room) – I was 100% sold on self-driving cars, just as I was 100% sold on iPhones within an hour of unboxing my first ever smart phone many years ago. It just worked. It felt intuitive. It felt reliable.

Now, we cannot always say the same of legal AI. For many use cases you do indeed get that instant sense of success and accuracy – which is one of the reasons why genAI has been so rapidly taken up by so many lawyers. But, ‘autonomous legal work’ is a hard problem, and also stretches out into many directions, sometimes with exponential increases in difficulty.

For example:

Summarise this short client email – easy.

Write back a quick thank you to the client – easy.

Now look at the legal question in that email and find the case materials that relate to the issues – first from general legal sources – medium difficulty.

Now do the same, but from this law firm’s own data from the last ten years, but avoid files that should not be viewed by those not permitted. Also, don’t just drag up everything. Only show the top two or three past matters that really match the ‘ideas’ within the client email – getting harder now.

Now draft an appropriate legal document that encapsulates the legal needs – at this stage in the process – of the client, making sure you use the right data, use ‘what’s market’, stay within the firm’s own way of working on such matters, and don’t miss out any key clauses, dates, names and other data, and make sure every single word is meaningful and relevant – rather than just making it sound as if it’s correct – very hard.

Now, without asking for any human lawyer help, engage with the client, take their feedback on the document and make your own judgment as to what needs to happen next and make any new changes. All the time, remember that this needs to be near perfect accuracy and contain no mistakes or omissions that would impact the end outcome – very, very hard.

But, is driving in the hyper-complex world of city traffic, with other cars, motorbikes, pedestrians, changing road layouts, potholes, sometimes wet weather, mixed with the inevitable randomness and errors of human drivers in the vicinity, as well as driving among other autonomous vehicles, more complex?

They’re different types of very hard problem: one in the physical, structured environment of the 3D world, the other in the unstructured and subjective world of human meaning, i.e. ‘words on a screen’.

Waymo has solved its problem, albeit with 20-plus years of work and billions upon billions of dollars spent solving it. Legal AI has not had quite as much time, nor has it had the same level of investment.

AL asked Gemini (part of Google) how much has been spent developing Waymo. The answer was: ‘Some reports have estimated Alphabet’s total spending on Waymo (including external funding) to be around $30 billion by July 2024. This suggests a significant portion of the raised capital has been continuously invested into the research, development, and scaling of Waymo’s autonomous driving technology.’

Then if you add in the investments made by many other companies in the same field, e.g. Uber, Tesla, and others, as well as investments into some of the underlying technology that supports these cars, from the latest Lidar to the ‘physical AI’ backbone, then we have a gigantic investment globally.

But, it’s still less than what it cost, (in 2023 dollars), to put humans on the Moon in 1969, which totally was around $300 billion.

How much has been spent developing legal AI? It’s hard to tell, plus money alone isn’t always a perfect measure of likely success – although it certainly helps. Moreover, genAI’s ‘backbone’ today is not made by legal tech companies, the LLMs are from OpenAI, Google, Anthropic and many others. We piggy-back on them. So, our investment is in fact much smaller.

Nor can a legal tech company truly control what OpenAI, or other model makers, do, even if one or two folks out there have Sam Altman on speed dial. They can refine outputs, they can combine perhaps with some much smaller models and other large models, and they can be really good at RAG and ‘pointing’ the LLMs involved at a very well curated collection of data….but when it’s for the most complex use cases in legal, we are still stretching the state of the art to get to where we need to be.

That is, stretching if we aim for ‘autonomous legal production’ – as opposed to AI just acting as a friendly little assistant that gets us only a small part of the way along a task and even then we don’t fully accept it on its own.

What Does This Tell Us?

Personally, I found the Waymo experience a real watershed moment. Like everyone else, I had watched with interest, then disappointment, over the years as promise after promise about self-driving cars was made and then failed to deliver.

But, now, in California, this promise has been made real. And it was empirically testable. I sat in a Waymo and it drove me home, safely and securely. It worked, for real.

And that gives Artificial Lawyer a surge of optimism about where legal AI is heading as well. While many lawyers are happy with LLMs and related applications providing some kind of half-way outputs, which help, but can neither do a whole task, nor are totally reliable, the goal has to be on a Waymo scale, on a Waymo level of ambition: to enable an entire, highly complex, legal workflow to be safely and accurately automated.

I am sure we will get there.

The only question is: how long will this take? But, we will get there.

After that, the next question is: what impact will this have?

P.S. it’s worth mentioning that some estimates suggest that Waymo’s automated taxis already account for about one quarter of all the rides in San Francisco, and they only have a fleet of about 300 vehicles there at present – although they operate much longer hours than human-driven taxis.

—

Many thanks to Todd Smithline for AL’s first Waymo ride. You can find more info about the company here.