Demystifying GPT and AI for lawyers with Mary O’Carroll and Cai GoGwilt

Foreign.

Hello and welcome to another episode of Legal Ops podcast where we talk about everything, legal operations, legal business and legal technology. My name is Alex Rosenrouch and I'm Elliot Lebue.

In this episode we interviewed Mary o' Carroll and Kai Gogwilt. They are from Ironclad. Mary is very well known. She is the former president of the Clock Organization and now the chief community officer over at Ironclad. And Kai is the co founder and chief technology officer at Ironclad. We spoke to Mary and Kai about GPT and sort of all things GPT and AI in legal and legal workflows. We wanted to speak with Kai in particular about what GPT is, what the different versions of GPT are and what it can do for sort of digital contracting and contracts. We wanted to sort of touch on what it's really good at at the moment and also what its limitations are. And then we spoke with both of them about the sort of hype. Like is the hype real here? Are we in? We in or approaching the peak of inflated expectations or is there something more substantial there? And then we focused on a bunch of challenges and as sort of mentioned what the limitations are but also of course where this might be going in the future. Excellent conversation, Alex. Thoughts?

Oh Elliot, it was a real treat. And you know, the reason why we wanted to have this conversation is because there's simply just so much information about GPT at the moment and also the information about GPT is changing almost on a weekly basis because it's getting better and better. So we actually wanted to get someone on the podcast that really understands this technology, which was Kai and look at it organizations that actually applying it to the day to day right now. We wanted to go a little bit further than just being able to read something online. We wanted to, like I said, talk to someone that's actually doing it right now in a real life scenario. The thing that really got me excited was the conversation around this, like inflated expectations. You mentioned it before, but the thing I took away from it was we actually just don't really know when that's going to end. I think in the podcast we talk about there's probably going to be another couple of months where those again expectations are inflated and we're not sure exactly where it's going to land. And look, another learning I had is we don't really know what the full potential is yet. And we talked about what it can do now. We even talked about what it could potentially do in the future, but what that future looks like. Is a pretty big question mark. And I loved the conversation about this idea around emergent behaviors. And we'll talk about it in the podcast. But effectively it's where a model, the LLM, which we'll talk about, starts to do things that it's not trained to do. It's effectively getting smarter, which was, again, like I said, a super fun conversation.

Yeah, and a couple of other things I'll just mention. So one of the reasons why I think this worked particularly well is that we had Kai coming at the technical aspects of this from a sort of very technical background. He's a physicist and engineer by background. And then we had Mary able to talk to the sort of business side of things. How does this actually work in a business context? And I think that's, that sort of reflects the marriage of, you know, what we're trying to do here for legal operations, bringing together tech process, etc. Into a sort of a business and, you know, legal context. So I think that was a really good sort of layering. The other thing I'll just mention is that I've had this persistent thought since we had this conversation and also since one of our previous Pulse episodes about what this means for other parts of the market. So, you know, we've, we've got here access to the, one of the most advanced sort of language models that exists in the world today. And it's available to be used by a number of sort of anyone out there who wants to and who can sort of pay the subscription fee. And what does that mean for organizations that either historically or at or currently have invested in building their own model for a particular purpose? So now that one of the most sort of leading models out there is available to everyone, what does that mean for those other businesses? Not sure if you have any thoughts on that, Alex.

Well, it basically puts everyone almost on an even playing ground, if I can call it like that. I mean, there are definitely technology solutions that have invested in a model that might look similar to what GPT is doing. And they've been investing in that for three years and that's been one of their key differentiators. But what happens when everyone else can use, let's say, the same engine under the hood? You know, those, those key things that are going to differentiate now is no longer the engine, it's what sits on top of the engine, it's the, you know, the, the shape of the car, et cetera. I'm trying to think of a cool analogy there, but there, there has to be something else. I think is the interesting part.

All right, enough of the intro. Let's get into the conversation.

Mary o' Carroll and Kaigua Gwilt, welcome to the Lealots podcast.

Thanks so much.

Thanks for having us.

Now we want to hear about you guys. Let's just get some intros. So, Mary, just tell us about you, who you are.

Hi, everyone. I'm Mary o'. Carroll. I'm the Chief Community Officer here at Ironclad. Before joining ironclad, I spent 13 years building and leading the legal ops and technology team at Google and was also part of building and leading CLOCK for over a decade as well.

Cool. And Kai, how about you?

Yeah, hi, I'm Kai Gogwelt. I'm the chief architect and co founder here at Ironclad. My background is in software engineering. So before starting Ironclad, I was at a company called Palantir, working on big data applications in the intelligence community and military space. And then before that, I did my undergraduate and master's degree at MIT in computer science and physics. And yes, that is the Massachusetts Institute of Technology, not the Melbourne Institute of Technology.

Believe it or not. I have gotten that question a couple of times.

Oh, there you go. Right. So we want to talk today about artificial intelligence and GPT. Obviously, it is all over all the news media at the moment, and we know that Ironclad has recently developed some additional products that incorporate GPT and all it can sort of do. So we want to hear about that. But I've got some sort of broad questions to start with, and I'll just leave it to you guys to work out who's going to take it. But the first question is, what actually is it? What is GPT? What is GPT3? Yeah, who wants to answer that one?

Oh, that's Kai for sure.

Yeah, absolutely. So GPT3 is a large language model, or LLM, developed by OpenAI. And it powers the kind of cultural phenomenon called ChatGPT, which is a conversational chatbot that I think probably most of the listeners have heard about at this point. Large language models have kind of been this. I think they originated with, I want to say the 2017 paper attention is all you need out of Google. And they've kind of revolutionized the area of artificial intelligence, focused on language called natural language processing, and are just capable of some really, really incredible things.

And so I think you touched on this a little bit. So I think a lot of people throw around the phrase or the word chatgpt, and then we've got this sort of separate ish concept of GPT. So am I right in understanding that sort of GPT is the model itself, so you feed in something to it and it gives you some kind of response. And ChatGPT is essentially a chat feature which incorporates that model. Is that sort of broadly right?

Yeah, that's broadly right, yeah. The way I think about it is GPT3 is like the foundation upon which ChatGPT and other things are built on top of. I will say ChatGPT does a number of things that are more than just an interface on the chat. It, like, optimizes response times. It kind of thinks for longer or shorter, depending on how complex the question you're asking it are. It kind of has a personality a little bit and tries not to insult anyone or make. It attempts at least not to make factually incorrect statements, but. So there's a little bit more going on there. But, yeah, broadly speaking, ChatGPT is built on top of this foundation of GPT3.

And I'm just going to chase that rabbit down that hole a little bit there, because firstly, I'm keen to understand because we don't just have GPT3, there's other models as well, I think so. For example, I believe Microsoft with Bing have recently incorporated one of the GPT models. I think it might have been 3.5 or maybe even 4. Do you know enough to tease that out a little bit?

I guess I do. I would say. I think the thing that's problematic here is actually the nomenclature that we're all using around GPT3. GPT 3.5, ChatGPT. So I think the way to think about it is GPT3 is the foundation, and then all of these other cool things have been built on top of it. So when people say, you know, ChatGPT is built on GPT 3.5, which is built on GPT 3, I think it's all getting a little bit confusing because we're just adding nomenclature for different things. In my mind at least. The big Advance Right before ChatGPT was something that happened about nine months ago called instruct GPT, where OpenAI basically helped GPT3 better understand how to follow instructions well, hence instruct GPT. And being able to follow instructions well kind of opened up this world of possibilities that helped us better understand what this large language model is now capable of. Before, it would vaguely follow instructions, but sometimes it would just act as a really advanced autocomplete. And now with this kind of very good ability to understand and follow instructions, we're really starting to understand just how powerful Its capabilities are. So it's a very roundabout way of answering something very simple, which I guess my answer should have just been. My understanding is that Bing integrates the most advanced version of GPT.

That is available right now.

Elliot, how's that rabbit hole? You want to go any further or. I got a question.

Yeah, go for it.

Yeah, cool. So, Kai, I've been reading some commentary that you've had in different interviews and you talk about this race for novelty and it's I think a bit of a comment around the macro trends around AI, as in AI within the legal industry. So can you just help me understand this concept around this race for novelty and kind of where we're at from a precipice, from moving away from that.

So part of my role at Ironclad is to kind of follow new technology and disambiguate the hype from the reality. And so there is a lot of hype right now around AI, but I think that hype is actually backed up by a ton of substance. And so I think you could describe it as a precipice, as if we're about to come off of this high of like hype and inflated expectations and things. I don't think we're at peak yet. I think there's going to be more, more inflated expectations. But I think that inflation of expectations is going to come because there's so much substance behind what this technology can do.

And I think that's probably a good segue into what actually does it do. And we want to probably get into some of the meat around this and obviously Ironclad being a clm, so let's start there. So how does it work in contracts, digital contracting? And it's kind of a two pronged question. And what does it offer to digital contracting? What parts of the process does it touch? What are the touch points? Give us the kind of macro view. Again, I'm not sure if that's a.

Mary or Kai question I can start. I'd love to hear Mary's take on it as well.

The first feature that we've launched that's GPT powered in Ironclad is called AI Assist. And AI Assist is geared at solving the problem of repetitive contract negotiation. So oftentimes I'll talk to people on legal teams who have to review the same type of contract over and over again, whether it's a vendor agreement or an NDA. It's coming in from a third party stating the same thing in different ways and they're making the same kinds of changes to these to make sure that they stick to the correct organizational standards. And so AI assist and GPT3 are allowing us to solve that problem by having the AI identify the issues and take a first pass at fixing them in concert with the contract manager or negotiator. So that's where we are today and that's what it's capable of today, where it goes from here, what all the touch points are.

I can't even start to predict that. Actually we're exploring the boundaries of what GPT3 is capable of. And as we talked about a little bit earlier, GPT3 is basically getting new abilities every week.

Starting with its ability to follow instructions and then realizing, okay, actually it can also translate things from one language to another language and then realizing, actually it's exhibiting chain of thought reasoning. And so all of these things mean that, you know, the boundaries of what we think are possible today with GPT3 are actually just the beginning. We can push the envelope here.

Kai, before I might be jumping in front of Mary here, so apologies if I am. Kai, can I just pause on this point for a moment to make sure that I understand what the role of a language model and GPT is in that process? So you described, I think, a sort of functionality where the system will look at, I believe, a contract on someone else's paper and help you to work out how that complies with your own sort of policies or whatever it might be. How does a language model help there? What's the role of a GPT in that process?

Yeah, great question. So in the AI assist functionality where it's helping to kind of read through the contract and suggest changes that might be made to it, the large language models are using text classification to identify different sections of the document that might pertain to a certain type of part of an organization's playbook. And then we're instructing GPT3 to make changes in very specific ways based on the organization's pre configured setup to actually change that language in a minimal way to comply with their standards. The easiest example for me to give is around governing law. The large language models are able to identify that this is a governing law clause and the governing law seems to be in this jurisdiction. And then given that information, we can construct an instruction to GPT3 to basically change this text to actually be in a governing law of a different jurisdiction. And GPT3 is very happy to make that change for us.

And we've had AI based.

Sort of language models for some time now. Is GPT something new? Is it a new type of thing or is it just a sort of much more advanced version of what we've had previously?

It's both something that we've. It's both something entirely new and something that we've had for a little while now, and just a new version of it. So maybe to explain a little bit of how can it be both something entirely new and something that's just a new version of what we've had already. All of the different large language models, I think, trace back to this attention is all you need. And this transformer model that was originally published by Google back in 2017, and so they're all kind of like doing the same underlying thing, but we're discovering that as you scale that, as you make it larger, as you add more training information and as you add more parameters to the model, it actually exhibits really interesting different behaviors. And so in parallel to all of, I would guess, the more practical work that we at Ironclad and others are doing in terms of applying GPT3, there's a whole other set of whole other body of academic work going on right now trying to understand at what inflection points of model size, of training data size, do these models actually exhibit different behaviors? This is a very, very exciting area of research called emergent behaviors of LLMs.

And Mary, I think this is where I'd like to bring you into the conversation because obviously we're getting the tech lens from Kai, which is really awesome and important to understand, but I think just as important is why should the legal industry care about this? What does the business need? So I'm curious to get your thoughts on that.

Yeah, and just first, I'm so glad Kai is on this podcast with me because I couldn't answer any of your original questions. So what is happening in legal is that, as you guys all know, it's been several years now that we're spending finally kind of feeling that legal departments are getting that pressure and that we have to transform. And as we see AI and GPT and other types of technology now impacting even our consumer and personal lives.

We have no choice. We're sort of being forced into this world where we've got to change the way that we work. And that's great because I've long said the hardest part about legal transformation is that mindset shift and getting people to even believe that there could be a better way to do things than the way we've done it for decades.

What's interesting about what's happening now is that it's here. And as Kai has mentioned, there is substance behind the tech. We've talked also for the last decade about the hype. Is it reality? Is it hype? What is AI and now what is GPT as a buyer? And I think I'm a pretty informed buyer, even more so now that I've joined a CLM company. It is very difficult to navigate through all the news and the press releases and the oh yeah, we have AI and REI can do that too. It is very difficult to navigate through that noise and to know which provider is doing something substantive, where it actually works and where there's just a bunch of Me too, I'm checking that box. So if you ask me if we have AI as an offering, you know, I can say yes.

So I think it's challenging as a buyer in this time and you've got to really do your homework. And another reason why I really believe in people needing to talk to others in the community and seeing what's real and what's hype. I mean, that's the only way that I really know to get around it. But in general, I think that as you mentioned, maybe as Kai mentioned, the peak will continue for a few months and then hopefully we'll have a better sense of what's possible. It's also very exciting to be part of a time that's transforming and a company that's transforming so quickly because none of us know what it's going to look like. That's frankly very exciting, very thrilling. And you want to find technology partners that are going to be future proof, that are constantly thinking about what tomorrow is going to look like, and not just tomorrow, but ten years from now. And I think that many of you, the companies out there, have been shortsighted in thinking like here is our product and it's great and it works perfectly. But that may be true for yesterday or today. But if you're not thinking about what your product needs to do and adopt to the new technology coming out for the next 10 years, then it's something you're going to have to rip and replace very shortly.

Yeah, and I just want to unpack a few things there, actually. So you mentioned adoption and.

Because of these advancements in technology, I've read again some of your commentary online around this is going to help with adoption. Yes, we don't know exactly what the future looks like, but we know that we're a little bit better than what the past looks like. So help me understand your thoughts around what kind of impact this is going to have on adoption of clm, particularly because of some commentary in the market, at least at the moment, around some of the, I guess, expectations falling short of what the market thought CLM could do.

Yeah, I think that the adoption. My philosophy on adoption has always been about the user experience and the value that you're offering to the user and whether or not they can see that value quickly. So the example is you give somebody a login, can they figure out what they need to do, request a template, get a contract signed, figure out where to go to find something in the repository without training? So that's part of it. Can they just open it up and just like your consumer apps on your phone, figure out how to get something done and then if they can do that, is it a good experience, are they immediately able to say, wow, this is so much better than the old way I used to do it? Going through this app or using this software is so much better than needing to send an email or to document something on a spreadsheet. If that value and that experience with that training happens very quickly, it's walk up usable and you get it, then you don't have to worry about adoption and change management because people will want to use it and they will tell the team next to them and the other team, you guys should do this. It's great. It's so much better. I think the challenge we've had with adoption in the past is that we were trying to shove really clunky tools onto our legal teams and say, I know this is going to take five extra clicks, I know it's not great, but if you could do this for me because it'll really help us get clean data or it'll get us everything in one place, it's for the greater good. That's a really hard sell and that's really hard to get adoption. The comments that I've had recently about the new AI and the GPT that is coming to contracts technology is that the experience is so much better for your professional satisfaction. You're not doing a lot of the mundane work like the example that Kai just raised. If you're constantly trying to get your contracts in the right jurisdiction and you've got to make that copy paste over and over, day after day, day and search for it, that is not exciting. And you know, no one's going to tell you, gosh, I wish I could do more of that. So if you've got that automated and you can trust it and it goes away, that's a win for everyone. And you See that as valuable.

Yeah, if I can jump in here, completely agree with what Mary said and want to emphasize something she said. Again, I empathize with buyers this coming year. It's going to be really difficult to understand what's hype and what's reality this year. And I don't know if there's a good way to make it better or advice I have for kind of disambiguating that hype because truly even the people developing this technology, even OpenAI, I don't think they fully understand the capabilities of the technology that they're building. So how can you disambiguate hype from reality when the people building the technology aren't even fully themselves aware of like all the amazing possibilities of the technology? So yeah, it's going to be really tough. But I think one thing that I'm really excited for is the question today seems to be.

Is this platform integrated with GPT3? Is this, you know, powered by GPT? And I think the question that people are going to start realizing that they need to instead ask is, you know, what value is this GPT powered feature bringing to me? And as as quickly as we can get back to asking the question of like what value is being created by this tool? And away from the question is this tool integrated with GPT? I think the better.

And Kai, I think you've answered it to some extent already, but what is the value? And also I think Mary touched on it as well. So obviously we are attempting to to some degree automate away that repeatable low value work. Is that the extent of it or is there more?

I think there's more. I think there's a lot more to come. But you hit the nail on the head in terms of kind of the initial foray with AI Assist, but I do think so. We keep coming back to this idea of changing governing law to a jurisdiction and honestly how soul crushing making that change dozens of times a day can be. So I think it's a little bit more than just hey now, instead of having to type those letters, you've now clicked a button and the AI has done it for you. Because what you're saving in there as well is you're saving this back and forth.

Between two different people who are busy doing other things, who have other things going on in their lives, who may be in different time zones in this current world.

And then you're also creating job satisfaction. I kind of think of it as you're allowing right now with AI Assist, you're allowing ironclad to clear out the underbrush of the routine contract negotiation, allowing you as the user, as the contracts expert, to kind of go one level up and better use the skills that you are so much better at than the skills that you trained for. Yeah. So by having the AI kind of clear the underbrush, we're allowing humans to what we like to say. We're allowing humans to do more humaning, to do the things that humans really excel at and to hopefully get more job satisfaction from that.

Yeah. If I could just add, I mean, just to bring it kind of back down to earth a little bit too. If we take AI and out of the picture, the future of where we're going with legal technology is really about data. And that can be for contracts, that can be for a lot of stuff. And then you layer AI on top of what you can do with data and, you know, it's truly limitless. But if we even think about just like digitizing contracts and turning them into pieces of data and okay, is that valuable? What value does that add to me in my department or my greater organization? And what decisions can I be making? What can I be doing with that data that I wasn't able to do before? These are all the questions that we should be asking, and I don't think we know the answer to even that question. You know, even before you bring AI into the picture, what can you do with all the data in your contracts? We're just starting to see what users, what the community is leveraging that data to do to answer questions for and think about using in different parts of their company that we weren't aware of in the past. And I don't think anyone has figured out what they're able to unlock. So it's an area that I think we're all watching and we're all going to learn from each other and, and share, and share best practices. But yeah, it's obviously valuable. What can we do and how valuable is it remains to be unseen. And that's what makes it fun.

And Mary, you've touched on a point that I was just about to take us into anyway. At the clock conference last year, you did a talk, and if I could summarise, and you can correct me if I'm getting this wrong, it was talking about the sort of evolution of legal operations and the idea being that sort of stage one is make the, make the legal team more efficient. Stage two is, you know, use the skills and knowledge and experience that we have to make other teams more efficient in relation to sort of Legal type services. And stage three was taking that, but at a more strategic level for the organization. Is that sort of broadly right?

Yes. So glad someone was listening.

Good, great. And so my question is, do you think this new capability of sort of marrying artificial intelligence, intelligence, GPT language models, et cetera, with nascent legal technology has a role to play in that story?

Absolutely. I mean, at the heart of Legal Ops, I've probably said a million times, it's about. Right sourcing, it's about optimizing the legal department, what resources you are using to deliver.

The legal services that you need to deliver to your clients. And should it be done by humans, should it be done by AI? Should it go to outside counsel? I mean, there's a lot of questions about the who. And now I think this brings in the what as well. It's like the who. Is it a who or is it a what? And if there's work that was formerly being done by lawyers or other humans that doesn't need to be or can be done better, faster, cheaper in another way, then it would be irresponsible for us as Legal Ops professionals to not consider how do we leverage that in our department. So it is absolutely a tool. And this is part of how you stay strategic thinking and in your role is not to just put systems and processes in place and let the trains run on time. I mean, that's certainly part of the job. But constantly iterating and continuous improvement on optimizing the legal department and optimizing how you're able to help the rest of the company. And this is a critical piece of it.

Yeah, I do feel like Legal Ops is probably in a particularly difficult situation right now, or exciting situation, depending on your mindset. With all these AI advancements, I think.

Would be forgivable for. For someone to say that they don't want to participate in all the hype and understanding the new stuff. Just let it settle down and, you know, I'll pick up the new way of doing things when someone else has figured it out. I think that's totally forgivable for maybe attorneys and other people in the legal profession. But for Legal Ops, I think. I think there's going to be a responsibility to try to get up to speed really quickly and to stay up to speed this year with all the advancements that are happening, because that is the duty of the Legal Ops function to allow the legal team and the company at large to perform at maximum available capacity. And AI is just completely going to change this year what that means.

And that is an excellent segue actually into a direction I wanted to take this, which is to ask about risks, actually. So I'm very sort of conscious and aware that in the self driving car space we have some providers out there, you know, launching what are being marketed as self driving capabilities. But actually they require the driver to still be aware and still be able to take control. I think that has led to a number of problems that I'm sure many people listening are aware of. I wonder if there's a similar risk here where we might have some sort of automated AI based system helping with contract review, but we actually still need the lawyers to be able to spot issues and not rely fully on the AI. Do you have a view on that sort of risk and how we might mitigate and balance those two competing priorities?

My thought is AI and technology is here to help you. I mean, I heard someone once say it's like lane assist, you know, in your car, you want it, it gives you an alert if you're in the, in the wrong lane. But you're certainly not going to take your eyes off the road or your hands off the wheel. And if it is your job to make sure that that contract is ironclad, no pun intended, you should use the technology as a help and a guide. But ultimately you do want to spend the time looking at what's important.

I think we're all still in a place where this is meant to help.

And it probably will get to a point where it's so good that the risk level of not looking over some of the things that AI has done for you is fine. And I'm actually positive we're doing that for a lot of things where the risk level is low. But that's going to be a department by department and lawyer by lawyer judgment. And my guess is it'll get better over time and people's risk tolerances with what they're comfortable with, needing review or not needing review, will continue to evolve. I'm sure that's what we saw with eDiscovery. I'm sure that's what we're going to see here.

But although that does make me think that there's a user experience or user interface sort of issue here, imagine you guys have thought about it in relation to your own product. So again, if you're going to have some sort of automated tool assisting with review, I'm assuming, and I'm interested to hear your thoughts on it, that you want to build into the user interface and the user experience prompts to make sure that the human user is still, you know, putting that effort in and still doing that review. Is that. Do you think that's a fair comment?

Yeah, absolutely. And that's something we spend a lot of time thinking about at Ironclad. How to responsibly integrate this technology and make sure that the UI is not kind of cutting the human out of the loop, but in fact inviting them in and making sure that they're a part of the decision making process, which is so critical.

I would say also kind of to go back to your earlier question and comparison between self driving cars and self driving contracts. I think one of the critical differences here is.

Going from human driven cars to self driving cars. We're going from humans spending basically all of their attention driving to less of their attention driving. And when we talk about self driving contracts.

What I hear sometimes is that the amount of work that the legal team is being asked to do is to a point where basically even without AI, there are self driving contracts happening whether they want them to or not. And so I think having AI Augmented legal review or AI Assist is actually just adding AI to these contracts that probably would have been driving themselves anyway. And now the legal team is able to get just a little bit more control over that process.

Yeah, that's very true.

I want to just play out a thought that I'm thinking about at the moment. So we're talking about elevating the legal team to focusing on that more strategic stuff, right? By not having to negotiate or redline those jurisdiction clauses that we're talking about. At what point do you think the tip is going to be between the legal team having to make that conscious decision based on what the system is saying around, hey, look, this could be what you're redlining to the business. As in at what point are we cutting out the legal function or at what point do we not need the legal function? Thoughts on that?

Yeah, great question. I don't think we're anywhere close to not needing the legal function.

The thing is, I feel like there's maybe there's like a Maslow's hierarchy of needs in terms of the legal function. And.

I think a lot of legal teams that I've talked to kind of pre Ironclad are very much in that bottom layer of Maslow's hierarchy and just kind of like getting by. But I think, and I think even with Ironclad, we're not at that, you know, the very top of the pyramid and you know, providing.

Much higher on that level of needs. And so I think the closer we can get to the top of that pyramid the better. And I think at least this current phase of AI is going to get us a little bit farther up that pyramid, but it's not going to get us all the way to the top where suddenly compliance with all of the regulations and understanding the full business context of how we're shifting our business strategy and therefore our entire contractual relationships. I don't think we're going to be quite there yet with this round of advancements in large language models.

No, look, definitely not suggesting that, but I was just thinking.

Can a business person mark up an NDA using this? I think so. Then go a little bit further. Some more complex agreements and some more complex obligations. Probably we're never going to get to that last 5% or 10%, but sorry.

Amir, I think I cut you out.

No, I was probably going to agree with you there. I mean, we already have the capability and what we encourage, I mean, as legal ops professionals, is to push as much self service as you can onto the business. So within, you know, the controlled language that's been pre approved by the legal department and that of course is where legal needs to come in and set that within these parameters, if you're a business person, you can go out and execute as many contracts as you like and get them signed without coming to legal. AI is just, you know, another lever that you're going to add to maybe raise the bar on what they can do on their own. But yeah, it's not going to eliminate the need for lawyers or legal.

Can I just touch on one thing you said there, Mary? You said essentially we're kind of setting a guardrail whereby business people potentially can.

Work through a contract themselves using pre approved language. What does that actually mean in the context of an AI based contract review where I assume you might not be relying on a pre approved language, but you're relying on a pre approved model. Is that sort of right?

I guess, yeah, I guess so. I don't know that. I don't know that anyone's doing that quite yet. You know, relying on AI to sort of look through let's say third party paper and allowing the business to say yes to it without that coming to legal. I suppose you could, yeah. It's a tolerant, you know, it's, it's a legal risk tolerance and review. And what are you comfortable with letting the business do by yourself? So it's going to really depend and vary on your business and how you want to run it.

I suppose what I was trying to get at there is.

We used to have pre approved language for certain clauses. I mean, we still do. Now we're introducing GPT, some sort of AI based model which can review and propose amendments to contracts on other people's paper. And so I imagine that the output of that review, particularly if it's proposing drafting, is not necessarily just the pre approved language because if it is, then that feels like the AI is not doing much. And so again, it feels to me that the next evolution of this is within some kind of guardrail to rely on the output of one of these models as opposed to just pre proof language.

Yeah, I think, I mean, I know nothing about technology but I think that's entirely possible and a very logical next next step of where things can go. Kai, would you agree?

Yeah, I was going to say just to riff on this, I think so this idea of like are you relying on pre approved language or are you relying on a pre approved model?

Feels like the sort of thing which today I think it would be very silly to rely on a pre approved model, but you know, in three weeks I might change my mind based on the advances in the technology. And I say this because the idea of relying on a pre approved model or maybe a model that tries to figure out how you as a lawyer or you as a company are negotiating different positions, the problem there feels like it'll always be a little bit behind because it's learning from what you're doing, doing and it's probably, you know, it's probably a little bit behind. So the more preferable thing would be to be able to tell the AI model like, okay, this bad thing happened last week, we're not going to let that bad thing happen again, so stop doing it the way you were learning to do it before. And so that's where I think maybe not pre approved language but pre approved positions and then having a model that can understand those pre approved positions and that pre approved language and apply it to the best of its ability and hopefully in a way that's useful is actually where we're going to go with this.

Yeah, it's fascinating how quickly the technology is evolving and I'll just give you guys a few examples. There have been a couple times recently where I've just been chatting with Kai and said, I've been talking to everyone in the community and I think it would be crazy cool if AI could do this. Is that possible? And Kai's like, yeah, not, not yet. You know, three weeks later he's like, can we meet? I want to show you something. Because things changed. And all of a sudden, you know, he's doing a demo for me on something that wasn't possible three or four weeks before. And that's happened a couple times now. Right. And that's pretty wild about how quickly things are changing. So, you know, depending on when you release this podcast, things could be different.

Truth.

So, Kai, I want to pick up a bit of a theme that's coming through here, and you've mentioned it right at the start of the podcast around this concept of emergent behaviors. And my understanding of that is that's kind of like the magic that sits behind these LLM language models, where they're kind of performing things that aren't necessarily or they weren't necessarily trained on. So when we start to use these technology solutions and they do start to do things that they're not trained on, like, obviously there's a positive to that, but there's also a negative to that. So help me understand the positives of it and then help me understand the negatives of it.

Yeah, absolutely. So I guess the positives of emergent behavior are when these emergent behaviors end up being useful for something. And so I think language translation is the perfect example here. Right.

And it's hard for me to emphasize how crazy this emergent behavior thing feels to me as a technologist.

Traditional AI, you train it on a task and you keep training it on this task and it gets better at that task, but you don't expect it to then be able to do this entirely different task. The emergent behaviors that LLMs are showing right now is mind blowing because that's exactly what's happening. You trained it on this unstructured corpus of text to be able to do some stuff, and then all of a sudden it's like translating from English to Russian or something like that, and it's just like, it's just wild. I referred also to some of the academic work that's going on that's trying to understand at what model sizes does this emergent behavior start to happen? At what point, at how large of a model do you need in order for this translation capability to come into play? At what level do you need.

Chain of thought reasoning to be exhibited? And so I think we're going to continue to discover new emergent behaviors, and it's going to be really beneficial because a lot of the emergent behaviors that we look for are the ones that we.

Would find useful. And you referred to some of the emergent behaviors that might not be as good. I don't have any examples of those. I think that's just because we simply don't ask the LLMs to do the things that are problematic or we try not to.

Yeah, I guess the train of thought, and perhaps it was a load question, is just around this explainable AI concept that's been kind of floating around, particularly within the legal industry, like where we have decisions as a lawyer, we need to back those up by rationale. And if the systems are getting so smart to the point where we can't really easily explain how it's either made those connections or made those decisions that could be problematic. I don't have a particular view on it, but I'm not sure if you've looked at the kind of explainable AI angle to GPT and to ChatGPT.

Yeah, no, absolutely, that's a great point. And I actually think of it more from more as the hallucination problem. So in kind of LLM, the idea of hallucination is the idea that these AI systems can often produce very, very convincing answers to questions that are completely unsubstantiated by any facts. And so, you know, that's my specialty.

There's a lot of. Yeah, there's a lot of research going on in terms of trying to understand how to this hallucination problem. But the funny thing I'll point out is just the hallucination problem is also, frankly, a very human problem. We're criticizing the AI for coming up with convincing answers unsubstantiated by fact. But I'm sure all of us have friends who make statements all the time that are entirely unsubstantiated by fact.

But it is a real problem and it is something that I think there's some really amazing research going on right now around, and I think there's a good chance that by the time this podcast releases, there are some good precedents already set for how to work around it.

Now, this issue of emergent behavior I think points us in the direction of what's coming in the future. And so I want to ask that question. It's a big sort of picture question. So what is coming in the future? And here, I mean, both from a sort of technology perspective and also from a How is this going to help our sort of workflows in the legal space? So I might start, Kai, with you. Any views that could be, you know, completely, completely big picture, completely out there, but any views on where this might be going in the future?

Yeah, absolutely. I think, you know, at the end of this year, we're going to look back and realize that the nature of work around contracting, the nature of legal work, is fundamentally different. Like, I think this is the year where AI changes the nature of legal work and the nature of contract work. And I think they're kind of like two parts to this. The first is this year is going to be really rocky. We've already talked about good emergent behaviors in terms of creating possibilities that we all thought were impossible, at least for the next couple of years, but also inflating expectations to the point where it's almost difficult to or it's almost impossible to tell when people are representing things truthfully and when people are just making outlandish claims.

And then the other problem here, I think is the hallucination problem that we just talked about. If we are not able to, you know, work around it appropriately, I think there are going to be a lot of products built without kind of the hallucination problem correctly accounted for. And that's going to create just a lot of like incorrect answers in an already in a space that's very sensitive to incorrect answers. So we're going to have a very rocky year is my prediction. But at the end of the year, I think we're going to realize that we've completely changed the way that we work. And I think the change is going to be so complete that we'll actually think it's pretty pedestrian. We're going to be using these crazy features that.

12 months ago we would have said, I don't think that's going to exist for the next five to 10 years. And we're going to use them the same way we use our. I happen to be an Android phone user, but the way we use our phones today.

And we're just going to say, okay, this is just the way we work now.

So okay, that's a big prediction there, Kai. So radical change in the legal industry throughout this year after a sort of rocky time. Mary, what's your take on the answer to that question?

So my answer is probably going to be the non technical answer, which is that I think that all this hype, whether it's based in reality or not, is welcome because it is going to be accelerating more change and more conversation and more pressure on the legal industry to start doing more. I've often said that the biggest challenge we have in transforming legal is just education.

Most people think that the way that they've done it is fine and they just need more headcount, they need more budget. But being able to open people's eyes to the art of the possible for them to have that aha moment of, oh, I mean, it was fine, but it could be so much better is what we need more of. And if this is putting more pressure on people. I remember the sort of adventure of AI and legal, right. Where everyone was getting pressure from their GCs, who would just show up at our doors and say, I heard about this AI thing. Where are we using AI and legal? Where can I use AI? And I'm like, what?

That's.

What kind of question is that? You know, like, what problem are we trying to solve? That's where we should start.

But if that is giving us more pressure to start looking at how to do things differently, at giving Legal Ops.

The mandate to go create and initiate change in a department, then I think that is going to just create again, more acceleration, more conversation, more education, and faster transformation in our space, which is great.

Yeah. And if I could just summarize, I feel like we're on the precipice and the future is exciting and we'll see where it's going to go and it's going to be rapid. Yeah. Okay. I think we've probably covered everything we wanted to, sort of. So, Mary and Kai, thank you very much for coming on the Legal Office podcast.

Thanks for having us.

Yeah, thanks for having us.