The law is one of those most foundational vehicles of our societies – and it’s something complicated, something complex. It’s something that relies on semantics, and the effects of words and rhetoric. It’s also something that gives us the delineations and guardrails that we need to move forward both in business and in our personal lives.
In business, that’s taking on a new face when it comes to artificial intelligence. We need to think about how we apply AI to legal systems. There were early reports of LLM results being used in reckless ways – producing arguments and documents including hallucinations, unchecked by human oversight.
So how do we move forward with this?
Thoughts on AI and the Law
A recent panel at the IIA event at Davos went into this in great detail.
Gabriele Mazzinin, who helped with the European AI Act, made the point that the intention with the law was to empower more innovation and adoption, by making these processes safer to go through.
Pablo Arredondo, who has a legal tech company called Casetext, talked about the need to provide tools that show oversight, for example, with transparent sourcing.
“We started realizing that we had to think about: how do we design the product to … minimize the chances of harm?” he said, going back to the early days of ChatGPT4. “And that’s … building oversight directly into the system. And then how do you go about training, teaching people about it? The law has its own need to make something regulated within it, because you can do a lot of harm if you don’t do it well.”
Julia Apostle of Orrick pointed out that she doesn’t hear people trying to avoid regulation, but instead are embracing the idea of clarifying rules for the use of AI.
She made the point that regulation isn’t going to be a one-size-fits-all solution, and won’t apply to each business the same way, also referring to a greater framework that can be challenging.
“I think one of the challenges for smaller companies is the volume of parallel of legislation,” she said. “So it’s not just the AI act, it’s the Digital Services Act, it’s the Cyberresilience Act, and it’s also the copyright regime, and so forth. That creates more of a challenge than one particular piece of legislation.”
Gabrielle talked about a “plurality of legal frameworks.”
“It all depends on, in my view, also the implementation,” he said, “for instance, the function of the sandboxes, where normally SMEs should have better access.”
Starting Out Moving Uphill
In response to questions about challenges, Arredondo talked about how to begin with crafting a good response to AI regulation, in the area of law, also noting the rapid advancement of AI models.
“If you understand law and you worked as a lawyer, I think that could help,” he said. “It doesn’t do everything. And I think another interesting aspect of this now is, when we started out in 2023, GPT4 was essentially the only game in town for the kind of quality that we thought we needed for law. And there was debate about (whether) open source (would) ever even come close to GPT4. And now what we’re seeing is that there are open source models that are coming out that are quite cutting edge, including things that OpenAI just released late last year.”
Apostle mentioned scope:
“There is scope, obviously, for compliance tools,” she said. “That’s going to be something that everybody will need to think about: how can we actually build tools that help us with these tasks? And identifying the risk is a big one, right? Because it’s one thing to say: here are your risk categories, but then you actually have to find how they play out in the product.”
Embracing the Regulators
Mazzinin, in talking about business and the future of AI, suggested that companies don’t have to fear overly broad regulation, in general.
“I think there is a bit of a miscommunication or a misunderstanding about the European legislation, that it’s a huge, onerous law that will apply to every single company, every single product, equally and onerously. And so that’s the first message to correct, and then it’s about developing tools that will then help identify how they are actually impacted.”
Debate on Open Source
Eventually, talk turned to open source models vs. closed ones: Arrendondo mentioned the input of Yann LeCun, former head of research at Meta, and his zeal for open source systems.
“My sense is (that as) entrepreneurs, (what) we need to need to do is … ask yourself, maybe you’re not directly under the letter of this thing, but if your AI is doing the kinds of things that they’re targeting, (it) would behoove you to at least start engaging in council or thinking about that, because that’s probably going to be the best we can do, given how fast everything’s moving.”
In weighing in on rules for open source, Mazzinn pointed out that with the AI Act, the focus was on high risk applications, with different sets of rules for different models.
“I think that debate led to a new chapter where we have rules applicable to all foundation models, with some exceptions when it comes to open source,” he said. “But for instance, provisions around copyright protection are applicable also to open source models, and then there are additional sets of rules for (anything that) poses systemic risk, and in that case, the rules (are) applicable to all models. So certainly, the debate about open source (models) and the impact that they may have on certain risks, including existential risk was definitely there and shaped the outcomes.”
Apostle talked about how regulation will look.
“There hasn’t been a lot of regulation of software in any legislation in Europe,” she said. “So this is already (a challenge) to include software and legislation. It’s not just the AI Act, but the cyberresilience Act, NIST 2, those also tackle software as one thing, and then to tackle open source software (is a) whole step beyond – how will enforcement in relation to open source software look? It’s going to be a very interesting development and also, I think, quite a challenge for regulators, as though things weren’t hard enough.”
She also pointed out that lawyers have the benefit of seeing risk across different industries. “That’s a real insight and quite a valuable body of knowledge that can be then used to develop tools that address a specific risk,” she said.
Working With Stakeholders
In response to questions about collaboration, Apostle suggested law firms should be able to make these kinds of partnerships.
I think all professions need to be thinking about how we can evolve in the smartest and the best way to serve our clients,” she said. “So if that is, you know, partnering, then yeah, all law firms should be open to partnering and I think that will be a big differentiator, also, between those firms that can, and those like that can’t. And maybe, actually, we’ll create more opportunities for smaller firms to evolve in different ways, because they can offer one part of the service, and then partner with a different service, for a provider to offer an entirely different system, a type of service that is not incompatible.”
Arrendondo agreed, providing this picture of what a modern firm’s activity might look like:
“I’ve seen fantastic in-house solutions from firms for certain things, for things like legal research and other content needs, for certain things like that. But I think it’ll be a mix, right, where you’re going to have some things you completely outsource, mostly either because of content, or a bunch of nuances, some things that perhaps you do partner and kind of develop, and some things that somebody just sort of comes up with, and it ends up being a really useful way to do something inside the firm.”
He also talked about the analogy to electricity, where AI will spur so much proliferation of products and services.
Mazzinin explained how the European law affects U.S. companies and others outside of the Eurozone.
“The E.U. says, essentially, we have an internal market. We regulate the product that enters into our market. Therefore, if you are an American company and want to sell a product in the E.U., and that product is regulated in E.U. because it’s high risk, or, let’s say it’s prohibited, or is subject to transparency obligations, then you have to comply. So in that sense, it actually imposes obligations also on developers outside the EU that want to enter into the market.”
More Insights
In closing, Mazzinin talked about enabling business.
“I think … it’s important to, going forward, to ensure (that parties are) enabling innovation in Europe, you know, with the consistent application of the plurality of E.U. laws that apply, and so making sure that, you know going forward, companies have a clear path for innovation,” he said.
“I would say it’s all in the implementation,” Apostle added, “and we’re at a certain stage, (where) it will be very interesting to watch the space on implementation.”
Arredondo ended by suggesting that, at the end of the day, regulators and companies can be on the same side of innovation.
“I think (with) all of the folks behind this regulation, it’s because we see (AI) doing so much good, and the potential for it, that these things are set up,” he said. “So I think society is on your side overall.”
That gives companies a lot to go on when thinking about where the rules will be for AI applications, and how the laws will work.