Turns out, courtroom battles over AI aren’t the only way for lawyers to capitalize on the generative AI goldrush. Because of their particular skills — reading, summarizing fast, answering questions — the new generative AI platforms seem a perfect fit for certain kinds of jobs, like reading and summarizing contracts, even writing boilerplate legal documents and answering questions based on a body of text. This has led to a lot of anxiety about robots coming for white-collar jobs. It also leaves big companies thinking hard about what they can use these platforms for, and how soon. Earlier this year, an AI startup called Harvey partnered with law firm Allen & Overy to automate some of the firm’s legal document drafting and research tasks. Harvey’s backers include Sequoia Capital and an investment fund managed by OpenAI (yes, that one) and its raison d’etre is creating custom generative AI software for law firms. Allen & Overy is far from the only firm to use Harvey, but it’s certainly got one of the highest profiles in the world of big law. So what will AI in a big law office really look like? I sat down with David Wakeling — who is leading A&O’s AI rollout as head of their Markets Innovation Group — to get the lowdown on how adoption is going, and how they are mitigating some of the risks. The biggest and best language models, like GPT-4, are also known for confidently making stuff up when they don’t know the answer — “hallucinations” is the technical term — and that could be quite dangerous in a business where law firms need rock-solid information and arguments. “It’s not a chatbot,” Wakeling clarified at the outset. “We're using it as a single question and answer,” he said. So the kind of emotional back-and-forth that the New York Times’ Kevin Roose had with ChatGPT is not what A&O’s AI tool is built for. To Wakeling, the most interesting thing about fine tuning a large language model for lawyers to use “is not so much adding huge amounts of data,” he said, but “more about suppressing the areas of GPT4 that are not very helpful for the business use-case.” And those use-cases are “boring,” Wakefield said, like “productivity gain on drafting a provision for a loan agreement or something. It’s not really about individuals. It's not reflective of societal trends. It's simple.” So how does it work? Well, the base GPT-4 large language model that Harvey provides is “tuned for law,” Wakeling said. And unlike Google’s Bard or the new GPT4-powered Bing search, where the AI tool combs through the open internet for answers to user questions, the Harvey AI model looks through a restricted dataset each time it needs to find an answer. “So you're keeping the base model, but forcing it to answer with respect to a book on my shelf,” Wakeling said. Still, “we work on the assumption that it hallucinates and has errors,” he said. Wakeling compared Harvey’s GPT4 model to “a very confident, extremely articulate 12 year old who doesn't know what it doesn't know.“ Down the line, Wakefield thinks that firms who are able to leverage AI to shore up the “the skills and knowledge of the institution” will stand a better chance at survival than those that don’t adopt AI and let institutional knowledge stay only in the brains of humans. And how about the impact on jobs? Right now, Wakeling said the AI service that he sees on his desktop would not result in job cuts and that the resulting productivity gain would be “at all levels — at the most junior and senior.” But he also added that he “could not crystal ball gaze” about how AI usage would affect A&O’s workforce in the coming years, given how quickly the technology is evolving. While everyone can agree that AI will be a game-changer for businesses (and the individuals working for them), not everyone is so sanguine about the kind of approach A&O is taking. “It's a shame that the benefits of AI are going to big companies and law firms rather than to consumers and customers. And it's not clear whether the savings will be passed on,” said Joshua Browder, the CEO of DoNotPay, which provides low-cost legal advice via an AI-powered chatbot. Wakeling said the use of Harvey is not currently reflected in what they bill their clients — but that might change later on. “In the future we imagine AI being used to support services at cheaper fixed fee rates than achievable in the past for e.g. due diligence,” he said over email. “It's not going to change billing in the short term,” said Browder, “especially because some of these AI partnerships are exclusive,” meaning big law firms have no real incentive to share the cost-savings with their clients. But on a longer timescale, Browder has faith in the “tens of thousands of lawyers” trying to apply AI tools to their individual market areas to correct the balance in their clients’ favor. Someday, A&O’s clients might see fewer billable hours on their invoice as a result of the great Harvey AI experiment, even though AI isn’t actually going to practice law anytime soon. “Eventually, the market is efficient,” Browder said.
| 
