Having worked in a law office, I can say confidently that the future of legal AI is human+AI. AI can rapidly spot all the issues, saving expensive lawyers from spending their time wading through the documents. But the human lawyer brings strategy, judgement and face-to-face consultation to the process.<p>In other words, I don't think lawyers are in danger. But law clerks? Junior staff? The people who are usually forced to do all the grunt work to allow the top lawyers to focus on their strengths? Those jobs are at definitely on the chopping block. There's still an insane amount of rote repetitive work in your standard law firm.
Time wasted reviewing pointless variations on the same old mutual NDA is a scourge. It's hard to take the profession's performative technophilia seriously while the most obvious, widespread inefficiency continues every day.<p>On the other hand, teaching computers to parse pointless variations of the same substantive terms in legalese accommodates bad practice, instead of correcting it. It's symptomatic treatment at best, enablement at worst.<p>Lawyers shouldn't be sending and receiving a thousand different NDAs for the same fundamental deal. They should be defining and evolving a shared vocabulary of common forms, and invoking them whenever cost-benefit favors a standard.<p>That's the idea behind a very recent project of mine, The Canting Tribe NDA:<p><a href="https://nda.cantingtribe.com/" rel="nofollow">https://nda.cantingtribe.com/</a><p>It's a "viral" NDA. The website publishes a versioned form. The form includes a signed certificate guaranteeing that the proposal is identical to the version published on the site. On the site, newcomers see a list of firms using and accepting the form. The experience <i>shows</i>, rather than <i>tells</i>, how standardization can work.<p>I soft launched the project just a couple weeks ago. A couple firms signed on almost immediately. Several more are reviewing. My top priority is more firms on the website. Once the site provides enough baseline social proof, the form can spread word and create new users on its own.
I am a transactional lawyer and I definitely would find value in an application that could issue spot an agreement in seconds. That said, just yesterday I spoke on a panel on the topic of how things can go wrong in a contract. We spent the majority of time talking about the dynamics and challenges that exist outside the agreement in the process of trying to memorialize the parties’ intent in a clear, concise, precise and reasonably complete manner. There are often significant challenges in terms of clearly obtaining the intent and relevant issues from the various stakeholders. And there are dynamics like relative negotiating leverage and psychology or other issues that can drive what the deal will look like regardless of pure legal issues. Also, since one never starts with a blank page, there is the contract template one starts with that must be evaluated against all this – what stays, what goes, what must change and how. Navigating these requires intangible skills, instincts, sensitivity to human dynamics, etc. It’s very much a human endeavor. So a key question is to what extent AI could help with all of these external issues. I have to think that’s much farther down the road. But having help assessing purely legal issues within the document would be a great supplement.<p>For a good thread on training to be a lawyer through apprenticeship, see this <a href="https://news.ycombinator.com/item?id=16255023" rel="nofollow">https://news.ycombinator.com/item?id=16255023</a>
With an AI, you don't know what it is that the AI <i>doesn't</i> know. A trivial example: Some "nondisclosure" agreements (NDAs) also include invention-assignment agreements [0] and/or non-competition covenants. <i>I'll</i> recognize such provisions if I see one in a putative NDA, but I don't know whether the AI will recognize it. Sure, the AI could provide a list of everything that it <i>does</i> know, but when I review the list, I won't necessarily notice that a missing item is missing.<p>[0] Stanford University got trapped by an invention-assignment agreement contained in an "NDA" that one of its researchers signed when he visited someone at Roche — see <a href="https://en.wikipedia.org/wiki/Stanford_University_v._Roche_Molecular_Systems,_Inc" rel="nofollow">https://en.wikipedia.org/wiki/Stanford_University_v._Roche_M...</a>.
Statement from Professor Yonatan Aumann, advisor to Tel Aviv-based LawGeex, on LawGeex AI:<p>“The technology has been developed through a combination of supervised and unsupervised learning techniques. Unsupervised learning was used for teaching the AI engine the core legalese language. Thereafter, supervised learning, using deep learning multi-layer LSTM and convolution technology, was used to train the system for the fine-tuned issue-spotting. Supervision was performed based on human-annotated documents, using legal experts. A unique augmentation algorithm was applied to boost learning from these examples. The overall result is the most advanced technology for the automatic analysis of legal documents. The p-value for the statement that accuracy of AI is above that of these lawyers is 0.0068 (using MannWhitney’s U test).<p><a href="https://images.law.com/contrib/content/uploads/documents/397/5408/lawgeex.pdf" rel="nofollow">https://images.law.com/contrib/content/uploads/documents/397...</a>
Clicking on the link, I expected to discover a watershed moment akin to AI beating a human at Chess or Go. Alas, since contract review is not an adversarial process, no human lawyers were "beaten". A less click-baity headline would be: AI slightly more efficient at issue-spotting contracts than lawyers.
The labor cost savings aren't going to make it to clients. I've worked in a law office, I know how billing goes. Those extra billable hours will be found elsewhere. I hope I'm wrong, and maybe I'm just cynical, but my opinion of lawyers and the legal profession in general took a hit after 3 years of working in it.
>> It took the lawyers an average of 92 minutes to complete the NDA issue spotting, compared to 26 seconds for the LawGeex AI.<p>Issue-spotting an NDA is not a typical task for an actual lawyer. It is the sort of thing handled by a legal intern or paralegal... then signed off on by a lawyer. It isn't typical legal work. The results are therefor of limited application in the real world.<p>Better test: Client asks "Do I need an NDA?" or, conversely, "can I break this NDA I signed last year?". Ask C3PO to handle that question. That is the sort of problem where the real legal expertise happens.
I just joined a company last year that works in this space almost exactly - due diligence contract review - and we are hearing this pretty constantly from people who evaluate our software. Not only is it much faster, but frequently it finds clauses that they had missed themselves.<p>It's a pretty exciting time to be working in this space.
NDAs tend to be a straightforward type of document with limited scope and a fairly well-known universe of issues. So having a computer program parse the document and itemize issues seems like a good idea.<p>But what does it mean that the AI “beat” the humans? That the program simply spotted the issues and listed them faster and more completely than the lawyers? Not that impressive a victory. It’s how the issues get resolved that are important.<p>Did the AI interact with a lawyer on the other side and get issues resolved? Not likely, but the story doesn’t tell us exactly what the test consisted of and how “victory” was determined. (It is possible to download the 40-page report and analyze it, but I am not willing to spend the time to do that, and I bet most commenters, like me, are just responding to the headline take-away.)<p>While AI can be a great tool, one of the problems with it is that often no one knows how the program makes decisions. Not knowing what you don’t know can present serious blind spots. For example, it may be possible for a counter-party to a transaction who has access to the same AI as you to construct a document with text containing an issue that the AI won’t spot but that the counter-party planted in the document for his or her advantage. If the document is reviewed only by AI (and given a cursory review by a bored low-paid worker who expects AI to catch all the issues), then there is an open back door to your legal castle.<p>Computers haven’t put accountants out of business, and they won’t put lawyers out of business either. They will make lawyers more efficient, but whether that translates into lower fees remains to be seen.
Is this technology somewhat akin to a code linter? It automatically finds that problems that would be tedious for humans to find and greatly speeds up the work of the human expert?
Lawyers have traditionally been able to secure their own payout by law, so they ll just thakn you for the extra free time , but will this lead to actual reductions of legal costs?
this is interesting point in time for me....im thinking about becoming a lawyer....im caught between my love for machine learning and deep learning and the desire to practice law.
"Predictive Coding".<p>You don't need AI.<p>You just need collusion between law firms i.e. to write a program, and then agree, that whatever it interprets, is the way it will be semantically enforced in court.<p>To be clear: the study of law is bullshit; so why not take advantage of it, and simply agree, that computer generated bullshit, is what determines it...<p>... of course while still retaining high rates and billable hours for "attorney review" (for billable hours/padding).
51-156 minutes is crazy to review an NDA. If an NDA is well drafted, I can do it in about 10 minutes. If it's a bit of a mess, maybe 30 min. If it's worse than that, I can assess that in about 5 minutes and propose using a better form.