AI For Legal Cases
Our portfolio Casetext was in the news yesterday for raising $12mm, but the more interesting thing about Casetext is their product, called CARA.
CARA is a research assistant for lawyers that offers a super simple proposition:
Securely upload a brief and discover useful case law
CARA uses Casetext’s wikipedia-like database of >10mm court cases and annotations and sophisticated natural language analysis and artificial intelligence to understand the brief and recommend related cases for a lawyer to analyze and possibly cite in their brief.
Lawyers seem to love CARA. According to Silicon Valley Business Journal:
Casetext’s customers include Quinn Emanuel, Fenwick & West, Ogletree Deakins, Greenberg Traurig and DLA Piper.
“CARA is an invaluable, innovative research tool,” Quinn Emanuel partner David Eiseman said in a statement. “With CARA, we can upload a brief and within seconds receive additional case law suggestions and relevant information on how cases have been used in the past, all in a user-friendly interface.”
We think the legal business is ripe for AI-driven innovation. Much of legal research can and will be automated with tools like CARA.
If you are a lawyer and do a lot of legal research, check out CARA. Securely upload a brief here and check it out.
Comments (Archived):
law firms, their names make me chuckle. they always sound so made up.
Dewey, Cheatham, and Howe
Doolittle & Prosper
The a16z podcast had an episode recently about Machine Learning Startups that featured the CEO’s from two companies Textio and EverLaw. EverLaw is building Machine Learning tools for legal business. My favorite line from the podcast, “Machine Learning isn’t just about predicting the future. It’s about helping us change the future.”a16z: The Product Edge in Machine Learning Startupshttps://overcast.fm/+BlzFexQEQ
There is a lot of conflation between AI, Machine Learning and Data Mining.
this is great all around, from the value prop to the business model. impressive!
Nice. The fit into the lawyer’s / paralegal’s human workflow seems crucial. This isn’t just about AI / NLP but about being a total solution. TextIQ for legal discovery and compliance is cranking as well.
Robot lawyers are coming. Oh, wait, I think lawyers have always been described that way.
Law school deans should be aware of this…
I am really pleased with the homepage marketing, particularly this:With CARA, you don’t miss important cases, you catch the cases opposing counsel left out, and you research faster so you can focus on providing best-in-class legal services for your clients.One thing though (well w/o further analysis). For the intended audience it is not crystal clear why this product is actually better. Lawyers who would be potential customers don’t understand AI and machine learning [1] [2] so they need to get into that in a better way by highlighting it and further explaining somewhere convenient.Also the use of ‘CARA’ is problematic in the way it is referred to.[1] As most people don’t, they aren’t @twaintwain[2] It’s not publicly understood like ‘GPS’ or ‘Database’
LOL! Some of my friends were convinced I’m a cyborg even before ‘Westworld’!Legal AI is something I’m looking into because I got approached to do some consultancy for a London-based legal startup exploring how to automate their current system.There’s lots of room for improvement, especially Natural Language UNDERSTANDING. No AI can currently do that. What they can do is Natural Language statistical inferencing.So the AI can pattern match words in the text and surface sections of case law that might be relevant.Human lawyers are still needed to interpret the meanings and subtle nuances of the case’s specifics and to have the EQ+IQ needed to make wise recommendations to their clients.I love lots of things AI can do which free up people’s time. So, for example, junior lawyers don’t need to be reading and sifting through hundreds of tomes when they could be learning more about client interactions or how to interpret complex cases to arrive at a successful outcome.BUT I hate when some folks claim AI is more intelligent than human just because it can correlate data faster or when they say silly things like, “AI is an emulation of the human brain and simulates evolution” — especially when they’re Professors of AI who should know their Science 101.AI is based on Logic+Probability. Logic was invented by Aristotle 2000+ years ago to model the “idealized, perfectly rational person.” Meanwhile, Probability was invented in 1654 to model the random (stochastic) behavior of fair and unbiased dice.The last time we checked, no perfectly rational person exists and no one thinks like dice. So AI is not reflective or representative of OUR intelligence.This fundamental difference doesn’t change no matter how much some AI folks say Neural Nets (and indeed Graph Theory) have been architected to replicate our neurons (nodes) and synapses (edges).There’s a ton of other organic mechanisms happening in the ecosystem of our embodied and integrated intelligence that the mechanistic, mathematical engines of AI simply can’t emulate.Now, Logic+Probability are good for pattern matching text in case law and doing Machine Vision.However, their limitations are becoming clearer in AdTech where 250+ brands have pulled their ad $$$ from Google because the AI basically can’t classify and discern between good and bad content and, importantly, WHY it’s good or bad.This is literally a fortnight after I was at a talk by Google Research in which they showed they can’t infuse intuition into Formal Reasoning (which is Logic+Probability, by the way).Meanwhile, intuition is de facto in human reasoning.Maths as a language is useful BUT it doesn’t (at the moment) map over to Natural Language Understanding.So … someone would need to invent a better system for NLU because Logic+Probability can’t get us there and they’d need to re-engineer 2000+ years of Aristotle’s Logic and 350+ years of Probability to do that.Fun stuff! Needs lots of ice cream! LOL.
which free up people’s time. So, for example, junior lawyers don’t need to be reading and sifting through hundreds of tomes when they could be learning more about client interactions or how to interpret complex cases to arrive at a successful outcome.Very possible that that reading and slave work is part of the process of learning though. Gorsuch did it. [1] Reminds me of when I used to slave to 7 hours to solve a problem that I can google the answer to today. I know I am better for it. [2] So good for the junior lawyer. Not good for the client or senior partners.One other thing is this type of work is dredge automaton partly in nature So it can be done with half a brain late at night when you are tired. With chinese food in the conference room that you eat directly out of the containers. Like in all of the movies and tv shows. Complex thinking (such as you suggest) can’t be done like that. It takes effort in prime time. (Really I am serious the other work is really autopilot work and we’ve all done it…)Also the question remains of whether the law firms will simply pocket the savings (but still bill for hours in some way .. I say they will) or do something with the junior lawyers that provides value to the client. My feeling is they will markup the work in a way that only nominally reduces the cost to clients.[1] He looks like George Clooney a bit.[2] And to that point had to stop ‘mom’ from helping the 12 year old with math problems last night and providing what I thought was to much assistance.
I get your point about how human intelligence strengthens when we put in the hours and slog through the learning rather than having the machines hand it to us on a dashboard.In our household, the calculator wasn’t allowed until we got to Quadratic functions, tanh curves and the such. It may be why I got an A in Maths in high school.Solving a problem mentally without a computer probably does train the brain differently.Computers have pros and cons:* https://www.theatlantic.com…* https://nplusonemag.com/iss…I do know that to invent my system I had to fly to the Accademia in Venice to pay homage to Da Vinci — even though there’s lots of information online.But, there’s something about seeing and experiencing things IRL and absorbing+transforming info through the lens of our intuitions and how our mind’s imagination very rapidly connecting and contextualizing the dots that the computers simply can’t do — despite all their data and the speeds of their probabilistic correlations.Anyway, I am for the humans-in-the-loop approach to AI. I simply believe we need to invent better tools to enable humans and the AI, especially in NLU.
Good points here. Casetext is a useful supervised learning classifier, but If you think about “understanding a document”, we need new models of unsupervised learning. The newer models will need a multiple step approach.
In my early 20’s, I was part of a team that built an M&A database in a joint venture with the ‘Financial Times.” That’s why I know how important data classifications and database architectures are for Machine Learning.The limitations of supervised learning are two-part:(1.) It can only pattern match based on historical data and how the ontology trees have been structured.So, if the law firm does a lot of commercial cases involving M&A in which terms like “white knight,” “revolving credit facility,” “all cash offer,” etc appear then they can certainly order-rank and weight those in the pattern-matching algorithm of that pre-classified data in Supervised Learning.However, if a new case appears with a word like “convertible equity” and that doesn’t exist in the data corpus then the AI would have a blind spot in the Natural Language Processing and so wouldn’t surface the relevant case for the lawyers to look into.(2.) Biases in the design of the data and the architecture.Even with Unsupervised Learning (where data is unstructured), it’s about more than multiple steps.The current approach is to do some type of concepts clustering. That blunt tool is fine for financial terms (which tend to, of and in themselves, be “structured” and well-defined; an interest rate is understood and can be quantitatively defined).It’s less useful for language-based data where ambiguity is involved and words can’t be quantitatively defined.For example, in law contracts, how do we quantify the term “applicable jurisdiction”?So for that reason, proper NLU is vital — which is why we need to invent better tools than what currently exists.
We are really saying the same thing. What I was getting at with multiple steps is this that the algorithms used for scoring the document needs to be a function of the use case. The trivial case of “convertible equity” vs M&A could easily be solved by “translating” the document into N combinations of pseudo-documents. The more difficult case is actually the opposite, how to find the needle in the haystack. How to find the opinion where the judge was furious/impressed/confused about the choice of about the the language in the M&A document. The newly approached have to assign probabilities to the the cluster themselves and come up with methods for doing so (the age of the article, the author, prior articles of the author….)
Probabilities aren’t the way forward.State of the Art AI and the leading thinkers in AI are stuck on how to get to Natural Language Understanding.“One possible approach is to have a system which would have a sophisticated and detailed understanding of the meaning of text—which is something that cannot be done yet today—then it would read many different articles from many different news sources and would look for inconsistencies, the same way a human whose full-time job is determining whether something is fake or not.” says Ilya Sutskever, Director of Research at OpenAI. [He was one of Geoff Hinton’s PhD students and worked at Google Brain.]“When 8/10 people say it’s fake news, is it fake news? What about 7/10?” Richard Socher, Chief Scientist at Salesforce. [He was the one who built Metamind, based on his work with Stanford Sentiment Treebank language classifier.]* https://qz.com/843110/can-a…
fake vs real isn’t the what this is about, it’s haystack with needle or haystack without.
Not all needles are the same, :*).Yes, if we’re talking about traditional search it is a “needle in haystack” problem.If we’re talking about being able to sew with that needle (understanding the text and its meaning), different Maths is needed.
If you haven’t already familiarized yourself, you’d love Kahneman and Tversky’s reserach.
Thanks, Adam.I agree with Kahneman: “Our brains aren’t built to follow the rules of probability.” Where I think Behavioral Economics is missing a piece of the puzzle is still the “WHY?” do people behave in that way (click the mouse, read the content, buy the product etc).Maybe the “WHY” doesn’t matter in a world where Amazon does one-click purchase and that behavioral outcome is what counts on the balance sheet!
This is very similar to the JSTOR Text Analyzer (http://www.jstor.org/analyze/), which was just released to help academics find relevant material. Regardless of the profession, advances like these are fantastic.
The monthly rate of $119/month/user seems like a old pricing model for a new technology (particularly given the technology). Let’s unbundle it, $20/month per specialty area (IP, Estate Law, Securities, etc.)
How long will CARA need to complete an analysis of a brief?Also, this expense is ultimately passed through to clients who are paying multiples of that per hour for top shelf legal work because they want to win.… they will want this edge and will pay eagerly. It’s a bargain.Wining a court case might be expensive, but losing is much more so. Just $119 for an hour of AI processing?Not even an issue.
The price is $119/month.
Wow, then the price is literally pennies per hour, effectively zero for an attorney and not a factor at all.Amazing.
When you are unrivalled, you don’t unbundle.
No MOAT, No monopoly.
Could be, but until 10’customers come to you and suggest they may switch to someone else due to packaging / pricing model, shut yer pie-hole and keep selling.Non-issue until then.No Moat is a mother reason to max customer revenue early on.
….and the unsuspecting victims welcomed the AI barbarians into the gates. Little did they know far from coming to help them, their real purpose was to replace them….
The AI can pattern match words in the text and surface sections of case law that might be relevant.Human lawyers are still needed to interpret the meanings and subtle nuances of the case’s specifics and to have the EQ+IQ needed to make wise recommendations to their clients.
So you’re part of the conspiracy too eh. Double agent for the barbarians.
Haha, LIAD.No. AI should not be replacing human employees. There’s tons of AI and automation in AWS, for example.However, notice that all the Solutions Architects and trainers are human. “And they’re hiring!” (Everyone in SV says this at their end of their presos.)I’m not a Singularity disciple. I don’t believe in Kurzweil’s ideas: “That leads to computers having human intelligence, our putting them inside our brains, connecting them to the cloud, expanding who we are.””Your personality, your skills are contained in information in your neocortex, and it is information,” Kurzweil says. “These technologies will be a million times more powerful in 20 years and we will be able to manipulate the information inside your brain.”* https://www.marketplace.org…The idea of intrusive Neural Lace (http://gizmodo.com/scientis… IS FREAKY.Yet some $billionaires are investing in that area — maybe with a view of making ‘The Matrix’ a reality.
They said the same thing about offshoring.Nope, never happened.
Newspapers did this for the internet you know. So pure so fresh and now free. Not only did they talk about it they talked it up. And in a big way.
Or… to free them.
“free them”
My first job out of college was at an L.A. law firm*, and my 2nd and 3rd were with SoCal-based litigation software firms that were doing basically exactly this– but it was 1990-94, running on Wang minicomputers, transitioning to PCs, so the results were primitive, but still lawyers would relish them.Would like to have helped you diligence this one. Maybe I still can help this company– I definitely get the need and business, core of which really hasn’t changed much since I was doing this.Unrelated: Quinn Emanuel is the firm now famous for helping ousted founders, such as the 3rd Snap guy. I’m a client.*When I went to Chinois on Main! (inside joke w/ Fred) 😉
Great application. These digital assistants are going to get stronger and transform the profession.
CONTRIBUTORS:ENTRAPMENT ALERT: BE AWARE OF THIS REAL CHARMERhttp://www.msn.com/en-us/mo…
Sounds like a great tool — lawyers will charge less now, right? In reality it’ll give them more time for other work and give them the ability to make better decisions faster.On a side note.. @fredwilson:disqus – Is there a possibility that SoundCloud could acquire TurnTable.fm’s technology and relaunch? I think it could add a lot of fun value.
Curious what the list of “AI for x-profession” is going to be? If law now, what’s next ?
Two thoughts. One, the legal profession is already an arbitrage between Really High Priced Partner, Still High Priced Associate, Moderately Expensive Paralegal and Cheap-ish Admin/Secretary.The client’s bill is, in part, baked using a combination of the above resources needed to get the job done, and the billable target that the client can stomach. The notion of further segmenting this with a Research Layer that is currently done manually at a Paralegal or Higher cost is an economic no-brainer.Two, and more fundamentally, law is by definition an industry where past precedence in terms of case law is a primary tool — if not the primary tool — in shaping legal outcomes. The notion that systems can add value to identifying patterns and specific “historical” candidate case laws, excerpts and the like is fundamental.”Show me what the winning outcomes look like, and how they argued that look like my case” is an easy to digest, intuitive narrative.Feels like a winner, though also a category with not a ton of defensibility.