YT YouTube2Text

YouTube2Text → Transcripts → Artificial Intelligence and Law – An Overview and History | Guest Speaker: Harry Surden

Artificial Intelligence and Law – An Overview and History | Guest Speaker: Harry Surden — Transcript

by Stanford Law School · 10,672 words · 1,588 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:04all righty thank you all for coming
0:07Kotick speaker series
0:09tomorrow is our code X future law
0:12conference just Harry pointed out is
0:15sort of the coachella of legal tech
0:19activities at least for us
0:22and Harry's our opening act so Harry was
0:25our first fellow in code X over 10 years
0:29ago by now and and he went on to become
0:33a law professor at University of
0:37Colorado in Boulder he's still actively
0:40involved in codex as affiliated faculty
0:43he started many important initiatives on
0:47the codex has been a very actively
0:49involved in our work around computable
0:51contracts many of you will know that
0:52Harry wrote the seminal paper and
0:55computable contracts and many other
0:57important works in technology but he's
1:01here today to give a general overview on
1:03a ai' in law overview in a history and
1:06so yeah so without further ado I'll just
1:09turn it over to Harry again and please
1:11join me in welcoming Harry certain all
1:17right thank you so much Roland it's so
1:19great to be back here at codex and at
1:22Stanford and all the great work that you
1:24and the other folks at codex do and all
1:27the great work that happens here all
1:30right so thanks for that introduction in
1:31terms of background as Roland mentioned
1:33I'm a law professor at the University of
1:35Colorado and also affiliated faculty
1:37here at the codex Center at Colorado
1:41I teach research and write about
1:43intellectual property law especially
1:45patent law and also the subject of this
1:47topic artificial intelligence and law
1:50why the interest in law and technology
1:53well prior to entering law I was
1:54actually a software engineer for several
1:56years I worked as a professional
1:58software engineer out there on the field
2:02so today I'm gonna talk about artificial
2:05intelligence and law artificial
2:09intelligence of course is also known as
2:11AI so I'll be using that interchangeably
2:13at points and I'm gonna try and
2:16understand how artificial
2:18Hodgins has been applied to law in the
2:21past and how it is currently being
2:23applied to law so I'll start with an
2:26overview of artificial intelligence
2:27generally because we are at at law
2:28school so it's not necessarily the case
2:31that everyone in the audience is
2:32familiar with the details of artificial
2:35intelligence the type of questions we'll
2:38be looking at is what is artificial
2:40intelligence what are the major branches
2:43and approaches where do we see
2:46artificial intelligence today outside of
2:48law oh and what are some of the limits
2:50and then I'll talk about artificial
2:52intelligence within the law
2:54I'll do a brief overview the history the
2:57thinkers who have thought about applying
2:59artificial intelligence techniques to
3:02different aspects of the law and then
3:04we'll move to the current where do we
3:06see artificial intelligence in law today
3:08why is going on in the state of the art
3:11and then finally I'll raise some really
3:13interesting and I think current and
3:15pressing policy issues in artificial
3:17intelligence in law that are really at
3:19the cutting edge of what policymakers
3:22are talking about and will serve as some
3:25fuel for thinking for all of us so
3:28without further ado let us begin with an
3:32overview of artificial intelligence
3:33generally so let's start with a basic
3:36question what is artificial intelligence
3:39well let me say at the outset that the
3:42phrase artificial intelligence has no
3:45universally agreed-upon meaning has lots
3:48of different meanings and means
3:49different things to different people and
3:51I won't attempt to get into all of them
3:53here however I find that one useful
3:57definition of artificial intelligence is
3:59the following
4:00AI is using technology to solve problems
4:04or make automated decisions or
4:06predictions for tasks that when people
4:09do them are typically thought to require
4:11intelligence so for example playing
4:15chess translating languages driving cars
4:17these are all tasks that are thought to
4:20involve some higher order cognitive
4:22processes when people do them they
4:25involve spiritual reasoning or thinking
4:27or abstract conceptualization
4:31vision etc so if a computer can solve
4:35such a task employing cognitive
4:40processes we call it an artificial
4:42intelligence task and if a computer can
4:44solve these same tasks even if they do
4:46it much differently than humans do it we
4:48think of that as an artificial
4:50intelligence task and a crucial point is
4:52that computers do solve these tasks like
4:55driving or playing chess very
4:57differently than humans do when they
5:00address these tasks a computer drives a
5:03car using techniques that are very
5:05different from the way humans do it all
5:09right so let me just start with a caveat
5:11that that is not a perfect description
5:13of artificial intelligence or AI not
5:15everyone would necessarily agree with
5:18that but I think it's a pretty good
5:21working definition automating tasks that
5:23require intelligence when humans do them
5:25and it's a pretty good working
5:27definition because many but not all AI
5:30tasks do involve the automation of
5:34activities that involve human
5:37intelligence
5:38so in describing what AI is I find it
5:43useful to start out by emphasizing what
5:46curt a AI current technology is not and
5:49this is to distinguish current a I
5:52technology from what many people think
5:55current AI technology is so when many
5:59lay people hear the term artificial
6:01intelligence they imagined that today's
6:05AI involves sneaking machines machines
6:08that think like people and this is a
6:10view of AI that sometimes called strong
6:13AI so let me emphasize this is the view
6:16of AI from the movies and from the books
6:18and TV of fully conversational computers
6:23or robots so if a computer had strong AI
6:27which they do not such a computer would
6:30it be able to think or talk at a level
6:32that would meet or surpass human beings
6:34so such an AI might be able to discuss
6:37abstract philosophy or engage in
6:40conversation and arbitrary topics let me
6:43just emphasize that strong AI is
6:45sigh totally not what we have today
6:47today's most advanced AI algorithms
6:50cannot think they cannot replicate human
6:53higher order cognitive skills such as
6:55abstract reasoning arbitrary
6:56conversations a typical two-year-old
6:58human has more advanced cognitive
7:00ability than the most advanced AI today
7:03so so why do I bring this up is because
7:06companies often advertise their AI as
7:09cognitive computing things of that made
7:12nature which lead people who are not in
7:15the field to think of AI is involving
7:18Thinking Machines but that is not
7:21accurate at all or sometimes you see AI
7:25startups talking about super intelligent
7:27robot assistants all that is just
7:31exaggeration and that's going to be a
7:34major theme of my talk so you'll also
7:38see on the news pundits are often
7:40worried speculating about super
7:42intelligent AI is that will you know in
7:45in the world super intelligent AI super
7:49computers so I spend a lot of time
7:51looking at the state-of-the-art in AI
7:53and in my opinion in looking at the
7:55evidence of the state of AI research
7:57there's no evidence that we're anywhere
8:00close to anything like strong AI or
8:02thinking computers anytime soon that's
8:04just my opinion but the point is that
8:07the term artificial intelligence at
8:09least as it is today is a bit of a
8:12misnomer given the state of today's
8:14intelligence what we call AI today is
8:16definitely not human level intelligence
8:18and it doesn't look anything like you
8:21would expect given the discussion in the
8:23media or the way the companies talk
8:26about it so we've talked about what a AI
8:29is not so what is AI well as I wish
8:33shortly describe in more detail what
8:35much of AI is today is what is called
8:39pattern based artificial intelligence so
8:41this involves the use of algorithms to
8:44search large amounts of data for
8:46patterns in that can be harnessed in
8:51useful ways and as we'll see this is
8:53goes under the formal name of machine
8:55learning now I don't want to
8:58understate today's pattern-based and
9:01machine learning
9:01AI technology it is an incredibly
9:04powerful tool in its particular context
9:07where it's useful so it has been used to
9:09automate things ranging from driving
9:12cars to language translation so in
9:15limited domains current AI technology
9:17can do amazing things it's just not
9:20unlimited what it can do it does not
9:23involve computers that can think or that
9:26can engage in human level abstract
9:28reasoning or cognition but rather it's
9:30really excellent pattern and useful
9:32pattern magic is also we will see as a
9:34lesser extent AI also involves knowledge
9:37representation and rules and deductive
9:40reasoning so why am I talking about this
9:43because a really important theme of my
9:45talk is we have to be realistic about AI
9:47we have to really understand what AI can
9:50and can't do what it's good at but it's
9:52bad at to allow us to have intelligent
9:56policy conversations because there are
9:58things that it can do and it can do
10:00quite well but we there are also things
10:03that it's very limited at so it's very
10:06important first of all to demystify
10:09artificial intelligence technology to
10:11understand what it really is in order to
10:14have sensible policy and law discussions
10:16and more to the point I think that
10:19rigorous analysis requires us all to
10:22resist speculation about the future of
10:26AI we really none of us myself included
10:28know where AI is going beyond in maybe
10:32the three to five year time frame so I
10:34would invite skepticism from all of you
10:37when you hear pundits out there in the
10:39world confidently predicting the future
10:43of AI and recommending law or policy
10:46choices on that basis because I think
10:47none of us know with any degree of
10:50certainty all right so let's with that
10:56high level view of AI let's dig a little
10:58deeper into the details of AI so I think
11:03there are a few useful ways of
11:05understanding artificial intelligence
11:07and one of the most helpful ways is to
11:10understand the major
11:11approaches that have been used
11:13historically to implement artificial
11:16intelligence and I think we can think of
11:19AI is really a set or a suite of
11:22different computer techniques and
11:25mathematics techniques that have been
11:28used to solve various AI problems so I
11:31like to think of it as a set of tools in
11:33a toolbox that can be picked out and
11:36adapted to various problems so AI is not
11:39one thing but it's more like a suite of
11:42things that have a family resemblance to
11:43one another but with that in mind let's
11:46talk about I think we can roughly divide
11:49a I approaches into two large approaches
11:53and this will be very approximate the
11:56first approach
11:58let me call computer logic or rules
12:01based systems and the second approach
12:03we'll talk about is machine learning so
12:08let's talk a little bit about logic
12:10rules and knowledge representation based
12:12AI which is the one approach so rules
12:18and knowledge based systems are part of
12:19a larger category of artificial
12:21intelligence of process sometimes known
12:23as knowledge representation and the goal
12:26of knowledge representation is to model
12:28real world processes or systems in ways
12:32that informs that a computer can use in
12:35process so often the logic and rules
12:39based method in knowledge representation
12:42involves a top-down approach and I'll
12:45explain what that means in a second but
12:46basically it means that you have
12:48programmers or groups of engineers
12:50observe some real-world tasks often
12:54involving experts which has certain
12:57real-world constraints and rules and
12:59then they attempt to model that
13:00real-world process in computer form in
13:03terms of computer rules and
13:05relationships so a really good example
13:07of this actually comes from law which is
13:09TurboTax so many of its tax season now
13:13so many of you are perhaps using some
13:15tax software like turbo tax for tax
13:19assistance and the turbo tax is actually
13:22a really good example of a knowledge
13:24representation
13:25based version of artificial intelligence
13:28so the company that created TurboTax
13:33when they're trying to create that
13:37software what they're trying to do is
13:38model something out there in the world
13:40and that something out there in the
13:42world is law it's the personal income
13:44tax code a set of real world constraints
13:47out there on people's legal liabilities
13:51and duties so what they're trying to do
13:54is model the personal income tax code in
13:57a computer in a way that faithfully
14:00represents the logic and meaning of the
14:03laws out there in the world so in that
14:06case the programmers from Intuit the
14:08company that makes TurboTax
14:10in consultation with lawyers engineers
14:13accountants and other will examine the
14:16personal income tax code try and
14:18understand it and see it as a set of
14:21rules legal rules out there in the world
14:23and attempt to translate the logic and
14:26meaning of those legal provisions into a
14:28set of comparable computer rules that
14:31the computer can follow so so say for
14:36instance there's a tax law out there
14:38that says if you make above ninety one
14:41thousand dollars your marginal tax rate
14:43is twenty eight percent we can all read
14:46that as attorneys and understand what
14:48that means you can also take the logic
14:51behind that and create a computer role
14:53that implements that logic which looks
14:56at what your income is and applies the
14:59correct marginal tax rate on that amount
15:02all right so that's just an example of
15:05taking a law out there in the world and
15:08modeling it as a legal rule faithfully
15:11representing the logic um so what's the
15:14point why is that useful well I think
15:16what's really interesting is that once
15:18you have laws represented in computer
15:21processable form you can do all sorts of
15:23interesting things with it you can
15:25engage in computer deductive reasoning
15:29that is basically difficult or
15:31impossible for humans to do on their own
15:33so computers can figure out things by
15:36processing rules
15:38in very complex chains that humans can't
15:41figure out and there's sometimes
15:43non-obvious conclusions in there so for
15:46example say you're on a business trip
15:49and you're recording all your tax
15:52expenses right there is a certain level
15:56of rules about what you can deduct or
15:59not for your business expenses a human
16:02can determine these rules but not
16:04without a lot of laborious calculation
16:07and reading but once you represent the
16:10laws in computational form it's kind of
16:13trivial for a computer to understand
16:15that an expense has been marked as a
16:17business expense and then apply the
16:19correct laws to that to correctly assume
16:22your tax liability so what's interesting
16:27about this is that this is basically a
16:29model of the personal income tax code as
16:32determined by into it but it's such a
16:33good model that over sixty million
16:35people in the United States use this or
16:38something like this and the IRS
16:40basically defers to the computations
16:43that have been created by this expert
16:46system so this is an example of what is
16:49known as an expert system where you've
16:52modeled you've taken some of the
16:54knowledge of experts like lawyers and
16:56certified public accountants out there
16:59in the world attempted to replicate it
17:01computationally and then you can use
17:03that knowledge to produce useful
17:04real-world results such as tax liability
17:09so as we will see a lot of the early
17:12history of artificial intelligence
17:14involved knowledge representation and
17:17also in law creating expert systems in
17:20medicine law and other domains where
17:23expert knowledge could be codified and
17:26usefully represented computationally
17:30okay so that was the knowledge base and
17:34rules base approach let me talk about a
17:36very different approach to AI known as
17:39machine learning so in the last twenty
17:43years or so a relatively new suite of AI
17:47techniques called machine learning began
17:49to become really popular and
17:51well it is what I believe really caused
17:53artificial intelligence to take off in
17:56the way that we see it today
17:58so at a very high level machine learning
18:03involves the use of algorithms that
18:05examine large amounts of data detect
18:07patterns in those data and then you can
18:09often harness these patterns to do many
18:12useful things machine learning is
18:15sometimes abbreviated ml so I'll be
18:18discussing that using that term
18:20interchangeably but here are some
18:22examples of machine learning out there
18:23in the world that you're probably all
18:25familiar with when a car drives itself
18:29some or most of the aspects of that
18:31involve machine learning I'll explain
18:34that more in a minute when Netflix or
18:37Amazon suggests a movie or a product
18:40that you might like based upon your past
18:42hurt purchases that involves machine
18:44learning when your email as we shall see
18:46tells you that an incoming email from
18:49Nigeria is likely spam that is also an
18:53example of machine learning it's been
18:55used in many different things you can
18:57find non-obvious relationships in data
19:00such as the risk between your credit
19:03score your FICO credit score and your
19:05risk of getting in an accident the
19:06correlation is very high you can spot
19:09fraud credit card fraud automatically or
19:12spot tumors and x-rays machine learning
19:14is all over the place and it's fair to
19:17say that machine learning is responsible
19:20for most of the AI headlines that you
19:22hear about in the news today so there I
19:25think they're rival of machine learning
19:26has made artificial intelligence what it
19:29is today so it gets most of the
19:31attention um so again it's helpful not
19:34to think of machine learning as a thing
19:37but rather a suite of different
19:39approaches that also share some family
19:42resemblances to know one another we can
19:45think of that as a toolbox of machine
19:47learning techniques so you might hear
19:51some of these different techniques out
19:52there in the world like neural networks
19:54or deep learning neural networks or
19:56logistic regression these are all
19:58different techniques that roughly share
20:00some same characteristics so let's talk
20:04a little
20:05what these characteristics are that make
20:08something considered a machine learning
20:10algorithm the first thing to understand
20:14is that machine learning algorithms are
20:15intentionally designed to learn and will
20:19see that I put that in quotes for
20:21reasons I will explain but roughly
20:24they're learning because these
20:25algorithms are designed to be able to
20:27improve their our performance over time
20:29on some tasks often by showing it good
20:33and bad examples of what you want it to
20:36do so I put the term learning in quotes
20:39because learning is really a metaphor
20:41for human learning it does not mean that
20:44it is actually learning in the way that
20:46humans learn that's a common mistress
20:49misconception it doesn't mean that it's
20:51artificially replicating the higher
20:54order cognitive processes that are
20:56thought to me involved in human learning
20:59rather we can think of it learning is in
21:02a functional sense so what do I mean by
21:06that well we say that a human is
21:08functionally learning when he or she is
21:11able to get better over time on some
21:13task through experience by that same
21:16token if we see a artificial
21:19intelligence machine learning algorithm
21:21being able to improve its performance
21:24over time on some tasks like driving or
21:27credit card fraud detection with more
21:30experience we can say that's kind of a
21:32functional metaphor for learning so in
21:34that sense we could think of it as
21:36learning but it's not human level
21:39learning the second point about machine
21:42learning is that it involves pattern
21:44detection so at a very basic level much
21:47of machine learning can be thought of as
21:49very sophisticated pattern detection
21:51systems designed to detect non-obvious
21:54patterns or correlations in large
21:56amounts of data and then we try and
21:59generalize these patterns to use them in
22:01other related contexts to make useful
22:05decisions so these patterns are encoded
22:08in a computer model in some ways and
22:10we'll talk about sometimes the how
22:13they're encoded is not very
22:14interpretable to people and other times
22:17it is
22:18but so what do I mean for patterns well
22:22let's just say that Netflix is trying to
22:24figure out what movie you're gonna like
22:27it's gonna look through all of its data
22:29and try and find people who are like you
22:32according to some machine learning
22:35measure of likeness maybe the past
22:38movies that you liked or maybe some
22:39other demographic data and what's
22:42interesting is the patterns are all
22:44there in the data already is just the
22:47machine learning algorithm that is the
22:49ability to pull it out and show it to
22:52you so almost always the data is the
22:56patterns that you want are in there it's
22:58the the art and science of machine
23:00learning to pull out those patterns in
23:02useful ways another point about machine
23:05learning is that they require data often
23:08large amounts of data to detect the
23:10patterns that can be used to automate
23:12decisions so the ability to have machine
23:15learning goes part and parcel with the
23:17availability of relevant data we're
23:19gonna see in law that's kind of a
23:21problem because not a lot of legal data
23:23is available in highly processable form
23:27in this day much of it is behind
23:29paywalls or non easily processable
23:33structures so that's something to think
23:36about and then finally I won't dwell on
23:38this but I think it's an interesting
23:40point machine learning often is thought
23:43of as software they can program itself I
23:46think it's a little overstating the
23:48point but I think there is an
23:49interesting idea in there so machine
23:52learning programming is contrasted often
23:56with the top-down structured programming
23:58that you typically see software done so
24:02if you're creating a piece of software
24:03like Microsoft Word you get a bunch of
24:05programmers in a room and they write a
24:08bunch of software code that give the
24:10software different capabilities in a
24:13very top-down
24:14structured approach and the computer
24:17ultimately Falls these instructions and
24:19these instructions create a web browser
24:21or a word processor depending on what
24:24instructions
24:25by contrast machine learning is a very
24:27different approach
24:28you basically often will just really
24:32the algorithm on some data and then give
24:34it a bunch of examples of good behavior
24:36and bad behavior and the Machine kind of
24:39figures out on its own what activities
24:43lead to good behavior and what leads to
24:46bad behavior so for example if you were
24:48programming a car by the traditional
24:51top-down approach a self-driving car
24:53programmers might have a bunch of rules
24:55whenever you approach a stop sign you
24:57stop whenever you know you're at a green
25:00light you go etc etc in a lot of cases
25:04in self-driving cars you can program the
25:07cell functionality but by machine
25:09learning the machine itself figures out
25:11example by looking at examples of good
25:14behavior and bad behavior it figures out
25:16how to drive or you might give it a
25:19bunch of examples of pedestrians and
25:21cars and without telling it the
25:24particular characteristics of what makes
25:26a pedestrian a pedestrian and a car a
25:28car it can figure out on its own and be
25:32able to tell the difference between the
25:34two so in a way it's programming itself
25:36to determine the distinction between
25:38pedestrians and cars as it's driving
25:40down the road so that's just a good
25:44example so that is one of the
25:46characteristics of machine learning it's
25:48much less top-down rules based
25:50programming but more bottom-up
25:52programming by often by giving examples
25:57I'm gonna just give you a quick example
25:58of machine learning and then we will
26:01move on to a I as it is in law but a
26:06really common example that people give
26:08because I think it's just very
26:09understandable is email spam filters so
26:12you we all know that we're out there in
26:15the world looking at our email and when
26:17email comes in you most software
26:20programs have the option for you to mark
26:22it as spam
26:24so how it does that tell your email
26:27program that a future piece of email is
26:31likely to be spam or not well basically
26:35they're often using a machine learning
26:37algorithm that's designed to detect
26:40telltale patterns of spam based upon
26:42what you've given it so what every time
26:44you mark something is
26:46you're telling the machine learning
26:47algorithm hey analyze this and look for
26:50what characteristics this email has that
26:54makes it spam and then every time you
26:57don't mark something as spam you're
26:59telling it hey look for patterns here
27:02that make it tell me that it this is not
27:06spam so that's an example of training
27:09and algorithm your training it by giving
27:12it known examples of spam or not spam so
27:16let's imagine for example that after 100
27:19spam emails you're a machine learning
27:21algorithm determines that the phrase or
27:24in cash is statistically really likely
27:28to be associated with spam right why
27:30because it's detected earn cash in know
27:36wanted emails and it detected or in cash
27:40in a huge amount of spam emails so the
27:42next email that comes in that has earned
27:45cash it's gonna think it's statistically
27:47at the high likelihood of being spam
27:50right so there's a couple interesting
27:52points in there first of all that's an
27:55example of a machine learning algorithm
27:57detecting a pattern in law you can also
28:01it's an example of the machine learning
28:02algorithm learning right it's learning
28:05that earn cash is associated with spam
28:08and you can use that pattern to make
28:11other predictions another interesting
28:16point about that is that the
28:17identification improves over time so it
28:20might learn another pattern that that
28:24words that have emails that have free in
28:29them are statistically more likely to be
28:31spam or emails from the country Belarus
28:34or Nigeria are much more likely to be
28:37spam so it's able to learn over time
28:39improve its ability to distinguish spam
28:42and email spam and wanted email so this
28:48is the idea I want to get to is if
28:51machine learning does not have an AI in
28:55general doesn't have the cognitive
28:56abilities of humans how is it
28:59that were able to have AI making
29:03intelligent decisions and I think the
29:04answer is that we use proxies or
29:06stand-ins so basically whenever you find
29:09a pattern out there in the world that
29:13can be a useful proxy or stand in the
29:15computer doesn't actually have to
29:17understand the deep conceptual layer in
29:19order to make an intelligent decision
29:21for say what a spam or what is not spam
29:24right so for instance when an email
29:26software is looking at an email to
29:30determine whether it's spam it doesn't
29:31know what cache is or free is doesn't
29:34know where the meaning of that doesn't
29:35have to know the meaning of that why
29:37because it's found a useful proxy a
29:39statistical correlation between email
29:43and the presence of those words and spam
29:45so this is what I call intelligent
29:47results without intelligence and that's
29:49what we see largely in the state of AI
29:53today where we're using patterns and
29:56proxies that can be harnessed to make
29:58useful automated decisions without
30:00actually having underlying higher-order
30:03human intelligence another example of
30:07this is Google's statistical machine
30:09translation which is pretty good these
30:12days if you've used it that also uses
30:15patterns or proxies to translate so when
30:18so often Google famously figured out
30:22that you could use UN corpuses so the
30:25united nation has for 50 years
30:28translated many documents into the six
30:32official language of the United Nations
30:34using high quality translators so you
30:36have these identical documents in
30:38multiple languages saying the same exact
30:41thing and you have hundreds of thousands
30:43or maybe millions of these so Google
30:45figured out hey let us release a machine
30:48learning algorithm on this and maybe
30:51it'll start figuring out correlations
30:53that you know oficina de correo means
30:56post office in spanish without us
30:58actually having to tell it that and
31:00that's basically what what has happened
31:04and we have these really pretty robust
31:08translations but importantly the
31:10computer doesn't understand what
31:12talking about it doesn't know what a
31:13post-office is it doesn't have to know
31:15what a post-office is it just has to
31:17know that certain words appear together
31:20with statistically high frequency all
31:25right so that's AI today as I see it
31:29summary of two major approaches you have
31:33the logic and rules based knowledge
31:35representation approach which leads to
31:37expert systems and embedded law and
31:39computational law like TurboTax and you
31:42also have the machine learning or
31:43pattern based approach one final point
31:46is I want to say that you don't have to
31:48choose one or the other many successful
31:50systems use both of these approaches so
31:53self-driving cars use a mix of rules
31:56based systems and machine learning
31:59systems very effectively so that's one
32:03point those are hybrid systems I don't
32:06want to imply that one or the other
32:07another point about hybrids I want to
32:10point out is that we often have this
32:12idea that artificial intelligent systems
32:15are replacing humans but that's not
32:17necessarily the case rather it's often
32:20better to think about it as enhancing
32:23humans it's often providing humans with
32:26the ability to make better decisions so
32:30computer chess can be the best human
32:34computer player in the world but often
32:36when humans you combine a computer chess
32:40program with a human chess player they
32:42can beat many of the best computer chess
32:44programs because the humans chess
32:46playing ability is being enhanced by the
32:49computer another important point is that
32:52we often talk about artificial
32:54intelligence as being fully autonomous
32:57that's the image which means that the
33:00system is on its own making the most
33:03important decisions about its own
33:04activities like a self-driving car right
33:08so a fully autonomous system is out
33:10there the system's making its own
33:11decisions the reality is many important
33:14AI systems still have humans in the loop
33:16which is the idea that the computer will
33:20often ask for help from a live human to
33:24make a judgment decision
33:26or an abstract conceptualization about
33:28an area where the computer just is not
33:30able to perform so a good example is the
33:33state of the art in self-driving cars
33:35these days is to have a call center
33:37where a computer self-driving car gets
33:39stuck it calls a human in a room and
33:43then a human basically takes over the
33:44car and steers the self-driving car out
33:47of trouble and then gives it back to the
33:49self-driving car and that's just a very
33:51common conceptualization just a point
33:54you see that almost in many AI systems
33:56it's not fully autonomous but it still
33:58involves a lot of human judgment all
34:03right
34:03so let's talk a little bit about AI and
34:08law so I spent a lot of time talking
34:10about artificial intelligence generally
34:12which i think is necessary before we can
34:14understand a I applied to law you just
34:18briefly talk about the history of AI and
34:21law alright so we can think of AI and
34:24law as going as far back as live myths
34:28perhaps even further so live myths was
34:31one of course one of the co-inventors of
34:34calculus of really famous mathematician
34:36he was also trained as a lawyer and as
34:39far back as the 17th century
34:41Leibniz expressed the idea that perhaps
34:43we could use math or algorithms to make
34:47law more predictable more manageable
34:50more useful and that's really the heart
34:53of AI in law the heart of the law is
34:56using computers and mathematical
34:58formalisms
34:59to make law more understandable more
35:00accessible more useful more predictable
35:03more manageable so he was kind of on to
35:06something there and I've got a fun aside
35:08here for the law students here live Nets
35:12was a lawyer who contributed major
35:14things to math so it turns out many
35:17lawyers have made important mathematical
35:19contributions out there so for instance
35:23the mathematical topic of linear algebra
35:25which is basically the mathematical
35:28backbone of most of machine learning
35:30these days was partially founded by
35:33lawyers two of the founding fathers of
35:36linear algebra arthur cayley and james
35:39Sylvester
35:39we're both lawyers and if you look in
35:41the history of mathematics many of the
35:44most famous mathematicians were actually
35:46lawyers as well like live Nets Kaylee
35:50and Sylvester so there actually is a
35:52long tradition of lawyers influencing
35:54math in this interconnectivity between
35:56law and technology so I'd say the real
36:01modern era of AI law happened right
36:04before the dawn of computing it might go
36:06all the way back to Lila Berringer who
36:09came up with this idea of Jura metrics
36:11which is basically saying hey we should
36:13use more statistics and technical
36:19analysis to understand the law that was
36:21back in the 1940s and then really the
36:25heart of AI in law happened from the 60s
36:28to the 80s there are a lot of early
36:30pioneers here I won't name them all but
36:32some of them are listed up there and I
36:35call that the error of knowledge
36:37representation because that was really
36:39what was going on in AI at the time it
36:42was a rules-based and knowledge
36:44representation based system and many of
36:47these early computer scientists started
36:50applying these ideas of representing
36:52roles to law trying and in this era you
36:58saw a lot of attempts to model legal
37:00argument formally in ways that computers
37:03can understand they attempted to model
37:05legislation or regulation ways that
37:08computers can understand there were
37:10early to text extraction and some of
37:13these people might even be here in the
37:16audience because there was a lot of
37:17pioneering work done at Stanford and I
37:21apologize if you're a pioneer and you're
37:22not up there there are many more people
37:25out there important idea important
37:29conferences that came out of the 80s
37:31where iCal and Jurek's which still
37:32exists to this day which are conferences
37:35dedicated to AI and law most of the AI
37:38in law research however occurred in
37:40Europe and not in the United States
37:42Italy the Netherlands England and
37:44Germany really their hot beds and even
37:46to this day remain important centers so
37:50that's the error of knowledge
37:52representation
37:53I'd say the next era is kind of mm to
37:56the president which is the era of legal
37:58tech and machine learning and that's
37:59where we see ourselves now a founding
38:02moment and that is the founding of
38:04Stanford codex by roland michael generis
38:07Josh Walker and all the Codex fellows
38:10who came in back in 2006 before AI was
38:14really the rage legal tech startups
38:18we've seen a lot of legal tech startups
38:20in the last 12 years that have come out
38:22some out of Kodak's some in four other
38:27venues a good example of a legal tech
38:30startup that came out of Stanford is Lex
38:32machina which was known as the Stanford
38:35intellectual property litigation
38:36clearing galas back in the day which I
38:38was involved in when I was a fellow here
38:41spun off and is now a company so that's
38:46kind of where we are in the history of
38:48AI and law so let's talk about AI in law
38:52today where are we seeing artificial
38:55intelligence being used actually being
38:57used as a lot of hype but it actually is
38:59being used in law today I'd like to
39:03think about there being three categories
39:05of AI and while use there are the
39:07administrators of law these are the
39:10people who create and apply the law so
39:13judges legislators government officials
39:15regulators how they use AI we have the
39:18practitioners of law these are attorneys
39:20and other people who navigate the legal
39:22system and then we have the users in law
39:24people like you and me businesses out
39:27there who are bound by the law and have
39:30to comply with it and are looking for
39:32help to deal with the law so let's talk
39:35about each of these in term so the first
39:41case is AI used by administrators of law
39:43so these are people create and apply the
39:45law and often we're thinking here of
39:48legislators or judges so one important
39:51area we're seeing AI quite a bit comes
39:53in the criminal sentencing area where
39:56criminal defendants I'll talk about this
39:58more in the in the couple minutes but
40:00you're seeing judges starting to rely on
40:04sentencing and bail reports
40:06and these bail reports have been
40:08produced by a computer machine primarily
40:10machine learning algorithms that give
40:12risk of reoffending and the judges are
40:15not bound by these reports but often
40:16they follow them and this is in bail and
40:19sentencing another example we see this
40:22is in government benefits decisions out
40:24there legal officials or using this to
40:28make decisions about certain whether
40:29somebody qualifies to get a certain
40:31benefit or not there's certain AI
40:33components the policing we're starting
40:35to see this quite a bit we're starting
40:39to see predictive policing for likely
40:42crime areas so a lot of police systems
40:48use AI to try and predict areas where
40:50they should devote more resources based
40:53upon past climb patterns facial
40:55recognition for suspects is happening
40:57really quickly a lot of police areas are
41:00using AI to try and detect suspects in
41:03crowds or from group photographs so that
41:07there's some I'll talk about this more
41:09in the Q&A another really important area
41:12and part of the legal tech revolution is
41:15the use of AI by Google practitioners
41:18primarily attorneys by others in the law
41:20the most famous example is youth the use
41:23of machine learning in technology
41:26assisted review document review used to
41:28be the bane of first and second-year
41:31associates law sources who would sit in
41:34a dark room physically looking through
41:36documents and being built out by their
41:38partners at $400 an hour to basically do
41:42pattern detection at this point we've
41:46started to use machine learning
41:47algorithms which are better and faster
41:50at detecting documents that are likely
41:54to be relevant or privileged to
41:56litigation however humans are still in
41:59the loop quite a lot so this is kind of
42:01a common misperception it's not that
42:03we're taking lawyers out of the job
42:04we're just kind of shrinking the
42:06haystack of irrelevant documents and we
42:09bring in lawyers to look at the
42:11documents that are most likely to be
42:13relevant to the proceedings at hand
42:16I'll go just go over these quickly but
42:18we're seeing AI and legal analytics
42:20out there trying to predict how a judge
42:23might rule on a particular case or how a
42:25venue might treat a favorable case based
42:28upon past outcomes case outcome
42:30prediction you there a lot of hedge
42:34funds or not a lot but some hedge funds
42:36are starting to use various algorithms
42:38to try and bet on litigation finance by
42:43using various predictive algorithms a
42:45big area that we see or legal document
42:48assembly automated patent writings
42:51automated contract assembly again these
42:53often do a first cut of a document
42:56putting together and then a lawyer will
42:58put it together
42:59automated document analysis and
43:02diligence so you might sick your machine
43:04learning algorithm out there in the
43:06world or if you have a merger for
43:09example and you're trying to locate
43:11relevant contracts instead of having
43:13first-year associate look through ten
43:16thousand documents you might unleash
43:18your machine learning algorithm to try
43:20and find documents that are possibly
43:23relevant to the merger natural language
43:26processes there's a bunch of other
43:29they're computable contracts starting to
43:30happen out of codex legal research is
43:35starting to have some basic AI law firm
43:38management and case docketing so a lot
43:42of activities going on in the lawyers
43:45area and then finally the poor people
43:48like us who have to go out there in the
43:51world and comply with all these laws
43:53that rule us on a daily basis there we
43:56can divide this up into two areas
43:58businesses and ordinary people
44:01businesses are starting to make a lot of
44:03use of machine learning and rules based
44:07system into automated compliance
44:08particularly the gdpr we can thank is
44:12the full employment act for privacy
44:14lawyers but also for legal tech startups
44:17it's created a lot of compliance issues
44:20and it's created a lot of interest in
44:23using machine learning and rules engines
44:26for complying there's a lot of expert
44:28systems companies are using kind of you
44:32know compliance
44:34expert systems for ordinary people there
44:38are these expert systems like TurboTax
44:40that we looked at legal chatbots
44:42like do not pay you may have seen that's
44:45sort of a common thing where they help
44:46to these chat BOTS
44:48and self-help systems try and help you
44:51navigate the law and people are starting
44:53to use automated document assembly
44:55that's you know will maker and things
44:58like that that help people put together
45:00their own personal documents okay so
45:06limits and artificial intelligence I
45:08want to return to this because I think
45:10is a really important point we don't
45:12want to exaggerate what AI can do
45:16because I think that serves nobody here
45:19except for maybe the legal tech startups
45:20there are still many things beyond the
45:24realms of current AI technology so first
45:27of all a I still requires data right so
45:32if it's an expert system it requires
45:34expert data which is expensive to
45:37collect and if machine learning it
45:40requires some sort of data whether it's
45:43contracts or rules that the algorithm
45:45can look at there's no thinking
45:48Computers Computers can't engage in
45:51abstract reasoning and also there are a
45:53lot of accuracy limits so if you look at
45:56Google Translate it's pretty darn good
45:58but you wouldn't want to have Google
46:01Translate translate your billion-dollar
46:03merger agreement into French right why
46:05because the accuracy is not quite good
46:08enough you'd want an expert human
46:12translator doing that and then one of
46:14the biggest problems is sometimes the
46:16systems are hard to interpret
46:20alright so finally I just want to raise
46:22a couple of current policy topics with
46:25AI and law and they're now happy to talk
46:27about these in the question so these are
46:29what our law professors and other people
46:32in the policy area are really thinking
46:33about in AI and law so one thing that
46:37will be of interest to law students in
46:39this room is the automation of legal
46:43jobs and so there's this idea
46:47is artificial intelligence gonna Auto
46:50automate away most or all jobs in the
46:53world and also in law that same question
46:56is here and I'll talk about some of the
46:58economics research just briefly here so
47:02I think it's useful to separate this
47:04into two ways of thinking there
47:06short-term impacts and long-term impacts
47:09in the short term AI like other
47:12technologies whether it's automated
47:14forum implements or factory robots does
47:17displace workers in those areas so that
47:20is a real short-term change we're gonna
47:22have to deal with as a society for
47:25automated truck drivers and we'll see
47:28just some degree although a lesser
47:30extent in law as well but there's a
47:33different question is where there law
47:35whether technology destroys jobs
47:39permanently and I'll talk about that in
47:43a minute but the the important point is
47:48that most the research most economists
47:50have show this is that technological
47:51change like AI ultimately creates more
47:55new types of jobs than it displaces
47:58right so the net number of jobs
48:01tends to be more how that's the
48:03long-term view about people individual
48:05people tend to be displaced which is a
48:08real social problem that we can't ignore
48:11so in law there's there's a similar
48:13story I think certain legal tasks and
48:16certain types of jobs in law will be
48:18automated away but as I will talk about
48:22I think new types of legal jobs will be
48:24created that we just can't imagine right
48:26now so often there's just a failure of
48:28imagination on our part so what I told
48:32my boss students is lawyers do a mix of
48:36tasks from the highly abstract to the
48:39routine and mechanical so they do a mix
48:40of both and traditionally they could
48:42charge for both
48:44so today's AI technology is much more
48:47likely to be able to automate a task if
48:49there's an underlying structure or
48:51pattern that it can harness so I'd like
48:53to say in tasks where lawyers are acting
48:56like computers today those tasks will be
48:58replaced by computers tomorrow so
49:00of the routine mechanical work that
49:02lawyers are able to get away with
49:04charging for today is gonna be automated
49:07away um however that's not a lot of what
49:10lawyers do a lot of what lawyers do
49:12involve all these other things that a I
49:14is actually pretty bad so lawyers do
49:17abstract thinking problem solving
49:19advocacy client counseling things that
49:22involve human emotional intelligence
49:23Policy Analysis big picture strategy
49:26creative thinking etc etc all of these
49:29things are things that a I current AI
49:31technology is really bad at so I think
49:33those things are less likely to be
49:36automated away in the near term at least
49:39in the foreseeable future so that's I
49:44think that's an important way to think
49:45about it so if you're a law student I
49:47would think about building up your
49:50skills in those areas that we just
49:52looked at there the other point is that
49:57as I mentioned the creation of new legal
49:59jobs so when computers came along people
50:03used to do mathematical calculations
50:05right they were calculators or computers
50:08those jobs went away but entirely new
50:11categories of jobs were created so
50:13things like slop game computer game
50:16designer that didn't exist 60 years ago
50:18or data analysts these are entirely new
50:20categories of jobs that were enabled by
50:24new technology I think similar things
50:26will happen in law legal data analysts
50:28who knows we just have a failure we're
50:30good at imagining the bad we're hard
50:32it's hard to imagine the good so we just
50:34don't really know what new jobs but I'm
50:38pretty confident like most technologies
50:40there will be new categories of jobs
50:41created by new technology in law like AI
50:46this is something I'm happy to talk
50:48about during the QA increasingly as I
50:52mentioned judges are using AI or at
50:55least reports created by AI is a famous
50:58compass system you may have heard of
51:00that will give a defendant's risk of
51:04reoffending to a judge it's a 1 to 10
51:05score and judges aren't required to
51:08abide by that but often they just refer
51:10to it why after all if it says it's got
51:12a 8 out of 10 chance
51:14offending are you gonna be the judge who
51:16wants to bet on the two out of ten
51:18probably not
51:19you're probably going to defer but
51:21there's a lot of issues that I think are
51:23worth talking about
51:27first of all these can these reports are
51:29created by private companies who use
51:31machine learning algorithms maybe so in
51:34a sense we're sort of outsourcing some
51:36of these decisions to private companies
51:38often we don't know the data that went
51:40into it
51:44so is that problematic another issue is
51:48what if there are biases in the data
51:49this is a real issue so a lot of these
51:52risk of reoffending scores are
51:55calculated based upon police data so
51:58they record crime and characteristics of
52:01criminals but what if the police
52:02practices themselves are biased all
52:04right so say you have two defendants one
52:06who is a minority and one who's not and
52:09a police force systematically arrests
52:12the minority defendant for a similarly
52:16situated non majority defendant right
52:18there's gonna be a bias embedded in the
52:21data and then if that data gets recorded
52:24and it is used to train a machine
52:26learning algorithm those biases will be
52:28subtly embedded in the machine learning
52:31model and might cause us to overestimate
52:33the risk of reoffending of certain
52:36groups so that bias is another major
52:39concern another concern is that these
52:42machine learning models are often hard
52:43to interpret we don't know what they
52:46mean the kind of there are more of a
52:48black box and that might be an issue
52:54another problem out there is that these
52:57machine learning systems give you the
52:59illusion of objectivity so when a
53:03machine learning report tells you that
53:06somebody has you know eighty seven point
53:08three nine four percent chance of
53:10reoffending
53:11that sounds pretty precise and damning
53:14but in fact there's all sorts of
53:16uncertainties and subjectivity that go
53:18in this but there's research that shows
53:20that people tend to defer to things that
53:23sound mathematical and have an air of
53:25false precision so when government
53:27they're making decisions based upon this
53:29is this something we should be concerned
53:32about that's something I call the
53:34illusion of mechanical objectivity in
53:37fact all these systems have tons of
53:39subjective decisions that go into them
53:41what data to choose what data to exclude
53:44how far back you go
53:46what do you emphasize for the judge what
53:48do you deep prioritize how do you wait
53:50things vary a lot of subjectivity in a
53:53seemingly objective term another problem
53:57is AI is becoming easy to use really
54:00easy to use we're packaging up systems
54:02for government officials to use but
54:04these are often lay officials they don't
54:06really know what they're doing so if you
54:07give a police department a facial
54:10recognition system with 20 different
54:12parameters that can be tuned are they
54:15going to know how to tune that in order
54:16to reduce the likelihood of in correctly
54:21identifying minorities probably not
54:23they're just gonna use the off-the-shelf
54:26settings and that's a real problem
54:28because a lot of times these systems can
54:31only be used when they've been properly
54:34calibrated by people who really know how
54:36to use them
54:37moreover laypeople don't often
54:40understand the subjective nature of the
54:43design finally there are privacy issues
54:45but I won't go into that because I want
54:47to leave time for questions so
54:56any questions anyone in the audience you
55:02sir had a question before it was just a
55:06terminology question on one slide when
55:08you talked about using AI to determine
55:11the distribution of government benefits
55:14benefits is a very strong term I was
55:16wondering if you could explain what what
55:18kind of benefits that's referring to or
55:19give an example
55:20yeah certain government benefits so
55:24there are certain states that might
55:26determine whether you qualify for you
55:30know getting some state government
55:32monies for various things and then
55:34they'll use it used to be a human who
55:37would go through the criteria list and
55:39now it's often an algorithm which will
55:42do the first pass at that yeah yeah so
56:18you're right not necessarily in the same
56:19industry on net for the economy right so
56:23right we could lose jobs in law
56:25potentially but on net maybe we'll get
56:27other jobs out there yeah
57:04yeah so that is that is the argument I
57:07tend to be skeptical of that just
57:09because there have been a lot of wide
57:12changes technological changes in the
57:13past like you know electricity and
57:16various engines and on net they have not
57:19led to and people were worrying about
57:21this 100 years ago with the
57:22mechanization of Agriculture like what
57:24are all these farmers the the US economy
57:27used to be 40% farming now it's
57:29something like 3 or 4 percent so what
57:31are all these unemployed farm workers
57:33going to do and things happen I'm not
57:36saying it won't happen I just think that
57:38that the technological history tells us
57:42to be wary of concluding that jobs will
57:44go away broadly it might happen still
57:47yeah it was just a request to display
57:53the last slide on privacy oh yes ok yes
58:01could you comment a little bit more
58:04about the self-programming aspect of
58:06this for instance in the bail
58:07determination program my understanding
58:12is that as new data is fed into the to
58:15the system it changes the weights that
58:18are assigned to these things is there a
58:20way to track that how do you argue that
58:22on appeal yeah that's a great that's a
58:26great question so this is one of the
58:28high issues right now in fact I'm having
58:31a conference on this very point next
58:33month that Colorado you guys want to
58:35come to the University of Colorado is
58:37what happens when a algorithm makes a
58:41substantive decision about somebody's
58:43rights what do we have the right to know
58:47and how do we think about that given
58:49that those decisions are basically a
58:52bunch of mathematical weights in there
58:54and well learning those mathematical
58:56weights tell us anything and then how
58:59far do we have to trace it back so the
59:02answer is it's complicated and people
59:04are thinking about it but the use of
59:06this is moving forward quicker than the
59:08policy thinkers are thinking about it
59:10for better or for worse
59:12you know following up following up on
59:14the thread about what will the role of
59:17attorneys and legal providers be in the
59:20future in a judicial context why
59:23wouldn't why would we necessarily think
59:26that there'd be you know a model and
59:28algorithm and system that the court uses
59:30to come up with like things from
59:32sentencing guidelines or other parts or
59:35even discovery as opposed to the role of
59:37the lawyer in the future would be as
59:40part of like you know preliminary
59:41motions saying you know advocating I
59:43think we should use this algorithm and
59:45you know trained by this data set and
59:47and then you know the other counsel
59:49would argue for something else and then
59:51there'd be a ruling and so we'd be
59:54really quite in the middle of this and
59:56then also maybe avoid the the likely
1:00:00outcome of you know once technology gets
1:00:02embedded in public sector and especially
1:00:04judiciary it's not prone to adapt and
1:00:07evolve very quickly and so instead of
1:00:09having it be kind of clunky and
1:00:11out-of-date quickly and not evolve with
1:00:12a law maybe it could be another way also
1:00:15to keep it up to date but also always
1:00:16fair to the facts and circumstances of
1:00:19the case yeah that's a great point I
1:00:22think the only issue is that the law has
1:00:24not caught up with this idea yet but I
1:00:27think something like that is something
1:00:30to think about as a way forward I do
1:00:33have a nice abstention question the
1:00:36legal system that we know the human
1:00:38legal system a major component of it is
1:00:41that legal decisions emerge from an
1:00:44adversarial process another thing is
1:00:50that the humans also have a hard time
1:00:54explaining the actual basis of their
1:00:57decision largely because we actually
1:01:00don't even have introspective access to
1:01:03to the actual decision making process
1:01:05instead our explanation process are post
1:01:09facto rationalizations stories and for a
1:01:14neural network making a decision I
1:01:16suspect the the actual decision-making
1:01:20cannot be explained in a way that some
1:01:23other agent could actually under
1:01:25and the explanation but rather one could
1:01:28imagine adversarial post-facto story
1:01:32accounts regarding justification of the
1:01:36outcome of the decision yeah I think
1:01:38that's a really good point and I think
1:01:41one of the points you raised in your
1:01:42question too is a really good point is
1:01:44we don't we sometimes or holding
1:01:48algorithms up to a higher standard of
1:01:50explaining themselves when today we have
1:01:52human decision makers who have to
1:01:55provide an explanation but it might not
1:01:56be the true explanation or it might be a
1:01:59misleading explanation or they might not
1:02:00even know why they made the decision
1:02:03algorithmic decision-making gives us the
1:02:05ability to peer in at a very deep level
1:02:08in ways that we can't peer inside the
1:02:10human brain and get in some sense see
1:02:13the decision but in other sense it might
1:02:15not be interpretable and strategies like
1:02:18that may be a way to deal with some of
1:02:20that yes yeah so thanks for for a nice
1:02:24survey of the history and current state
1:02:28of this area I was just going to make
1:02:30one comment about the risk assessment
1:02:33algorithms there composts stuff so so
1:02:36it's great the risks are significance
1:02:38it's generated this whole literature
1:02:39about sort of studying the varieties of
1:02:46sort of bias and fairness and they sort
1:02:50of multiple incompatible possible models
1:02:53that you might want to optimize but one
1:02:55downside of it is it is it sort of
1:02:58sucked the oxygen out of the discussion
1:03:01of AI in support of administrative
1:03:05processes which which is that that
1:03:15there's a vast number of different
1:03:16governmental activities that can be
1:03:18improved in principle by by automation
1:03:22of some aspects and you know it's like
1:03:25the risk assessment is just barely it's
1:03:28only marginally AI at all it's easy as
1:03:30your statistical models that are based
1:03:32on mostly on sort of criminological
1:03:34theory they have an empirical basis
1:03:38right it's unfortunate I just I guess
1:03:41it's a kind of unfortunate that that's
1:03:43become representative sort of AI in the
1:03:47government when it seems like it's not
1:03:48representative at all other than how
1:03:51things can go wrong it's representative
1:03:53of that yeah no I think you're I think
1:03:54you're really right I mean one danger is
1:03:57that particularly academics can be
1:04:00become really entranced by the latest
1:04:03idea and this you know has a lot of
1:04:06interesting properties to it but they
1:04:09might be ignoring the more beneficial
1:04:12aspects of AI there might actually lead
1:04:15to helping criminal defendants or
1:04:17helping the under lawyers and you want
1:04:20to be wary not to suck the oxygen out of
1:04:23the room to throw the baby out with the
1:04:24bathwater yes from our online audience
1:04:29do you think that the world needs the
1:04:31data producer right for non personal
1:04:34data or some kind of sweet generous
1:04:36related right for companies who generate
1:04:39or labeled ML data you think such a
1:04:43right would hinder or incentive eight
1:04:46innovation wait say it again
1:04:49the whole question or just the first
1:04:52part yeah do you think that the world
1:04:54needs a data producer right for non
1:04:56personal data or some kind of sweet
1:04:58generous related break I tend to not be
1:05:02in favor of more legal rights over
1:05:04information there's already quite a lot
1:05:07of protection in terms of copyright to
1:05:11some extent patent I think that a legal
1:05:13production right would just add another
1:05:15layer of friction in the system so I
1:05:19tend to be skeptical on that basis hi
1:05:24you mentioned abstract thinking and
1:05:27something that was safe for attorneys
1:05:29versus AI or machines I want to know if
1:05:32you can elaborate on that one of the
1:05:33questions was prior was that machines
1:05:37might be able to abstract down to be
1:05:39able to interpret discussion or
1:05:42argumentation better than humans so I
1:05:44found that contradictory it wasn't sure
1:05:46oh wait so say say more about the
1:05:48example you found confusing basically I
1:05:50believe that when you're doing
1:05:52computation
1:05:52no law and you're doing abstraction that
1:05:55a computer could do abstraction much
1:05:57better than a human that's actually one
1:05:58of the things that computers can do
1:06:00better and so I was wondering I was just
1:06:02for you to comment on one of your slides
1:06:04that you were saying that that was one
1:06:06of the things that was safe for
1:06:07attorneys yeah so abstraction is one of
1:06:09those unfortunate words which has 20
1:06:11different meanings and there's the
1:06:14computer version of abstraction and
1:06:16mathematical which I agree with you
1:06:17which is sort of representing things at
1:06:20different levels what I was referring to
1:06:22is abstract ideas like philosophy you
1:06:25know a computer couldn't knowingly the
1:06:29way the state of AI you know have an
1:06:32intelligent original conversation about
1:06:35an abstract idea you know like the
1:06:37freedom of the state or something all
1:06:39right that's the sense I'm abstraction
1:06:46I'm not yeah yeah I'd love to see it
1:06:50yeah thank you
1:06:51yeah so from a different member of our
1:06:55remote audience is legal personality
1:06:58part of the answer to how we should
1:06:59regulate more advanced applications of
1:07:02AI most policymakers currently seem to
1:07:05be going down a regulatory route that
1:07:07focuses on product liability but could
1:07:09granting some AI systems legal
1:07:12personality it can to say the way the
1:07:14law recognizes corporations as legal
1:07:17persons be part of the answer uh so I
1:07:20tend not to think that legal personality
1:07:22is that useful right now given the state
1:07:24of AI because I think at least you know
1:07:27when it comes down to it corporations
1:07:29although they have legal personhood or
1:07:31ultimately people at the end of the day
1:07:34and I think a lot of the complex complex
1:07:38issues of the day have to deal with
1:07:40people at the end of the day and I
1:07:41haven't seen any really good problems
1:07:44that would be solved by granting
1:07:46artificial intelligence legal personhood
1:07:49given that the state of the art today is
1:07:52basically this intelligent results
1:07:54without intelligence so I think a lot of
1:07:57the legal personhood arguments involves
1:07:59conflating what people think
1:08:01AI is versus what a lot of people think
1:08:04AI involves intelligent
1:08:06thinking machines and they have an urge
1:08:08to grant a personhood on that basis but
1:08:12I think if they were to look at the
1:08:13details they would find that it's not a
1:08:16very useful route that's just my
1:08:17inclination at this stage professor have
1:08:21you had an opportunity to think about if
1:08:24the legal system are to be adversarial
1:08:27for example if you were to look at small
1:08:29claims and in countries where you have
1:08:32you know a typical judge would see 500
1:08:35claims a day would you recommend that a
1:08:38I would perhaps be have you had a chance
1:08:41to look at it it'd be more efficient and
1:08:44how would you be with bias if that were
1:08:46the case and how would you sort of
1:08:48separate the data bias feasibly the
1:08:52pattern recognition bias that the system
1:08:56would sort of throw out yeah no I see
1:08:58that as the way things might go these
1:09:02will be framed as decision support
1:09:03systems that's the way they're framed
1:09:05now for judges to deal with their
1:09:07workload but under the hood there will
1:09:10be aspects of artificial intelligence
1:09:12and machine learning and and I don't
1:09:15really have any answer that other than
1:09:16say this is something we just need to be
1:09:17very cognizant of it's all too easy for
1:09:21governments to adopt systems and put
1:09:24them into place without really thinking
1:09:25of the larger implications of that this
1:09:30is another question concerning bias you
1:09:33referenced California judges using legal
1:09:34software to inform their sentencing
1:09:36decisions obviously that you mentioned
1:09:39the data can also reflect some bias so
1:09:41do you have any proposals or any
1:09:43suggestions to deal with that is that
1:09:45something that is comes down to
1:09:47educating judges or is that is there a
1:09:50policy approach that needs to be taken
1:09:52in working with the private companies
1:09:54that are creating the software and using
1:09:56the data what are some suggestions there
1:09:58yeah so there's a lot of thinking going
1:09:59on in this area right now in the policy
1:10:01and legal world I don't know that people
1:10:03have come to any solutions but some of
1:10:06them do involve having sort of neutral
1:10:08third parties maybe academics or others
1:10:11as part of the process of analyzing and
1:10:15looking at systems and more ideas and
1:10:17more transparency because right now it's
1:10:19a pretty murky
1:10:20world of government vendors the kind of
1:10:23people who sell this software often do
1:10:27it because they have they're already in
1:10:30the government system and they have the
1:10:32ability to sell so they're not
1:10:33necessarily the best or people who are
1:10:38situated to create the software but the
1:10:39ones who are inside the system and have
1:10:42the easiest path to selling it so all
1:10:46right Terry well you're pretty much at
1:10:47the end of our time so let's wrap it up
1:10:49here please join me in thanking Harry
1:10:51certain for a wonderful talk Thank You
1:10:52great to have

About this transcript

This page contains the full transcript of Artificial Intelligence and Law – An Overview and History | Guest Speaker: Harry Surden by Stanford Law School, generated from the public captions YouTube serves with the video. The transcript has 10,672 words across 1,588 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.