Latest / CLEs You Actually Want to Hear / The Human Baseline: AI Displacement and the Silent Recalibration of Legal Doctrine
Transcript
- 0:02For being here, I am Mark Hoag, chair of the
- 0:07Beverly Hills Bar Association's AI and the Law
- 0:09section, along with my co -chair, Amanda Harris.
- 0:11And I'm also the founder of Axiomic LLC, an AI
- 0:15first web and iOS app development company. And
- 0:19for my day job, of course, I also do run my own
- 0:21law practice, hoagelaw .ai, where I serve as
- 0:24fractional GC to AI first SaaS companies and
- 0:27tech startups. I am thrilled to introduce our
- 0:30speaker today, Jim Koh. Jim is the founder and
- 0:34executive director of the AI Rights Project,
- 0:36an independent non -profit that educates students,
- 0:39creators, and citizens about their rights in
- 0:42the age of AI, anchored by its flagship Know
- 0:46Your AI Rights ebook. He also serves as principal
- 0:51attorney at Koh IP and AI Law PLLC, bringing
- 0:55more than 20 years of experience across in -house
- 0:57law firm, and non -profit roles to help clients
- 1:01navigate intellectual property and AI related
- 1:03legal issues. Combining civil rights advocacy
- 1:06with practical legal counsel, Jim is dedicated
- 1:09to helping people and organizations understand
- 1:11their rights, protect their ideas, and thrive
- 1:14in the algorithmic age. And with that, let's
- 1:18dive in and get started. Take it away. Thank
- 1:21you, Mark. And good morning, everybody. I'm Jim
- 1:24Coe. I run the Coe IP and AI law, and I direct
- 1:28the AI rights project, as Mark was just saying.
- 1:32In June of last year, Judge Alsup in the Northern
- 1:35District of California held that training a large
- 1:38language model on copyrighted books was fair
- 1:41use. Here's part of his reasoning. The author's
- 1:45complaint, he wrote, is no different than it
- 1:47would be if they complained that training school
- 1:51children to write well would result in an explosion
- 1:54of competing works. Let's dig a little deeper
- 1:58on this. Whether you think it's right or wrong,
- 2:01notice what the sentence does. It resolves the
- 2:04case by treating what the machine does and what
- 2:07the child does as the legal equivalent. There
- 2:13was really no further analysis provided. It arrived
- 2:17as an analogy and it was, apologize, and it was
- 2:33the assumption the entire fair use analysis was
- 2:36built on. That's the pattern we're going to spend
- 2:39the next hour on. American law is built on background
- 2:42assumptions about human cognition agency and
- 2:46authorship that are so deeply embedded they have
- 2:49never needed to be stated. Statutes were enacted
- 2:53against human scale conditions. Sometimes the
- 2:56law says person and the assumption is explicit.
- 3:00More often it says author or speaker or agent,
- 3:04or as here just learning or the equivalent. And
- 3:08that assumption there is silent. Those assumptions
- 3:11are being tested right now by courts. that may
- 3:14not recognize they're making a choice. And the
- 3:17path of least resistance is to keep the doctor's
- 3:19vocabulary while quietly lowering what it demands.
- 3:24I call that silent recalibration. Three predicates,
- 3:29four bodies of law. That's the ground we're looking
- 3:31to cover. And we'll finish in employment discrimination
- 3:35where I think the problem is sharpest and the
- 3:38least noticed. So here's the framework. I call
- 3:43it the human baseline. Three predicates embedded
- 3:47across human law. The first predicate is human
- 3:51creativity or the assumption that behind every
- 3:55work there is an author, a human author who made
- 3:58choices. Copy runs on this predicate. Our second
- 4:04human predicate is observational capacity. The
- 4:08assumption that watching people is expensive
- 4:10and it has costs. So surveillance is naturally
- 4:13self -limiting. The Fourth Amendment in particular
- 4:16is exemplary for the observational capacity predicate.
- 4:21And the last and not least predicate is the attributable
- 4:26agency human predicate. The assumption that behind
- 4:30every decision, there is someone who can be asked
- 4:33why that decision was made and to be held account
- 4:37in account for that. The First Amendment runs
- 4:40on the attributable agency predicate, as does
- 4:43Title VII employment discrimination cases. None
- 4:47of these is stated anywhere, but all of them
- 4:50are load -bearing. On the right is what AI systems
- 4:56now do. They generate text and content. They
- 5:01make decisions. They optimize, and they operate
- 5:04at scale. Line the columns up, and you see the
- 5:08collisions. Generation of content runs at creativity
- 5:12on the left. Scale runs at observational capacity.
- 5:17And deciding and optimizing both run at attributable
- 5:19agency. One word about what I mean by displacement
- 5:24because everything today turns on it. It is often
- 5:28said in popular discussions on this subject that
- 5:32AI is just a tool and just like any other technology
- 5:35tool that assists humans. And it certainly can
- 5:39be that. But one of the premises of this presentation
- 5:43is that AI can go beyond that and it can also
- 5:46functionally displace humans. It can perform
- 5:50functions that the predicate had always assumed
- 5:54a human was performing, because before AI, humans
- 5:59were the only things that could perform them.
- 6:02Tools that serve a human decision maker do not
- 6:05displace. but architectures that substitute for
- 6:09human decisions can. Hold on to that distinction.
- 6:14It's going to do a lot of work for us over the
- 6:16course of this hour. So three predicates, three
- 6:21baselines, and all three of them being pulled
- 6:25towards the same place by AI. Notice what the
- 6:30picture doesn't show. It doesn't show any agent
- 6:34or person deciding to move them, it's more of
- 6:38a gravitational pull in concept. The word doing
- 6:47the work here on this screen is the word silent.
- 6:52The issue here is not that doctrine changes.
- 6:54Legal doctrine changes all of the time. And that's
- 6:58the way the common law was designed. When a court
- 7:02extends a rule deliberately, It says so is one
- 7:07of the common marks of the common law. It gives
- 7:09its reasoning. Those reasons can be briefed against,
- 7:13they can be appealed, and they can be criticized
- 7:16in a law review. Most importantly, they can be
- 7:20overridden should the legislation decide to do
- 7:24so. Silent recalibration leaves none of that
- 7:28behind. The opinion reads as faithful application
- 7:33of settled law. The vocabulary is preserved.
- 7:37The court still says fair use. It still says
- 7:39the word author. It still may say the word job
- 7:42related. But what changes? What those words require.
- 7:47And because nothing announces that change, there
- 7:50is nothing to appeal and nothing for Congress
- 7:53to respond to. One more feature worth naming
- 7:56now. This displacement really runs only in one
- 8:02direction. Accommodation to AI displacement always
- 8:08is one direction. It lowers the bar. It never
- 8:11raises it. Each step is small and defensible,
- 8:14perhaps on its own, but they do accumulate to
- 8:17great effect. AI silently recalibrates the law
- 8:24across all of these domains, but it does it differently
- 8:29in each one. Same three predicates as the last
- 8:33slide, here showing the language that each domain
- 8:36uses for them. There are two mechanisms of silent
- 8:44recalibration that we have identified. The first
- 8:49is what we call the equivalence assumption, treating
- 8:54a machine function as equivalent to the human
- 8:58predicate that it displays. The best example
- 9:01came from the top of this presentation, that
- 9:05machine learning equals human learning, that
- 9:09the LLM trained information and content generated
- 9:14equals the school children who learn and create
- 9:18their own works. We opened with that concept.
- 9:22The second mechanism of silent recalibration
- 9:25we'll come back to at the end once we've seen
- 9:27why it matters. Here's why the first one, the
- 9:31equivalence assumption, is hard to catch. When
- 9:33the equivalent assumption is in operation and
- 9:36it carries the day, the doctrine, the language,
- 9:39actually survives intact. But when it doesn't,
- 9:44the words stay, but the logic underneath them
- 9:47has been quietly displaced. It loses its coherence.
- 9:56Here on the screen is the articulation. of what
- 9:58we call the human baseline principle. That is
- 10:01the basis of the entire framework we'll be discussing
- 10:04here today. It's an interpretive discipline.
- 10:08It is not a new cause of action. It's not novel
- 10:14in structure. There are lots of principles of
- 10:17law that this is very similar to. For instance,
- 10:22clear statement rules, the federalism canon,
- 10:25the major questions doctrine. American interpretive
- 10:29doctrine already resists structural transformation
- 10:32by implication. The shared logic is the same.
- 10:38Where a reading would shift the conditions Congress
- 10:40legislated against, you need clear authorization
- 10:44in the legislative text to get there. Otherwise,
- 10:50we need to defer to congressional intent, including
- 10:53with rewriting any laws as necessary. as it deems
- 10:57necessary. The human baseline principle applies
- 11:01that logic to a new kind of structural pressure
- 11:04that is AI. Notice what it does and what it doesn't
- 11:09do. If you apply this principle, as you see on
- 11:13the screen, it doesn't decide the outcomes. It
- 11:16doesn't decide whether AI employment screening
- 11:19decisions are lawful. It doesn't decide whether
- 11:23training is fair use. What it does is it decides
- 11:27who answers those questions, a court by extending
- 11:32the words of the law by implication or the Congress
- 11:37on the legislative record. Here on screen is
- 11:44the question that this whole hour turns on, not
- 11:47whether AI is good or bad, not whether it's accurate,
- 11:51just... Was a human function amplified or was
- 11:57it replaced? Spellcheck flags a word, but you
- 12:04decide whether to take the correction or not.
- 12:09Westlaw ranks cases, but you're the one who reads
- 12:11them. These are examples of using technology
- 12:15to amplify the human process. A person is still
- 12:19doing the judgment. They're still doing the application.
- 12:22They're just operating at a faster rate because
- 12:24they're helped, they're assisted by the technology.
- 12:29Let's look at the other scenario. A screening
- 12:33tool ranks 400 applicants and a recruiter interviews
- 12:38the top 20. Well, who decided which 20? It wasn't
- 12:44the recruiter in the scenario I'm providing.
- 12:46It is a technological tool. If that technological
- 12:51tool is following deterministically a bunch of
- 12:54rules that were set before it, that is one application
- 12:57of the technology screening. If, however, the
- 13:01tool is AI based as we commonly think of it today,
- 13:05and the AI tool is coming up with its own screening
- 13:10process, not one that was fed to it like a recipe,
- 13:14well, that's a different scenario. In this case,
- 13:19that recruiter at the end, while he or she may
- 13:22be making the final recommendations of who to
- 13:25interview and who to hire, she chose among 20
- 13:28out of 400 applicants that were pre -screened
- 13:32by that AI -driven employment screening tool
- 13:36that I was just describing. The issue here is
- 13:41that this same product, it really can land on
- 13:45either side of the line depending on its configuration,
- 13:48whether it's amplify or transform, whether it
- 13:52still fits within the bounds of human decision
- 13:54making where humans are accountable, and whether
- 13:56it kind of crosses that threshold into an AI
- 13:59driven decision process that is just fundamentally
- 14:02different. And all of this, where it falls on
- 14:06the line in any given case, well, that's for
- 14:08the lawyers here. It's a discovery question.
- 14:11That's something to be analyzed. in whatever
- 14:13cases you are litigating and defending or pressing.
- 14:18The main test question I would offer to help
- 14:23tease out that determination is this amplification
- 14:26or is this transformational? Can you point to
- 14:31the human who is responsible for this function
- 14:35and when an AI -driven process takes the responsibility
- 14:41for some of that process, Can that human supervisor,
- 14:46if you ask them, can they explain how and why
- 14:51that AI decision making process came out the
- 14:54way that it did? If the human can explain that
- 14:57and explain how the parameters that were set
- 15:01for the tool were designed by humans and the
- 15:05execution all fell within human expectations,
- 15:09well, that's kind of one. category, one bucket,
- 15:11so to speak. But if that human supervisor, when
- 15:15put on the stand, he or she may be responsible
- 15:18for the tool and its outputs. But when asked,
- 15:23he says, I have no idea how it works. AI is,
- 15:26it works in mysterious ways. It's inscrutable.
- 15:28It's a black box. Don't you know that? Well,
- 15:32right or wrong, that kind of categorically should
- 15:34be evaluated differently, or at least that's
- 15:37the That's what we would submit here in this
- 15:40presentation. When the law's machinery is working,
- 15:47every doctrine we look at assumes there's someone
- 15:51who is accountable, someone you can ask, someone
- 15:54who knows its operation. The framework. The human
- 16:05baseline principle. We have set up a five -step
- 16:10framework, a disciplined step -by -step that
- 16:15the bench and bar can and we submit should be
- 16:19applying when addressing these AI issues. This
- 16:24framework provides the interpretive discipline
- 16:26for the courts, and we respectfully submit that
- 16:29it should be added to the tradition of the clear
- 16:32statement rule, substantive canons. and other
- 16:35legal interpretive constraints. Everything from
- 16:39here on in this presentation runs through it.
- 16:42It's going to be the same five steps for every
- 16:45application of this framework. We'll walk through
- 16:49four domains here today, copyright, the Fourth
- 16:52Amendment, the First Amendment, and then employment
- 16:56discrimination. Our first application of the
- 16:59day is copyrightability. the copyrightability
- 17:03side of copyright law. Can an AI generated work
- 17:07be copyrighted at all? Applying my framework
- 17:13one step at a time, the first step is to identify
- 17:17that human predicate that is at issue. And here
- 17:21is the predicate of human creativity. The question
- 17:26presented is, is creativity for the purposes
- 17:30of having available the right to protect a work
- 17:36that is created by copyright law, is that prerequisite
- 17:41creativity, must it be an exclusively human activity?
- 17:46Must a human be the one with the brush, so to
- 17:49speak, or the pen? On this issue, the courts
- 17:55recently came back with an affirmative on this
- 17:58in Thaler versus Perlmutter. The DC Circuit ruled
- 18:03that the Copyright Act of 1976 does in fact require
- 18:08all eligible work for copyright to be authored
- 18:12in the first instance by a human being. So on
- 18:16at least this question, the predicate held. And
- 18:19as importantly, the court actually examined this
- 18:21predicate expressly and directly. As we will
- 18:25see shortly though, such discipline and clarity
- 18:29is simply not something we can safely assume
- 18:31will happen on its own in every case. Step two
- 18:37of our framework is to describe the AI pressure,
- 18:40where it's coming from and what it's being applied
- 18:43on. In the case of copyrightability, the pressure
- 18:47is AI, of course, and it's more specifically
- 18:50machine scale generation and replication. It
- 18:56entails or includes the capability Whereas a
- 18:59human artist may be able to put out a couple
- 19:02of paintings a year, as we all know, AI can generate
- 19:06graphic digital works of art at a scale that
- 19:10is millions of times faster than any human could
- 19:14possibly contemplate achieving. The third step
- 19:19in our framework is to identify the risk of silent
- 19:24recalibration. In this copyrighted arena, copyrightability
- 19:29issue, that risk of recalibration can actually
- 19:33go both ways. You can see the courts over -correcting
- 19:37in two different directions. They could define
- 19:39authorship too broadly, thereby extending copyright
- 19:43to output that no human being ever really touched
- 19:48or authored. You can also see the court going
- 19:51too narrowly. and stripping protection from work
- 19:56that a human made with the assistance of AI tools.
- 20:00And to be clear on this, there is without question
- 20:03art being generated using AI and leveraging AI,
- 20:09but still leveraging human creativity in new
- 20:12and amazing ways that happens with every new
- 20:15technology. And there's the risk here that the
- 20:18courts and the copyright office could overcorrect.
- 20:21and not reward and not make possible for these
- 20:25types of artists the copyright protections when
- 20:28they're exhibiting the same level of creativity
- 20:31as has been exhibited across all areas of art
- 20:36and have been recognized and granted such protections.
- 20:42Step four, and this is the step that matters
- 20:45the most for the analysis. is to ask and for
- 20:50the courts to ask themselves, is the court authorized
- 20:54to carry out any recalibration? If that risk
- 20:59exists, that the coherence of the law is going
- 21:03to be altered if we replace machine learning
- 21:07for human learning, the key question that the
- 21:11court should be asking themselves in a form of
- 21:13judicial restraint is, Is the court authorized
- 21:17to make a ruling on this? Or is this a question
- 21:21that really is properly allocated to Congress
- 21:26in your Constitution? There are two main sources
- 21:34to look at for that question. One being, if it
- 21:38is a congressional piece of legislation, a law,
- 21:42that is the key issue here. that legislation
- 21:46may actually specify that the courts are authorized
- 21:49to fill in any gaps in the wording of the law,
- 21:55or it can make it very clear that the court's
- 22:00flexibility to do so is restrained, constrained,
- 22:03and that should be left to Congress. Outside
- 22:08of any statutory controlling laws, of course,
- 22:13we've got the Constitution that can play a role
- 22:16in at least some cases. Does the Constitution
- 22:19independently confer certain inalienable rights
- 22:23within its language itself that the courts are
- 22:26authorized to interpret and apply? And perhaps
- 22:32the court can apply these in ways that cabin
- 22:35limit or even overturn laws, even when they are
- 22:38contrary to the intent of Congress. because that's
- 22:42simply the allocation of powers to the court
- 22:45under the constitution. It should be noted, our
- 22:50paradigm case that we talked about within human
- 22:53creativity of failure, that answered the question
- 22:57that was directly in front of it. And that case
- 23:00was, you know, you'll hear a lot of people say
- 23:02in passing that, well, the courts have already
- 23:04decided this. They've decided that AI created
- 23:08art is not copyrightable. Let's move on to something
- 23:11else. Well, the simple fact is, it's important
- 23:15for people, especially non -IP practitioners,
- 23:18to understand that case was a very uniquely framed
- 23:25case. It was a test case framed by a plaintiff,
- 23:29and it was presented as 100 % This piece of art
- 23:36that we're filing for copyright was generated
- 23:40100 percent by AI. There was no human art involved.
- 23:47There was not even any human direction involved.
- 23:51It was all AI generated. That was the issue presented
- 23:56and the court ruled under those narrow facts,
- 24:00no copyright. It goes without saying, the real
- 24:05world cases are far different. There's gonna
- 24:08be some form of mix where AI contributed some
- 24:13to the process and humans created some to that
- 24:18process. And those questions, it truly is wide
- 24:23open. There is no guidance as of yet or very
- 24:27little meaningful guidance from the courts as
- 24:30to how to determine how much human contribution
- 24:34is necessary to meet the copyrightability bar,
- 24:40so to speak. And a next step that will evolve
- 24:46over time and with the case law is to just kind
- 24:50of for the courts to make a check against various
- 24:53guardrails that developed over time. With the
- 24:58starting one that we will respectfully submit
- 25:00would be that augmentation or transformation
- 25:04question as something just to set guardrails
- 25:08and parameters within which the court should
- 25:11operate. Let's look at the other side of the
- 25:17copyright coin. Let's turn to that. We've spent
- 25:20all our time on the copyrightability side or
- 25:23trying to get a copyright. The flip side is the
- 25:28copyright fair use and AI training side of the
- 25:33coin. And that is when individuals copy an artist's
- 25:39work without their authorization. And in the
- 25:43AI context, they use it without authorization
- 25:45to train their AI models. Does the AI provider
- 25:51in this case violate the original artist's copyright
- 25:55rights? Or do they get out of jail? a free card,
- 26:00so to speak, do they get forgiven, be under the
- 26:05doctrine of fair use, copyright fair use. It's
- 26:11the same statute, just a different section of
- 26:14it. Section 102 of the Copyright Act is for copyrightability.
- 26:19Our discussion here is section 107. A human reader,
- 26:34okay, sorry. This is where the equivalence assumption
- 26:39does its work with respect to the fair use doctrine
- 26:43for AI training. Put the two side by side here
- 26:47and they just don't add up. A human reader, is
- 27:01selective. When you read, you take what interests
- 27:04you, and you leave the rest behind. In contrast,
- 27:08an AI model ingests the entirety of everything
- 27:11that it reads exhaustively and at scale. When
- 27:16a human reads a book, she brings a lifetime of
- 27:20experiences to that book, whether they're six
- 27:23years old or 80 years old. Whatever you take
- 27:26from the book, lands on top of everything that
- 27:29you bring to the table already. When a model
- 27:33ingests works, it trains for the first time,
- 27:37there is no already there. It is a blank slate.
- 27:42Before training, there's nothing there. It's
- 27:46only with training that it has any content to
- 27:50begin with. A reader that is a human picks his
- 27:55or her own purpose. Why? They are reading this
- 27:58book. A model is pointed at an objective that
- 28:01someone else specified, and error corrects towards
- 28:04it through a process of AI optimization. Mind
- 28:09you, none of this comparison and compare and
- 28:14contrast, none of it tells you that fair use
- 28:17should come out differently for machine learning
- 28:20versus human learning. Our framework is intended
- 28:25to not be outcome driven. The purpose is to provide
- 28:30the bench and bar an interpretive framework for
- 28:33these entirely novel issues that are brought
- 28:36about by AI, and to ensure that if a court is
- 28:39going to rest its fair use determinations on
- 28:43the presence that all of these activities, machine
- 28:46or human, are the same, that this premise itself
- 28:50should be litigated and fully examined and not
- 28:54just assume, because the consequences on our
- 28:58law and our society are really massive. Coming
- 29:07back to the sentence that we started with, and
- 29:14now let's look at the one that follows within
- 29:16the same opinion from the Bart case we started
- 29:18with, with respect to AI training. Like any reader
- 29:24aspiring to be a writer, aspiring, turning a
- 29:29hard corner to create something different, all
- 29:33of this language attributes purpose, meaning,
- 29:38drive to an AI machine that doesn't have any
- 29:45of those things, at least not in the same way
- 29:47that human biological organisms have. And I want
- 29:53to be careful here because this is not a criticism
- 29:55of the judge in the Bart's opinion. Anthropomorphism
- 30:00is baked into how we talk about computers and
- 30:03for good reason. It's how the field explains
- 30:05itself. We say a program reads a file that memory
- 30:10remembers, that a computer system decides. Nobody's
- 30:15confused when we use this language. That's why
- 30:18we use it. It's a shorthand. and it makes an
- 30:21unfamiliar foreign thing like computers and how
- 30:25they operate graspable for people like you and
- 30:29me. Machine learning is applying that same principle.
- 30:36It's a useful metaphor, learning. It was useful
- 30:40when it was coined and it's useful now when you're
- 30:43trying to teach someone conceptually what's happening.
- 30:47But the risk that comes out of such a habit is
- 30:52when the metaphor crosses into the legal domain,
- 30:56into having legal consequence. In a textbook,
- 31:00learning is a teaching aid. In a summary judgment
- 31:04opinion, it's a premise, and these premises carry
- 31:07significant consequences. So taking a look at
- 31:12the Barthes versus Anthropic case, did Judge
- 31:16Alsop determine Did he examine and come to a
- 31:21reasoned conclusion that machine learning is
- 31:24the equivalent of human learning? Or did he just
- 31:27basically assume that premise and build his entire
- 31:31transformative fair use analysis on top of this
- 31:35unexamined assumption? Respectfully, I think
- 31:39a fair read of the opinion really does. I don't
- 31:44think you can avoid the latter being a conclusion.
- 31:48Then the next and harder question goes back to
- 31:52step four of our framework. Is a court authorized
- 31:56to make such a decision that effectively changes
- 32:02the scope of the original law as it was drafted?
- 32:06Let's examine that in this copyright fair use
- 32:08context. There is no constitutional right. There's
- 32:18no amendment, first or tenth, or what have you,
- 32:22giving creators ownership rights over their creative
- 32:26works of authorship. If there was a constitutional
- 32:29right of copyright, we'd be having a completely
- 32:33different conversation than we all have. Because
- 32:35in that case, the courts would unequivocally
- 32:38have the right to construe such language. But
- 32:42here, the power of copyright It does come from
- 32:47the Constitution. It's in Article 1. And the
- 32:54particular section in clause begins, the Congress
- 32:59shall have the power. Notice right away, it's
- 33:03not the courts that are given this power. The
- 33:07power to calibrate the copyright system is vested
- 33:09in one branch, and it's not the branch deciding
- 33:11these cases. Congress shall have the power to
- 33:17promote the progress of science and useful arts.
- 33:20Now there's a whole debate as to what this means.
- 33:24There's a robust line of Supreme Court cases
- 33:27saying that the purpose of the copyright is instrumentable
- 33:32for the good of society, not for the individual
- 33:35artist. There's also a Supreme Court opinion
- 33:39recently that, and a line of cases from there.
- 33:43Emphasizing, yes, this is good for society, but
- 33:46at the same time, the copyright law should promote
- 33:49the public interest, but without lessening the
- 33:52incentives of creators to create. The key question
- 33:57is, does this constitutional intellectual property
- 34:00clause give the Supreme Court or the courts the
- 34:03authority to make the determination of the scope
- 34:06of any of these issues and any associated distinctions
- 34:10or not? Taking a look at the specific verbiage,
- 34:18Congress shall have the power to promote copyright
- 34:21by securing for limited times to authors and
- 34:25inventors the exclusive right to their respective
- 34:28writings and discoveries. There is a very significant
- 34:39word in this clause. this work back to a particular
- 34:44someone. Now read it as a whole. Notice that
- 34:49step four is really answer on the face of its
- 34:51text. The power to extend this scheme is granted
- 34:54to Congress. Courts interpret these words, they
- 34:57always have and they always have to, but there's
- 35:00a difference between reviewing a judgment Congress
- 35:03has already made on these issues and making a
- 35:07new one that Congress hasn't. and effectively
- 35:10rewriting the law in a direction different from
- 35:14that expressed by congressional text. Here, Congress
- 35:18hasn't spoke. It simply hasn't spoken on this
- 35:22issue of whether the copyright law or fair use
- 35:25should extend to machine learning as opposed
- 35:29to stay within the constraints of human learning.
- 35:32Our viewpoint and our framework is respectfully
- 35:36that that the court should not be making such
- 35:43determinations. So let's take a look at the three
- 35:48substantive fair use opinions that have issued
- 35:52on the subject matter so far for an early preview
- 35:55of where the law currently sits in the AI age.
- 36:00The Barge case we talked about and its sister
- 36:03case also in the Northern District of California.
- 36:05was Kadri versus Meta platforms. And both are
- 36:10class actions or war class actions over training
- 36:13on copyrighted books. They actually, the Sunray
- 36:17judgment opinions for both of these cases came
- 36:20weeks apart in June or July of last year. And
- 36:23they both came out for fair use. In other words,
- 36:29Anthropic and Meta both had copied books. to
- 36:34train their AI without permission, and the courts
- 36:38held that was totally legitimate, the fair use
- 36:42principles apply, so long as the works were legally
- 36:46obtained in the first place. Bart's in particular
- 36:51drew the line between what the opinion referred
- 36:55to as pirated works, meaning those obtained illegally,
- 37:01no fair use protections apply, But so long as
- 37:05Meta and Anthropic, which famously or infamously,
- 37:09depending on how you looked at it, part of that
- 37:12case was they acquired something on the order
- 37:14of six million hard copy books. They bought them
- 37:18used, they bought them legally, and they ripped
- 37:20them off of their cover, digitized them, fed
- 37:22them into their AI trading process. And because
- 37:27they did all of this lawfully, the court ruled
- 37:30that that was fair use. The third case went the
- 37:35other way. It's Thomson Reuters versus Ross Intelligence
- 37:39out of the East Coast, out of the district of
- 37:41Delaware. And here the company Ross Intelligence
- 37:47took the Westlaw headnotes that we all know all
- 37:52too well as practicing lawyers. And they took
- 37:55Westlaw headnotes without authorization from
- 37:58Thomson Reuters to train their own competing
- 38:01legal research tool. And the court held, hey,
- 38:05this is not fair use. Finally, we have a win
- 38:08for the starving artist. Except, well, here,
- 38:12Les Lawhead notes aren't exactly a bastion of
- 38:15artistic expression. And of course, Thomas Thompson
- 38:18Reuters has a $46 billion market cap or so. As
- 38:23you can see from this map from Ed Lee's Chat
- 38:27GPT blog, This copyright fair use issue is a
- 38:31heavily litigated issue. We've only talked about
- 38:36three of over a hundred cases that are going
- 38:39on. In conclusion, that's going to wrap up our
- 38:44discussion of human creativity and the IP law
- 38:48and human creativity. If you put the two halves
- 38:52of our discussion together, you'll notice which
- 38:54way they both cut. With respect to copyrightability,
- 38:59Output works of authorship are protectable, but
- 39:04the line between AI assisted and AI generated
- 39:08is unsettled. With respect to fair use, creators
- 39:13are being limited in that direction as well.
- 39:18Their creativity is being used without their
- 39:21permission to train AI that at least in part
- 39:25can compete competing works against them. potentially
- 39:28weakening their market, and there's plenty of
- 39:31documented evidence that that is happening to
- 39:33various degrees. So on both sides of the copyrightability
- 39:37coin, creators are losing at both ends. They're
- 39:40less able to protect what they make with their
- 39:42tools, and they're less able to stop those tools
- 39:45from consuming what they've already made. That's
- 39:48what I call the IP paradox of AI. AI both spurs
- 39:54human creativity while undercutting individual
- 39:58creators' ability to protect them as IP. I'm
- 40:03gonna move on to the second human predicate we're
- 40:07gonna talk about, and that is the predicate of
- 40:10observational capacity. We've already been talking
- 40:13about a first application in our discussion of
- 40:16IP. We just didn't realize it at the time. But
- 40:20I'm gonna move on in the interest of time to
- 40:23the kind of paradigmatic application of this
- 40:27concept, which is the Fourth Amendment and our
- 40:31constitutional protections against the government
- 40:33conducting an unreasonable search and seizure
- 40:36requiring the police to secure a warrant from
- 40:40a judge before using any such invasive search
- 40:43and seizure techniques. Now, how on earth do
- 40:46we define what unreasonable means in this or
- 40:50any context in the law? The Fourth Amendment
- 40:52law has often been criticized as wildly subjective
- 40:56and inconsistent. And yet, rather remarkably,
- 41:01it's actually out of this body of law where proponents
- 41:04of preserving human rights in the face of AI
- 41:07should look to first. It is the first domain
- 41:11and easily the one with the most developed case
- 41:14law applying the concept of the human baseline
- 41:17principle and this human observational capacity
- 41:20predicate. Privacy law is another. The courts
- 41:24simply haven't given a name to this concept as
- 41:27of yet. We're here trying to provide the bench
- 41:31and bar an interpretive framework to help analyze
- 41:33these issues and a vocabulary to talk about.
- 41:39The human observational capacity predict kit,
- 41:42it's kind of a real world type of application.
- 41:48It's not a hard fast rule written in the law.
- 41:51before digital records, watching someone for
- 41:55police to track and follow a suspect. In the
- 42:01old days, it really did take a lot of effort.
- 42:04It was costly and it took a detective in a car
- 42:08hours to conduct a surveillance. And because
- 42:14it was labor intensive, such surveillance was
- 42:19inherently gonna be more episodic. One cop or
- 42:24two cops following a car for as long as they
- 42:28can be awake type of thing and self -limiting.
- 42:33The Fourth Amendment has nothing in writing saying
- 42:36that surveillance must consume a great deal of
- 42:40resources and it can cover unlimited resources
- 42:45and extending capabilities beyond etc etc. The
- 42:50law is not written to account for that. And it
- 42:53didn't, as a practical matter, need to. There
- 42:57were just inherent limitations until technology
- 42:59came along, so to speak. But as technology evolved,
- 43:03the court has applied these same principles,
- 43:09at least in principle, at least in general. In
- 43:12Carpenter versus United States, the government
- 43:15obtained four months of cell site location data.
- 43:20on Carpenter's movements from his wireless carrier.
- 43:24And they obtained them by court order. They did
- 43:26not obtain a warrant. They used a different avenue
- 43:30in the law that doesn't require a probable cause
- 43:34finding that Carpenter had committed a crime
- 43:39as would have been required in other contexts.
- 43:44The Supreme Court held that this was a Fourth
- 43:47Amendment violation. And its reasoning turned
- 43:50on the comprehensiveness of the surveillance
- 43:54that was at scale. It was four months of cell
- 43:57site tower communication saying where he was.
- 44:01And the court basically said, look, we don't
- 44:05want to set up the system where nobody can escape
- 44:09the watchful eye of the government in all contexts
- 44:13when they had weren't under suspicion in the
- 44:15first place. However, it is important to know
- 44:19that the Carpenter opinion, while it certainly
- 44:22gestures in these directions, the Supreme Court
- 44:25went out of its way to expressly limit its finding
- 44:29on its facts. It was not defining any sort of
- 44:32applicable general principle. What we are trying
- 44:36to do, and it ties into the human baseline framework,
- 44:40is to help the courts and the pension bar move
- 44:44towards the application of general principles
- 44:46of law. and develop those here. Let's look at
- 44:54this Fourth Amendment in determining the court's
- 44:58authority here. And this is to be distinguished
- 45:01from the copyright law that we looked at before
- 45:04that is primarily, if not exclusively, statutory.
- 45:09The Fourth Amendment talks about the right of
- 45:12the people to be secure in their persons. and
- 45:15to be secure against unreasonable searches in
- 45:19seizures. Particular requirement really does
- 45:27presuppose a sequence. Under the logical structure
- 45:32of how the Fourth Amendment is written, the government
- 45:34needs to first suspect a person committed a given
- 45:38crime, and then the government has to get a warrant
- 45:43if it wants to use any invasive search and seizure
- 45:46procedures. With that warrant, it can proceed.
- 45:50The point here being there has to be a human
- 45:53particularized judgment first. Now let's consider
- 45:57the more modern context of technology surveillance
- 46:01and now AI surveillance. With mobile fortify
- 46:05being a facial recognition application, that
- 46:09immigration customs and enforcement agencies
- 46:11have been using. This algorithm identifies people
- 46:14on the street before any human official has formed
- 46:17any judgment about whether that individual warrants
- 46:20an investigation. This inverts the order that
- 46:25the Fourth Amendment clause assumes. It's not
- 46:28a warrant problem, it's a predicate problem because
- 46:31the Fourth Amendment's machinery runs human suspicion
- 46:33first and here it's an AI system that flagged
- 46:38that suspicion first. The question I'll pose
- 46:41to the audience here today, should such AI assisted
- 46:46suspect identification be considered in compliance
- 46:49of or in violation of the Fourth Amendment? And
- 46:54this applies also to, I'm sure everybody's heard
- 46:58a lot of talk in the news recently of the flock
- 47:01cameras that seemingly popped up everywhere.
- 47:05and are being used to the same purposes? And
- 47:08those same big questions apply here as well.
- 47:12Is technology here, is AI merely amplifying ordinary
- 47:16human capabilities in the same sort of way and
- 47:20being used in the same way as those human search
- 47:23capabilities were used before? Or are these technologies
- 47:27fundamentally transforming them and therefore
- 47:31perhaps should be applied differently? So here,
- 47:42this is just a quick table to show that the observational
- 47:46capacity predicate, it touches on more than one
- 47:52different doctrine of law, and it interfaces
- 47:55with each in a different way. And when the human
- 47:59predicate is displaced in each of these four
- 48:02different areas of law, you get four different
- 48:04failure modes. Now, we're going to turn to our
- 48:10third and last and most expansive predicate,
- 48:14that of attributable agency. We've got two exemplars
- 48:20of that, institutional free speech and AI -driven
- 48:25employment screening. This predicate, simply
- 48:30put, is the notion that the basic idea in law
- 48:34that if a harm occurs, if somebody is injured,
- 48:39the law always provides an avenue to the buck
- 48:42has to stop somewhere and it always stops in
- 48:45a person or a group of people. This fundamental
- 48:51concept in the law is absolutely displaced in
- 48:55part and sometimes in large part by AI and how
- 49:00it's already used and in particular how it will
- 49:03be used in future. With respect to First Amendment
- 49:08free speech, we're going to look at it in the
- 49:10context of social media and how social media
- 49:14curates its personalized web pages for you and
- 49:19me on Facebook or Instagram or TikTok. And there
- 49:25are two powerful shields that social media providers,
- 49:28they benefit from that don't exist in other industries.
- 49:32The first one is the First Amendment. Congress
- 49:35shall make no law abridging the freedom of speech
- 49:37or of the press. And that's a protection against
- 49:40government regulation. The second shield is section
- 49:46230 of the Common Communications Decency Act
- 49:49of 1996, which has been credited as creating
- 49:54the conditions to protect the internet. And I'm
- 49:59realizing our time is. is moving, so I'm going
- 50:02to have to just move on from here with that as
- 50:05the background. Okay, the key for Section 230,
- 50:10though, is it provides protections for Internet
- 50:14service providers, going back to the birth of
- 50:19the Internet in the late 1990s and 2000s, to
- 50:23provide them with protection against lawsuits
- 50:27for serving as passive hosts for third party
- 50:31content that was put on their ISPs. And the basic
- 50:35notion was ISPs should be able to try to make
- 50:40good faith efforts to moderate the content that
- 50:43is posted on their servers, but without bearing
- 50:49the liability that if they failed to properly
- 50:52moderate something that they would be on the
- 50:55hook for such content. then there are practical
- 50:59reasons for providing them with the protections
- 51:02that we're not going to be able to go into right
- 51:04now. Both the First Amendment protection and
- 51:09Section 230 turn on what we call the human predicate
- 51:13of attributable agency. Here the specific application
- 51:17being whether or not there's the presence of
- 51:19human editorial judgment. However, they actually
- 51:25come at this and benefit and they come at it
- 51:29from completely opposite ends. Because the First
- 51:32Amendment, in principle, it protects these platforms
- 51:35because a human made the editorial decision,
- 51:40their expressive choice to publish or not something.
- 51:46With Section 230, the protection is for precisely
- 51:50the opposite reason. It's because no human made
- 51:52an editorial choice that they get such protection
- 51:57for liability. It's awfully hard for the social
- 52:01media companies to claim both at the same time.
- 52:05It's a have your cake and eat it type of circumstance.
- 52:14It helps to have an understanding of how AI systems
- 52:22make their decisions. In fact, it's necessary
- 52:25for any adjudication as to what liability exists
- 52:30and what protections carry over from human judgment
- 52:34to AI judgment. You really do need to understand
- 52:38be able to parse out how the AI decision -making
- 52:41process is happening in this case and how different
- 52:44that is from human decision -making process.
- 52:48As an example, well, let's take a look at the
- 52:51human. newspaper publication game, the business.
- 52:55How has it decided that all of the news that
- 52:59is fit to print is laid out on the page in the
- 53:02way that it is and what articles get placed where,
- 53:04and as importantly, what articles are, don't
- 53:07make it past the cutting room floor. The objective
- 53:12is set by a human being in the newspaper business.
- 53:18And the measure that they all apply, and reasonably
- 53:23so, is they're trying to increase subscriptions.
- 53:28In terms of how do we increase these subscriptions,
- 53:33well, there are certain rules that are developed,
- 53:37kind of objectives that are set up by the editor
- 53:40-in -chief and by the newspaper as a whole. They
- 53:43do market research. And really everything comes
- 53:46around editorial judgment. They take that market
- 53:49research and decide what they are looking for,
- 53:52all towards that goal of increasing subscriptions.
- 53:54And that editor does a very human process of
- 53:58determining and identifying for that newspaper
- 54:00different rules, specifying different rules that
- 54:05they are going to select for when deciding whether
- 54:09or not to publish a given article. And then that
- 54:15rule is implemented again in a human editorial
- 54:18decision -making process on an article by article
- 54:21basis. Do we publish this or not? Do the different
- 54:25rules that we've specified, are they met in a
- 54:28given article? And do they all combine towards
- 54:31our goal of increasing subscriptions? All of
- 54:35this is necessary for a framework and a starting
- 54:40point that we're going to launched from as we
- 54:43look at an AI process of the same. So in AI,
- 54:49we got the same general structure. Some of this
- 54:52looks the same, but I will tell you right now
- 54:54the objective specification part is exactly the
- 54:57same in that a human also has to define its overall
- 55:02objective with the difference being that instead
- 55:05of increasing on a subscription macro level,
- 55:08is much more micro with AI and with social media,
- 55:13where everything's measured based on clicks,
- 55:15hence the notion of clickbait. And notice that
- 55:20the objective here, the things informing the
- 55:25development of rules is no longer editorial discretion.
- 55:29It is combining training data and the... training
- 55:34of the models and the use of them to run inferences.
- 55:39And here is the key distinguishing point that
- 55:41will basically be leading you all with here today.
- 55:45Instead of the rules, we tend to think of AI
- 55:49as replacing human decision making, but we tend
- 55:52to think of it more on the level of, hey, when
- 55:55the boss gives me a set of rules and I make a
- 55:58lot of decisions to apply those rules, well,
- 56:01that's my job as a middle manager. and AI is
- 56:04replacing my job as the middle manager. Here's
- 56:07what I want to leave you all with today. The
- 56:10real defining characteristic of AI is not at
- 56:14the level that I just described. The level that
- 56:16I just described is basically what's been going
- 56:19on for the last 10, 20 years. The difference
- 56:22with AI and the reason it's so transformative
- 56:25and the reason it turns upside down our law is
- 56:29those rules that we were talking about. This
- 56:34criteria on which these decisions are made themselves
- 56:38are no longer made by humans. They emerge out
- 56:42of the AI training and optimization and AI model
- 56:46generation process. This radically transforms
- 56:51the analysis. And furthermore, not only are these
- 56:54rules developed by AI, not humans, they're also
- 56:59inscrutable. We cannot know real time and we
- 57:03can't figure out after the fact exactly how AI
- 57:07came to decide publish this and not that. That
- 57:12information is truly unknowable because AI is
- 57:18doing this massive processing optimization, testing
- 57:22out millions of parameters and the associations
- 57:26between each. and predicting and connecting and
- 57:29making predictions as to whether that's connected
- 57:32to more clicks or not in a fashion that the AI
- 57:36developers themselves cannot explain real time
- 57:41or after the fact. And that has massive implications
- 57:45on our attributable agency concept because how
- 57:50can we hold anybody accountable for the decision
- 57:54making if Not only the rules for those decisions
- 57:59were not made by humans, but we can't understand
- 58:03how they were made. And therefore, how can we
- 58:06hold anybody accountable for it? We can't even
- 58:09tell whether it was discriminatory or not. There
- 58:13are a lot of issues that this gives rise to.
- 58:16So I'm seeing I'm at the end of my time. I didn't
- 58:19time this out as well as I would have liked.
- 58:21I will pause here and ask if there are any questions.
- 58:26And I see one on the screen right now that AI
- 58:30generated deep fakes and revenge porn. Can I
- 58:34address the latest litigation in light of the
- 58:37Taken Down Act and the Defiance Act? I wish I
- 58:40could. I am not knowledgeable of the latest case.
- 58:43I will say that I give all due credit to our
- 58:48Congress and our executive branch for identifying
- 58:52and passing early legislation on this issue and
- 58:55getting ahead of it. It is one that I take near
- 58:58and dear to my heart and think is incredibly
- 59:00important. Having said that, though, it is once
- 59:03again one of those easier cases in the sense
- 59:05that everybody can easily get behind a law prohibiting
- 59:10this type of use of AI. It is those more those
- 59:16other areas that are much more complicated and
- 59:19harder to navigate through that I think we've
- 59:23got our work cut out for us as practitioners.
- 59:27I see another question here of what of strict
- 59:30liability and AI in terms of patent and trademark
- 59:34infringement and insider trading and finance?
- 59:39And I'm gonna ask the person who wrote that question,
- 59:43What do you mean in particular in terms of strict
- 59:46liability? Are you referring, you might very
- 59:50well be referring to the notion of if a harm
- 59:54occurs that somebody should be strictly liable
- 59:59for that harm. And well, I mean, I'm not sure
- 1:00:04if I'm getting to the heart of your question,
- 1:00:06but as much as I am sympathetic to and want to
- 1:00:13preserve our legal tradition of every time there's
- 1:00:16a harm, there's a remedy. The flip side of a
- 1:00:18strict liability regime is notable and important.
- 1:00:23It's hard to imagine if there's a strict liability
- 1:00:26regime across the board, AI developers being
- 1:00:30able to develop AI, which absolutely can and
- 1:00:35we hope will be used to the benefit of society.
- 1:00:39The questions we're posing today are somewhat
- 1:00:41very much analogous to those in the 1990s in
- 1:00:45terms of the birth of the internet and kind of
- 1:00:47similar considerations do apply. And I think
- 1:00:52I do see a follow -up, but my screen actually
- 1:00:54just blanked out on that front. So hopefully
- 1:00:56we'll come back here. Of a human directing AI
- 1:00:59without intent that gives rise to a crime without
- 1:01:03intending it to do so. And that's a fantastic
- 1:01:07question. It ties into that strict liability
- 1:01:11question, right? If a human is using AI, if I
- 1:01:15have clients who basically sell AI services,
- 1:01:20and that question of if that AI service presents
- 1:01:24harm to one of the end customers, should the
- 1:01:29vendor be held liable for that harm? If they,
- 1:01:33in good faith, didn't intend for that harm to
- 1:01:35happen, is one standard. The real guidance that
- 1:01:40I always give my clients is, well, nobody intends
- 1:01:45to harm anybody. The real question is, is it
- 1:01:50foreseeable that it can cause a harm to your
- 1:01:54customers? And if it is foreseeable, have you
- 1:01:57taken reasonable steps to mitigate against that
- 1:02:01harm from happening? In the abstract, that's
- 1:02:06got to be the guidance, and that's going to be
- 1:02:08how it's going to end up. If you took reasonable
- 1:02:11steps at a certain point, the courts will say
- 1:02:15your steps were reasonable enough, and even though
- 1:02:17the harm occurred, you tried to mitigate it,
- 1:02:20just applying basic tort principles, product
- 1:02:23liability principles, but in this AI context.
- 1:02:26And product liability really is the standard
- 1:02:29that we're looking for. It's clear that that's
- 1:02:32being applied. as we speak, most recently in
- 1:02:35the Medicaid status at all. Can I have a question
- 1:02:38of, if the result is discriminatory, then what
- 1:02:43does it matter who created the thought process?
- 1:02:47And that is a fantastic Title VII employment
- 1:02:51discrimination type case, or that those same
- 1:02:55principles can be applied across the board to
- 1:02:58college admissions and so forth and so on. And
- 1:03:02gosh, that is a massively important question.
- 1:03:05I think the questioner is asking, if AI provided
- 1:03:08the discriminatory effect, then that should be
- 1:03:12held liable just as if a human hiring process
- 1:03:16had a discriminatory effect, even if the employer
- 1:03:20didn't have actual intent to be discriminatory.
- 1:03:24And I absolutely agree with the commenter's question
- 1:03:28and their premise. As people, however, are generally
- 1:03:32aware, that question of whether intent is in
- 1:03:36fact not required under anti -discrimination
- 1:03:39law is a very contested one as we recently speak.
- 1:03:47Our federal government, our executive branch
- 1:03:48has taken the strong position that for discrimination
- 1:03:53liability to exist, proof of intent by the employer
- 1:03:59has to exist for disparate impact liability to
- 1:04:03impose. And obviously, that's a robust discussion
- 1:04:09that we cannot possibly give credit to in this
- 1:04:15session today. I see one more question here,
- 1:04:18at least on my screen. There seem to be a couple
- 1:04:20more. About what about the product liability
- 1:04:22theory approached in the recent case by the AGs
- 1:04:25against Facebook? How does my model assess that
- 1:04:29and the settlement announced yesterday? I have
- 1:04:31to confess, I haven't dug into the settlement
- 1:04:34itself. I will say that product liability, that
- 1:04:39basic principle, AI really does have a way of
- 1:04:43slipping through the cracks of all aspects of
- 1:04:45law, but the concepts of product liability universally
- 1:04:51apply. And the concept of that and joint several
- 1:04:55liability within it is going to be essential
- 1:04:59for all of these cases going forward. So I'm
- 1:05:03afraid I can't give you any of my insider knowledge
- 1:05:06or really informed knowledge on the opinion.
- 1:05:08I haven't read it yet. But I will leave you with
- 1:05:10that framing. I see a question that if AI replaces
- 1:05:17specified human rules with emergent rules, that
- 1:05:21even the developer cannot explain? Does accountability
- 1:05:24need to move away from intent and pretext and
- 1:05:27towards enterprise liability for the entity that
- 1:05:30chose to deploy the system? If not, who should
- 1:05:34bear the risk when no human decision maker can
- 1:05:36be identified? I tend to agree with you. I'm
- 1:05:42going to bring that back to the rubric that I
- 1:05:46described before. I do think, ultimately, all
- 1:05:49these cases will be resolved in principle, not
- 1:05:53on the intent factor, because, again, nobody
- 1:05:56intends to harm anybody. The product liability
- 1:06:00concept of foreseeable harm and reasonable steps
- 1:06:04to mitigate the harm. that is going to be where
- 1:06:08the rubber hits the road. That is the absolute
- 1:06:12devil is in the details type of statement. It's
- 1:06:15easy for Jim Cota, makes such highfalutin statements,
- 1:06:20practicing this in practice, applying it in practice,
- 1:06:23and having AI employers and users and vendors
- 1:06:28actually carry out this principle. in light of
- 1:06:33those challenges we talked about in terms of
- 1:06:35insurability, in terms of a whole multitude of
- 1:06:39challenges, is very, very difficult to even conceive
- 1:06:43of achieving. But that is absolutely our task
- 1:06:47as attorneys. I see another question, and it
- 1:06:52is how do we decide whether it is or is not a
- 1:06:59human act or process And which of those should
- 1:07:02be preserved? I think I understand the spirit
- 1:07:04of the question here. And it is a really important
- 1:07:09distinction that I wanted to get more into, but
- 1:07:11I ran short on time. That question of when an
- 1:07:15AI process is used, under what circumstances
- 1:07:21should a determination come out that that AI
- 1:07:25is used to assist the human decision -making
- 1:07:28process? and therefore the human decider is directly
- 1:07:33accountable, A, and B, that that process should
- 1:07:39deserve and earn the protections of First Amendment
- 1:07:44protections as an example, or Section 230 protections
- 1:07:48as another example, that exist for human decision
- 1:07:52-making processes and are lost, or should in
- 1:07:58principle be lost, or potentially lost when AI
- 1:08:02displaces the human decision -making. And I really
- 1:08:06wish we had more time to go through the slides
- 1:08:08I had intended to do. The more you understand
- 1:08:12of how AI decision -making processes are generally
- 1:08:17run, and the more we understand how humans can
- 1:08:22insert themselves and provide the famous term
- 1:08:26or the commonly used term is guardrails, the
- 1:08:29more we can show that humans have set up not
- 1:08:32just guardrails in principle, but actually effective
- 1:08:35guardrails to constrain the decision -making
- 1:08:40process and in some circumstances, give us actual
- 1:08:46visibility of that constrained decision -making
- 1:08:50process that is auditable, that can be verified
- 1:08:53and tested over time. And perhaps the model can
- 1:08:57be retrained to mitigate against those instances
- 1:09:01happening in the future. Like the more that we
- 1:09:05apply these validation, testing, mitigation against
- 1:09:09risk principles that have existed in all of our
- 1:09:12history of law, if we move in that direction,
- 1:09:16then you can conceive of circumstances where,
- 1:09:20hey, you know what? a reasonable vendor has taken
- 1:09:25the necessary steps. And even though the harm
- 1:09:27occurred, even though the Waymo ran over a human
- 1:09:31being, maybe criminal liability should not be
- 1:09:35imposed. Maybe some sort of liability is still
- 1:09:38of course expected, but we can make more nuanced
- 1:09:43decisions as to degrees of fault. The more humans
- 1:09:48are actually and meaningfully involved in the
- 1:09:55AI decision -making process. I think that is
- 1:10:02probably going to wrap up our sessions. I don't
- 1:10:05see any more questions on the screen. For those
- 1:10:08of you who remained to the end here, I really
- 1:10:11appreciate. I appreciate you staying to the end.
- 1:10:17I'm going to ask you all, if anybody is here
- 1:10:20and interested, please do reach out to me. I
- 1:10:24would love to be connected with you, and please
- 1:10:28consider me a resource going forward. Thank you
- 1:10:31all so much. This concludes today's presentation.
- 1:10:35Thank you again, Jim, for an excellent presentation
- 1:10:38today, and thank you so much to our audience
- 1:10:40for joining us. Have a great day, everybody.