Latest / CLEs You Actually Want to Hear / Contract Drafting: Protecting IP in AI Deployment and Development Agreements
Transcript
- 0:03Hello, everyone, and welcome to this presentation
- 0:06of the Beverly Hills Bar Association, Contract
- 0:10Drafting, Protecting IP and AI Deployment and
- 0:14Development. We're very fortunate and delighted
- 0:17to introduce our speaker, Mr. Matthew Cole. Matt
- 0:20Cole represents clients in commercial litigation,
- 0:23intellectual property matters, and with data
- 0:25privacy issues. His experience with intellectual
- 0:28property matters includes claims for the misappropriation
- 0:31of trade secrets. trademark infringement, false
- 0:34advertising, the sale of counterfeit goods, and
- 0:37cases involving a variety of patent issues. In
- 0:40addition, Matt counsels companies that develop
- 0:43and use artificial intelligence on AI governance
- 0:46and strategy, intellectual property, data privacy,
- 0:50and regulatory and compliance issues. He writes
- 0:52regularly and is a frequent speaker and panelist
- 0:55on AI, especially as related to intellectual
- 0:58property. data privacy, and corporate policies
- 1:01on the use and management of AI in the workplace.
- 1:05Relying on his over two decades of experience
- 1:07as a litigator, Matt drafts and negotiates agreements
- 1:11and licenses for businesses with a focus on protecting
- 1:14intellectual property and proprietary technologies.
- 1:18Also, Matt helps companies develop and implement
- 1:21data privacy programs and is IAPP CIPM certified.
- 1:26And now, let's get started. All right. Thank
- 1:32you very much. I'm Michael. I am a partner in
- 1:36Seoul Ewing's Baltimore office. I lead our firm's
- 1:41AI team, and I'm really excited to be giving
- 1:43this presentation today. IP issues in AI use
- 1:47and AI contracting are really glitchy, important
- 1:52issues, and depending on whether you are signing
- 1:57an agreement for use, purely for use of AI within
- 2:00your organization or some kind of development
- 2:03agreement developing your own system or buying
- 2:06a system and building on top of it. The contractual
- 2:10provisions have different nuances to them related
- 2:14to intellectual property issues and so we're
- 2:16going to dive into those throughout this presentation.
- 2:19So if we could go to the next slide and we're
- 2:21likely just give an overview of today's agenda.
- 2:24So we're going to start off with a little background
- 2:28on some of the AI terminology, with a focus on
- 2:31generative AI as opposed to other types of AI,
- 2:35and a brief discussion about how some of the
- 2:37different technological aspects can influence
- 2:40the IP decisions and strategy when negotiating
- 2:44an AI agreement. We're going to look at some
- 2:47of the current law and guidances from the Copyright
- 2:51Office and the PTO on IP protection and AI use
- 2:55and how that will also influence contract negotiation
- 3:00and AI usage. with respect to IP. We're going
- 3:04to look at some governance considerations, because
- 3:07really, I think the best way to frame this out
- 3:09is to think about things in the right order,
- 3:11understanding your use case, get your governance
- 3:13in place, and then your IP strategy and contracting
- 3:15strategy should flow from that. And then we're
- 3:18going to do a deep dive into some of the key
- 3:21terms of many AI software agreements that relate
- 3:25to IP. And then if there's any time, I will answer
- 3:28some of the questions that the attendees put
- 3:30into the Q &A. So why don't we jump right in
- 3:33and go to the next slide, which is just an intro
- 3:37slide. And if you could go to the next slide
- 3:39after that, we'll start talking about AI or JAI
- 3:43and some of the IP issues at play here. So really,
- 3:46artificial intelligence is a computer system,
- 3:50a software system that can mimic human reasoning
- 3:54in some respect, whether it's identifying patterns
- 3:59in large volumes of data, making decisions, making
- 4:03recommendations. So the challenge, though, is
- 4:09there is really no universally accepted definition
- 4:11of AI. State statutes will often define AI differently.
- 4:17And even statutes within one state or regulations
- 4:20within one state may define AI differently, often
- 4:23is defined very broadly, and how an organization
- 4:27defines AI in its agreement can really influence
- 4:31some of the IP terms. because, you know, there
- 4:35are, there's now AI aspects into a variety of
- 4:37different kinds of software and the consequences
- 4:40of broad or blanket provisions relating to AI
- 4:44can have consequences or unintended consequences,
- 4:47for instance, at the end of an agreement when
- 4:49there are data deletion obligations. So we're
- 4:53going to talk about GenAI and everyone's probably
- 4:55familiar with large language models. But the
- 4:58key is the nature in which they work and how
- 5:01that impacts electrical property ownership. And
- 5:04so unlike traditional software, which is deterministic,
- 5:08which means that if you put in the same input,
- 5:11you're bound to get the same output every time,
- 5:14unless there's some kind of error. Large language
- 5:16models, GNI systems are probabilistic. which
- 5:20means that they are based on probabilities and
- 5:22you're not always necessarily going to get the
- 5:24same output every time even in response to the
- 5:28same prompt because the AI system is developed
- 5:32to It doesn't know anything. It doesn't have
- 5:35quote -unquote knowledge or a memory, although
- 5:37it has that in some sense. But an AI system,
- 5:40an LL system, is designed to pick the next word
- 5:43in a sequence based on the probabilities that
- 5:46it's been trained on. So it's probabilistic.
- 5:50And your output is going to vary depending on
- 5:52different things, such as the quality of the
- 5:54training data, the nature of the prompt. And
- 5:58is this system really enterprise system? How
- 6:00has it been trained? Or is it a custom system
- 6:02built for a specific department or use case?
- 6:06And as part of the LMS, I have various architecture
- 6:09that include the algorithm. Algorithms are not
- 6:12a concept exclusive to AI. It's really a set
- 6:15of rules designed to solve a problem. Maybe if
- 6:18condition X exists, then Y happens. Our desk
- 6:22example of an algorithm. The AI model architecture
- 6:25also includes the training data. It has a large
- 6:28corpus of information that the system is trained
- 6:31on to give it its knowledge base, to help it
- 6:33get language fluency and understand what should
- 6:36be that next word in a sequence in response to
- 6:39a user's prompt. And then the last piece that
- 6:41we're going to talk about are the weights, which
- 6:43are the connections, the internal connections
- 6:45within the neural network that drives the strength
- 6:48of the response between the information in the
- 6:51prompt and what output should look like. And
- 6:55so all of these various things, depending on
- 6:57the type of your AI agreement, the algorithms,
- 6:59the training data, the weights, those could be
- 7:01different. Those are types of intellectual property.
- 7:04And depending on whether you're developing themself
- 7:06in conjunction with a developer, your agreement
- 7:09will call out who owns various parts of the AI
- 7:13model, as well as the inputs and the outputs.
- 7:17So let's go to the next slide, please. And because
- 7:22we're going to talk a little bit today about
- 7:24development projects or AI development projects,
- 7:27I'm going to add some more terms here. You probably
- 7:29have seen agentic AI. It's a very common term.
- 7:32Every company wants to adopt or use AI agents.
- 7:36Not many really understand what the difference
- 7:38is, what the import is of an AI agent and how
- 7:40that's different from a typical LLM. But that
- 7:43can also influence intellectual property decisions.
- 7:47And so I think of an AI agent as being different
- 7:50from an LLM in one major respect, right? When
- 7:53you have your conversation with your LLM, you
- 7:55put your prompt in, you hit enter, and you get
- 7:57output, right? That's the end of the AI systems
- 8:00task as opposed to an agent. And I characterize
- 8:03an agent by using the acronym ORAEL. So unlike
- 8:10an LLM, an AI agent is designed to observe. reason
- 8:14on the observations that it makes in this environment,
- 8:18act upon those observations, and then do it again,
- 8:21an execution loop, perform that task again until
- 8:25the agent believes that it has met or satisfied
- 8:28the user's request, performed the task, et cetera.
- 8:31So that's a key distinction between an LLM like
- 8:34Chad GPT or Claude. They're just going to respond
- 8:37to your prompt, give you the information, that's
- 8:39the end. Whereas the agent has the ability to
- 8:41reason, go outside itself, and continue to do
- 8:45the task over and over until it believes it has
- 8:48satisfactorily completed the task. And another
- 8:51key distinction is connectors or MCP. MCP stands
- 8:56for model context protocol. And that's sort of
- 9:00an industry standard way for an LLM to connect
- 9:05externally, access external databases or data
- 9:08outside of its own training data. And that could
- 9:11be company databases, it could be the internet
- 9:14and allows AI systems to use tools and communicate
- 9:18externally, which of course can influence intellectual
- 9:20property if there's an unintended data leak.
- 9:23or if the AI agent grabs some third -party IP,
- 9:26incorporates it into your output. And so we're
- 9:29going to talk about why indemnity is such an
- 9:31important term when using very sophisticated
- 9:33AI systems, especially agentic AI systems. RAG
- 9:37or Retrieval Augmented Generation is another
- 9:40type of AI system that many companies are using
- 9:43or developing internally. If you think of a RAG
- 9:47system, a chatbot is a really good example, a
- 9:50chatbot that is curated or trained on a subset
- 9:54of information. I think of a RAG system as a
- 9:58library or a pre -selected universe of data that
- 10:01you give a pipeline to the LLM. And when you
- 10:04put your prompt into the LLM, the LLM will access
- 10:08the RAG database for the information to influence
- 10:10its answer, to generate the output. And there
- 10:13are IP implications there because RAG databases
- 10:16are specially curated with pre -selected training
- 10:19data for a very specific purpose. And it can
- 10:22be a general purpose, like helping employees
- 10:24answer questions. But it could also be for innovation,
- 10:27competitor intelligence, market assessment, identifying
- 10:31gaps in patents, et cetera. And the last type
- 10:34of AI we'd probably talk about today is a fine
- 10:37-tuned model. And that's a customized model,
- 10:40a custom -trained model, where you take a general
- 10:42foundation LLM. And you retrain it or you fine
- 10:47-tune it into a smaller model that uses a smaller
- 10:51set of pre -selected information. So it's different
- 10:53than a RAC database. A fine -tuned model is typically
- 10:56used for a very specific use case and can have
- 11:00departmental or employee or domain -specific
- 11:03information in it. And fine -tuned models and
- 11:07RAG models have important IP implications when
- 11:12a company is developing its own AI systems in
- 11:16-house or taking a third -party enterprise system
- 11:19and building on top of it to develop its own
- 11:22custom tools. Next slide, please. So now we're
- 11:26going to talk a little bit about the landscape
- 11:30of IP protection and AI in the US. So let's start
- 11:34with some general overview of the copyright and
- 11:37patent boundaries on the next slide, please.
- 11:41So it's pretty clear that output generated exclusively
- 11:47by AI cannot be either patented or copyrighted.
- 11:51Both the U .S. Copyright Office and the PTO,
- 11:54as well as federal courts, have already given
- 11:57us their viewpoint on this. They read the patent
- 12:02act and the copyright act very literally talking
- 12:04about the requirement for an inventor and author,
- 12:08a human being. And so if you're using AI to innovate,
- 12:13or if you're a content creator, you really have
- 12:16to understand the IP implications that third
- 12:20parties have the ability to use the information,
- 12:24to use your marketing materials, to use your
- 12:26innovation that is exclusively generated by AI.
- 12:29And so when you're developing your AI strategy,
- 12:33you really have to think about what type of output
- 12:37we're going to create. How do we... How do we
- 12:41protect it? What's our IP strategy? And how do
- 12:44we use the contract to help us get to yes when
- 12:48the law otherwise does not automatically give
- 12:50us the type of protections we would have if there
- 12:54was a human inventor or a human author rather
- 12:56than a automated or artificial one. And so the
- 13:03goal is to get your counterparty who is going
- 13:05to be probably the AI vendor right, or an AI
- 13:09developer helping you develop your own system
- 13:11to agree to give you ownership of the output,
- 13:14the IP through the contract. All right, next
- 13:19slide please. And so just doing a little bit
- 13:22of a deeper dive into the copyright and patent
- 13:25issues, exclusively generated information by
- 13:29AI cannot be copyrighted, but Information that
- 13:33a human has significant or substantial control
- 13:36over the expressive process can be copyrighted.
- 13:39So if you are a publishing house or you work
- 13:43with authors or you create content, you know,
- 13:47for schools or for trainings and you're using
- 13:51AI and you want to protect that information,
- 13:54you really want to understand or have, you know,
- 13:57what you need to do to get a, how much human
- 14:00involvement is enough, right? And so it's pretty
- 14:02clear that prompt engineering is insufficient.
- 14:05Your copyright office has told us, even if you
- 14:07revise a prompt repeatedly, That's not going
- 14:11to be enough. The AI generated output is still
- 14:14going to be controlled, quote unquote, by the
- 14:16AI. And so the Copyright Office has given three
- 14:19examples of how an individual or a human author
- 14:23can establish or demonstrate significant human
- 14:28control over the expression, the creative expression,
- 14:31by doing certain things. So the selection and
- 14:33arrangement. various aspects of the work, for
- 14:37instance, moving the graphical arrangement here,
- 14:40putting the words here, etc. So there may be
- 14:42some creativity in the selection and arrangement
- 14:45of the elements of a work, significant modification
- 14:49to AI -generated output, and then very expressive
- 14:53inputs. And so the Copyright Office's guidance
- 14:57also applies to derivative works, and the Copyright
- 15:00Office has been very clear that Authors or applicants
- 15:04must disclose their AI use and copyright applications
- 15:07and there have even been instances where the
- 15:10Copyright Office is using its investigative abilities
- 15:14to go and look on social media to see if applicants
- 15:18are actually disclosing AI use and honestly and
- 15:21accurately disclosing their AI use. So let's
- 15:25go to the next slide please and talk a little
- 15:27bit. We're going to talk now about fair use because
- 15:29this is a consideration when you are signing
- 15:32up or using an AI system, signing a license for
- 15:35an AI system or developing it. And there are
- 15:38potential claims for infringement that could
- 15:41be leveled against you if, for instance, a third
- 15:43party's copyrighted information ends up in your
- 15:46output. And so you're probably familiar with
- 15:49the fact that there's dozens of lawsuits that
- 15:53have been brought against LM providers in which
- 15:55the primary claim is copyright infringement.
- 15:59And the copyright infringement is really twofold.
- 16:01It's copyright infringement by the LLM developers
- 16:05using the plaintiff's data for training the models,
- 16:09and then copyright infringement for protecting
- 16:12information ending up in output that's generated
- 16:16by the AI systems that may be used by third parties.
- 16:19And so last year, the court's out in California.
- 16:23in two cases handled or issued rulings on one
- 16:27aspect of the fair use defense, the aspect about
- 16:31whether training, whether training, AM model
- 16:35training is a fair use. And in fact, both courts
- 16:37found that using books to train AM models is
- 16:41transformative. It does qualify. as fair use.
- 16:46The AI isn't using the books to generate existing
- 16:49copies or to distribute exacts of the originals,
- 16:52but using it to break it down and turning it
- 16:54into numbers to identify patterns, to generate
- 16:58new text, to create completely new output. And
- 17:01so the course saw that as transformative, which
- 17:04is one of the most important and heavily weighted
- 17:06fair use elements. By contrast, the court found
- 17:10that maintaining a library of unauthorized information,
- 17:15a shadow library, keeping it, acquiring it without
- 17:18permission, that constitutes infringement. So
- 17:22the court is drawing a very broad, very sharp
- 17:25distinction between the act of training and the
- 17:28copyright implications around that versus the
- 17:31act of collecting and maintaining information
- 17:33without authorization that or ultimately, or
- 17:35may ultimately be used for training, which is
- 17:38a no -no and can constitute infringement, the
- 17:40collection of maintenance. And so as you're thinking
- 17:44through your IP strategy, your AI strategy, right,
- 17:47and you're gonna really focus on indemnity, the
- 17:50key question is, does your output, can your output
- 17:53include another third party's copyrighted information?
- 17:58All right, next slide, please. And it's a similar
- 18:03theme with patentability, although the PTO has
- 18:07changed its approach a little bit. But initially,
- 18:10it's very similar to what the Copyright Office
- 18:12said. It was, you know, an inventor must be human.
- 18:16Innovations, inventions, if you will, generated
- 18:18exclusively by AI are not patentable. The same
- 18:22inventor human requirement under the Patent Act
- 18:25exists as a corollary to the Copyright Act. And
- 18:29in February of 2024, The PTO issued inventorship
- 18:33guidance for AI -assisted inventions. They used
- 18:37the same language that you've historically seen,
- 18:39right, for there to be inventorship. There must
- 18:40be a significant contribution of a human. And
- 18:44in this guidance, the AI assist, the PTO measured
- 18:47a human contribution. The test was... How do
- 18:50you measure the human contributions against the
- 18:52GEN -AI output? How do you as a applicant and
- 18:55inventor really demonstrate to the PTO that your
- 18:59involvement in the inventive process was significant
- 19:02enough that you should get credit as an inventor
- 19:05of that invention? So next slide, please. And
- 19:11so late last year, the PTO issued revised inventorship
- 19:15guidance for AI -assisted inventions, recognizing
- 19:18that AI is going to be used quite a bit and has
- 19:21been actually for innovation, and inventors need
- 19:25a way to protect their inventions or innovations,
- 19:29right? And so... The AI must be viewed, the PTO
- 19:32now views AI as an instrument and legally equivalent
- 19:36to laboratory equipment, specialized software,
- 19:39or a database. It's there to assist the inventor,
- 19:42not replace the inventor. There's no higher conception
- 19:46test for AI -assisted patents, but the human
- 19:50must really have the definite and permanent idea
- 19:52of the claimed invention and used AI to solve
- 19:56that. problem, but doing it in a very specific
- 19:59and granular way. Simply writing a prompt that
- 20:02says, hey, AI, here's a disease. How do I fix
- 20:05this? That's clearly not going to be enough.
- 20:07But if you can demonstrate and document a specific
- 20:11hypothesis. the ground rules and the parameters
- 20:14you gave the AI system, whether it's uploading
- 20:17documents, proprietary information, other patents,
- 20:22research papers, product manuals, whatever it
- 20:25is, and then lay out in as much granular detail
- 20:27as you can the problem. what you want the AI
- 20:30system to do and demonstrate that you control,
- 20:32you mean a substantial contribution to innovation
- 20:34throughout the entire way, that is going to increase
- 20:37the likelihood of the PTO granting you a patent
- 20:41for using AI to assist with the innovative process,
- 20:44not to overtake or be solely responsible for
- 20:48the inventive process. And it's the same story
- 20:50with obviously creative process and obtaining
- 20:53copyrights. So next slide, please. So where does
- 20:58that leave us? That leaves us with trade secrets
- 21:00and AI because for trade secrets are very flexible.
- 21:03You don't have that human involvement requirement.
- 21:08There is no inventorship requirement. Trade secrets
- 21:11can be any information, right? The DTSA, the
- 21:14Defend Trade Secrets Act, and I believe all the
- 21:16state corollaries under the UTSA, the Uniform
- 21:18Trade Secrets Act uses the very broad term information
- 21:22that can qualify as a trade secret so long as
- 21:24it meets the specific criteria. Um, but you know,
- 21:28thinking about AI and trade secrets, right? You
- 21:30have, if you are assigning an agreement or deploying
- 21:33AI, you're going to want to use most likely AI
- 21:36for innovation, but you're also going to be most
- 21:39likely putting proprietary sensitive business
- 21:42information, marketing information, customer
- 21:45information, pricing information, et cetera,
- 21:47possibly even formulas into the AI system. So
- 21:50you want to make sure that it's secure and that
- 21:52your agreements clearly spell out who owns what.
- 21:55in the chain of submitting the prompt and then
- 21:59generating the output, but as well as using the
- 22:02output and any insights derived from that output,
- 22:06and then thinking about it as well. If you are
- 22:08developing an AI system, customizing an AI system,
- 22:12you're going to want to, you know... You're going
- 22:14to want to consider the AI system that you've
- 22:16developed that's been customized for you, from
- 22:18which you've submitted your proprietary information,
- 22:22that you've helped train, consider that you're
- 22:26Tracy for your proprietary information, you are
- 22:28IP. And we're going to talk about, too, through
- 22:31how to protect against, through the contract,
- 22:34AI leading to the loss of Tracy to protection,
- 22:36either through some unintentional data leakage,
- 22:40through data training by a third party, or even
- 22:43through a third party retaining your information
- 22:46indefinitely that gives a defendant, for instance,
- 22:48an argument that the plaintiff did not take reasonable
- 22:51measures to protect their trade secrets. So next
- 22:56slide, please. And of course, with everything
- 22:59else in AI and IP, there is a counter tension
- 23:03even with training data. And California is leading
- 23:05the way on this tension here. If you want to
- 23:09keep as much as we can secret. versus the trend
- 23:13towards AI transparency and disclosure. So just
- 23:16quickly, earlier this year, California statute
- 23:20AB 2013 went into effect. It's the training data
- 23:23disclosure statute. It applies retroactively
- 23:28to any models released on or after January 1,
- 23:312022. So that was designed to capture all the
- 23:34Frontier LLMs because CHAT CPT was commercially
- 23:37released November 30th of 2022. It applies not
- 23:41just to Well, it applies to developers, but broadly
- 23:44defines developers to be more than just the developers
- 23:49of the large foundation models, the Clauds, the
- 23:51Chad CPTs, the Gemini. So it's more than just
- 23:54Google or Anthropic or OpenAI, right? This law
- 23:57covers companies that fine -tune AI models or
- 24:02retrains AI systems for their internal uses.
- 24:05And the obligation is the publication of a high
- 24:09-level summary of sources owners, purposes, IP
- 24:13rights. And I believe there's about 12 categories
- 24:15of data that the AI system developer has to publish.
- 24:21So as you're thinking through your IP strategy
- 24:24and your contract strategy, you're going to have
- 24:27to think, OK, we want to develop our own AI system.
- 24:29But on the other hand, what are our disclosure
- 24:32obligations and how we maintain this information
- 24:34as a trade secret? New York has a very similar
- 24:37statute that just went into effect. Other states,
- 24:39I believe, have proposed legislation. Maryland,
- 24:42Ryan has a bill that's similar to California
- 24:44AB 2013. So really the takeaway from these first
- 24:48few slides is that signing an agreement for an
- 24:52AI system or to develop an AI system. the IP
- 24:56implications are significant. So you really have
- 24:58to think about this, you know, from the very
- 25:00big theme section of the process, not when you
- 25:02are negotiating the deal, but you really want
- 25:05to identify the use cases, understand the IP
- 25:07implications through ownership, through reliability,
- 25:10risk exposure, and the potential for loss of
- 25:13IP or third party use of what you would consider
- 25:15your IP because it is very nuanced when using
- 25:19an AI system. And so let's talk a little bit
- 25:21now, if you would, next slide, about how to develop
- 25:24that front -end AI strategy and the IP implications.
- 25:29And so let's go to the next slide, which really
- 25:31breaks down at a high level the divide between
- 25:36some of the IP considerations for an agreement
- 25:39to deploy AI. right, which is simply signing
- 25:43a license and enterprise agreement to use, you
- 25:46know, Claude at your company versus an AI development
- 25:48agreement, which can be something, you know,
- 25:51developing a reg system, fine tuning a model,
- 25:54maybe making a small language model or doing
- 25:58something, taking some existing system and customizing
- 26:00it for your use. So here are, here's some of
- 26:03the key considerations. The IP ownership and
- 26:07focus for the deployment agreement is now becoming
- 26:10pretty much standard. There are very few standard
- 26:13terms in AI software agreements, but ownership
- 26:17of inputs and outputs by the customer is one
- 26:20of them, especially in AI deployment agreements.
- 26:23Although from time to time, I do see agreements
- 26:25in which the AI vendor does try to retain ownership
- 26:29of the AI output, and it does become a significant
- 26:32negotiating point. in contrast with AI deployment
- 26:37agreements where you also have to consider the
- 26:40allocation of the pre -existing IP, the customer's
- 26:43IP that you're using for fine -tuning or model
- 26:46development versus ownership of the new IP, the
- 26:49foreground IP, the output, what's going to go
- 26:52in it, what sources it's going to pull from,
- 26:54who's going to claim ownership and who has licenses
- 26:57over all the information throughout the entire
- 27:00flow of data. Another consideration, right, not
- 27:04necessarily legal, but a pseudo business legal
- 27:06consideration is the customization complexity.
- 27:09So if it's just an AI deployment agreement, you're
- 27:11not customizing it per se, you're not fine tuning
- 27:14it or having it sync up to a ragmon necessarily.
- 27:17You may develop some custom workflows, but that's
- 27:20a little bit different. You're really just using
- 27:22an off the shelf AI model as learning. So your
- 27:27patterns, you're going to use it for your specific
- 27:29tasks, but it's not a true customization, like
- 27:32a development project, like fine tuning or a
- 27:35rag database, which you're going to have other
- 27:37IP ownership considerations and other considerations
- 27:40at the termination of the agreement. I want to
- 27:43talk about that, how to really step this up so
- 27:45that way when the relationship ends, you don't
- 27:48have any questions about who owns what. Who gets
- 27:51to keep my IP? How can it be used? And what third
- 27:54parties will have access to it? And how long
- 27:57will it be retained? These are all the key questions
- 28:00to consider as you're thinking through an AI
- 28:02strategy and the IP questions. And as I mentioned,
- 28:05termination, what's your off -ramping priority?
- 28:08If it's a deployment agreement, you just want
- 28:09to make sure that the vendor does not keep your
- 28:12customer data, or keeps very little of it, or
- 28:14just what they need for specific purposes to
- 28:17comply with their data privacy obligations. or
- 28:21for fraud or abuse monitoring, as opposed to
- 28:25a development agreement where a custom model
- 28:28has been developed or a database has developed.
- 28:31And it's most likely going to have customer data
- 28:34in it. And you're going to want to either claim
- 28:36ownership of it. Or if you can't get that, get
- 28:38some verified destruction of it so no third party
- 28:42gets access to your information. All right, so
- 28:45next slide, please. So here's some of the things
- 28:49to think about when you are formulating your
- 28:51AI strategy, some of the key IP questions. Obviously,
- 28:54you want to think about the IP issues and risks
- 28:57when developing the use case for the AI system.
- 29:00And so those are infringement, ownership, and
- 29:03use of IP, as well as validity, which we talked
- 29:07about. If you are using AI to create innovations
- 29:11that are potentially patentable, or you would
- 29:14like to protect them through copyright protection.
- 29:17Because AI systems use a variety of data, different
- 29:22for a development agreement where you may be
- 29:24licensing in third party data, you may be using
- 29:27publicly available data, you may be using your
- 29:29data or some combination of all three, you really
- 29:33want to understand the ownership and lineage
- 29:35of the data sets before using it to generate
- 29:38output or developing a custom system. You don't
- 29:40want to run into any licensing issues. You don't
- 29:43want to walk into any claims for infringement.
- 29:47So you have to really understand how the model
- 29:51is trained. If you're developing it with a partner,
- 29:53what rights they have, permissions they have,
- 29:57and just understanding the data flow from design
- 30:00and development through actual use and commercialization.
- 30:03If you are using a system externally, if you're
- 30:06under yourself developing an AI system. As I
- 30:09mentioned, ownership rights, usage rights, and
- 30:11deletion are key questions, key issues to consider
- 30:15as part of the AI system lifecycle, but you really
- 30:18want to hash those out in advance and go into
- 30:20the negotiation with your counterparty. uh you
- 30:23know with a plan but also with some backup fallback
- 30:26you know positions because vendors are very protective
- 30:30of data and AI negotiating AI software agreements
- 30:34continues to be a AI vendor data grab and we're
- 30:37going to talk about that in a few minutes the
- 30:40various categories of data and how complicated
- 30:42it is getting when it comes to protecting IP
- 30:44and AI agreements and just another consideration
- 30:47right you can you can permissibly Reversed engineer
- 30:51or competitors trade secrets using AI to do that
- 30:54is a great use case. And so if you know what
- 30:57are the considerations there when when when doing
- 31:01that, right? You don't want to make sure you're
- 31:03not walking into any trade secret claims for
- 31:05improper means. and thinking about really how
- 31:09AI can benefit you if you are going to consider
- 31:12it a trade secret, use it for trade secret purposes,
- 31:15use it for innovation, and maybe not necessarily
- 31:18for patenting inventions, using AI to reverse
- 31:21engineer competitors trade secrets is an increasingly
- 31:24popular use case among companies that have significant
- 31:28R &D budgets and departments. And so thinking
- 31:32about these questions, these are some of the
- 31:34key questions that will influence your negotiation
- 31:37strategy as you get into the weeds of the back
- 31:41and forth with your counterparty on the various
- 31:43provisions, but especially the intellectual property
- 31:45provisions. So let's jump ahead now to the next
- 31:49slide. And so we're going to look at some of
- 31:51the key contractual sections and provisions.
- 31:55And from my perspective, It really starts with
- 31:58the definitions and AI agreement. Traditional
- 32:03SAS and software agreements have a lot of definitions,
- 32:06but AI agreements take this to a whole new level,
- 32:09especially with the various data categories at
- 32:13play. So we're going to start with the three
- 32:15most basic ones for inputs. That's the information
- 32:18that the user puts into the prompt. typically
- 32:22in text form, though LLMs presently allow for
- 32:29attachments to be uploaded, if that's what I
- 32:32can mean, like your proprietary documents, your
- 32:35R &D meeting notes, photos of your factory, anything
- 32:41else that you have obtained, and it can even
- 32:45be publicly available information organized or
- 32:47compiled in a unique way. And the AI systems
- 32:51are not just able to see the prompt and the attachment,
- 32:56but modern LLMs can also review the metadata
- 32:59within documents. And so that's an important
- 33:02consideration if, for instance, you have track
- 33:05changes, comments in a Word file, other metadata
- 33:09associated with attachments that you are uploading
- 33:13to an LLM. You really want to have a broad definition
- 33:16of inputs to include the text based prompt, the
- 33:21attachment, and if you can, the metadata found
- 33:25in the attachments. Because that's your information.
- 33:29The input's going to be owned by the customer.
- 33:31So you want to make sure that all the information
- 33:33that goes into the LM remains your proprietary
- 33:37information. And you'll notice in the definition
- 33:40of inputs I have submitted to or accessed by
- 33:43the services. access, right, includes the new
- 33:47reality of AI systems communicating with API
- 33:50feeds or data pipelines through a RAG database
- 33:53or MCP connectors, right, it can access other
- 33:57databases. And so it's more now than just submitting
- 34:00a prompt. Inputs have to be comprehensive and
- 34:04include all the information accessed by the AI
- 34:07system to generate the output. And so here's
- 34:11a good definition. of output, right? It really
- 34:14is just the responses generated by the services,
- 34:17which is the AI platform based on the inputs
- 34:19and customer data, right? So if you are, for
- 34:23instance, developing an AI system, if you are
- 34:27providing a RAG database, if you're giving it
- 34:29access to some other information beyond the input,
- 34:33Your definition of customer data should capture
- 34:36that, and the output should cover both the inputs
- 34:39and any other proprietary information the AI
- 34:42system used to generate a response that the customer
- 34:46supplied. And in addition to that, it's not just
- 34:50the output itself. but it's the summaries, the
- 34:53synthesized takeaways, and all the insights derived
- 34:55from the outputs, right? That you want to own
- 34:58as your proprietary information, as your trade
- 35:01secrets as well. So don't just limit it. When
- 35:04you're defining output in the agreement to I
- 35:06hit X, hit enter, I put an X and I get Y. No,
- 35:09it's broader than that. It's all the insights
- 35:11derived from the outputs that you want to claim
- 35:14as well. And there's a tension there because
- 35:16vendors want to own as much data as possible
- 35:19and specifically various categories of derived
- 35:22data or data they consider the byproduct of AI
- 35:26use, things like performance data, usage data,
- 35:29some types of metadata. So really having specific,
- 35:34carefully drafted definitions in your agreement,
- 35:38your AI software agreement, is key to protect
- 35:41the IP that's pre -existing and any IP that is
- 35:44generated through your use of the AI system.
- 35:48And then customer data, which is a term you often
- 35:52see in all manner of AI software services agreements.
- 35:56So it could be all data content, blah, blah,
- 35:59blah, blah, blah, supplied by or on behalf of
- 36:01the customer to the AI vendor and generated through
- 36:04the customer's use of the services. So we're
- 36:07capturing output, anything we've given the vendor
- 36:11to create. the AI system for our purposes or
- 36:15to generate or when output is generated using
- 36:17our data. And it's very common now for vendors,
- 36:22software vendors, to add AI features to existing
- 36:26software systems. And if they do that, you really
- 36:30should. And then sometimes they'll provide you
- 36:32with an AI addendum. I think if you're a customer,
- 36:35it's a best practice to have an AI addendum at
- 36:38the ready for when this happens and be proactive
- 36:40and go to your vendor. and say, well, now we
- 36:43need to negotiate on these AI terms because,
- 36:46for the reasons we've talked about, the probabilistic
- 36:48nature, the different kinds of output, you really
- 36:50want to drill down on IP ownership and an underlying
- 36:54traditional SaaS agreement really isn't built
- 36:57to deal with innovation when using AI. So you
- 37:01should supplement your existing SaaS or services
- 37:04agreement or master services agreement. with
- 37:06an AI addendum that addresses these issues. And
- 37:09of course, it's important to make sure that customer
- 37:11data as defined in the AI addendum is consistent
- 37:15with how customer data is defined in the underlying
- 37:18software agreements. Next slide, please. So here's
- 37:25the IP ownership framework that really has taken
- 37:27place or is starting to take shape through negotiations.
- 37:32Customers retain all rights to input. That's
- 37:34a very non... controversial proposition these
- 37:37days and AI negotiation land, although there
- 37:40is very little that's not controversial, that's
- 37:42not contested. In a deployment arrangement, the
- 37:46vendor will retain ownership of the model of
- 37:49the services. but not the results of its use,
- 37:52not the output. It has become standard for customers
- 37:56to own the output, recognizing that the agreement
- 37:58is going to say it may not be copyrightable,
- 38:01is probably not patentable. Therefore, your contract
- 38:05should not just say, should really lay out in
- 38:08clear terms that the customer, to the extent
- 38:10permitted by applicable law, You're seeing that
- 38:12phrase a lot added by the vendor to protect themselves.
- 38:15And it's completely understandable for the reasons
- 38:17we discussed earlier. But the customer owns all
- 38:20rights in inputs, outputs. So you get not just
- 38:24the results, but you also get the derived insights,
- 38:28the key takeaways. And I think it's the best
- 38:31practice to not just declare ownership, but have
- 38:34the vendor provide an explicit assignment of
- 38:36all rights in the output to the customer. That
- 38:38also is not controversial. And if you're getting
- 38:41pushback from your vendor, all you have to do
- 38:44is point them to OpenAI's enterprise terms of
- 38:47service or Anthropics enterprise terms of service.
- 38:50And all the large LM providers, they all, in
- 38:53their standard terms, assign any rights that
- 38:56they may have in the output to the customer.
- 38:58So there's your framework, right? You own the
- 39:00input as a customer, all the data supplied, you
- 39:03own the output. And if it's a deployment agreement,
- 39:06the vendor retains ownership of the model. We
- 39:08are going to talk about how that differs in a
- 39:11development arrangement. So next slide, please.
- 39:15So one thing you hear quite a bit about is the
- 39:18risk of a third party, an AI vendor or some other
- 39:22third party, a sub processor training on your
- 39:25data. So really what that means is that these
- 39:27AI systems are designed to continuously learn,
- 39:31be updated based on the information that people
- 39:33submit in prompts. And for enterprise agreements,
- 39:38it has become relatively standard, I think, for
- 39:40vendors to promise to agree not to train. on
- 39:44customer data, not to train on inputs or not
- 39:48to train on outputs. But there's always a nuance
- 39:52there, and there's always a little bit of a pitfall,
- 39:55which is while the vendor on the one hand promises
- 39:58not to train on your data, it will often include
- 40:02a provision that will at the same time allow
- 40:04it to use your data to improve the services.
- 40:07So when you see that, you of course are thinking
- 40:10that this is inconsistent. And so you have to
- 40:12go back to the vendor and say, really, what does
- 40:14this mean? Because improve is vague. Training
- 40:17is a form of improving, because it makes a system
- 40:20better. It's got more data, more current data.
- 40:24Data is better suited to give the user the response,
- 40:27the output it is after. Because you really have
- 40:30to drill down with the vendor what improve means.
- 40:32Strike this term and really specify how the vendor
- 40:37can use your data, if at all. with regard to
- 40:40the services. So you can do that by allowing
- 40:43the vendor, for instance, to use certain categories
- 40:46of data for fraud monitoring, abuse monitoring,
- 40:51to get to comply with certain privacy obligations.
- 40:54You can agree to all that, but specificity is
- 40:56the key here in improving the services. It's
- 40:59really just a back door to letting the vendor
- 41:02train on your information. And so when you're
- 41:04defining training, You know, many agreements
- 41:07do not have a definition of training. I think
- 41:09it's a good practice to include it. And you can
- 41:12hear different types of training. You can even
- 41:13use this as a pretty robust definition. You know,
- 41:17it captures the training of the model at various
- 41:21stages, initial development, fine tuning, continued
- 41:25training, et cetera. And it has a pretty good
- 41:29litany of what might be considered trainings
- 41:31or development, improve, right? You really want
- 41:34to take the ability to improve the services on
- 41:37proprietary information out of the vendor's hands.
- 41:41And so here's a trap or here's the kind of provision
- 41:43that you might see in agreement that I was just
- 41:46talking about where a vendor is going to say,
- 41:47well, look, we're going to use your data to improve
- 41:49our services, but it's going to be aggregated.
- 41:52It's going to be anonymized. Those terms really
- 41:55apply. to personal information, to protection
- 41:57of people's privacy rights. But when you see
- 42:00that, you got to think to yourself, how does
- 42:02this help me? How does this protect my IP? And
- 42:05I'm not so sure that anonymizing the customer
- 42:08data is going to give me comfort that the vendor
- 42:11is going to use it to improve their services,
- 42:13not just from a risk exposure reason, but AI
- 42:18systems have the ability. They're very good at
- 42:21understanding patterns, and they can re -identify
- 42:26individuals when using data sets by looking at
- 42:28user behavior. It's called the mosaic effect.
- 42:32So there's always a risk when allowing vendors
- 42:35to use your data. Even if they tell you, don't
- 42:38worry about it. It's anonymized. It complies
- 42:40with the privacy statutes. There are still potentially
- 42:43implications out there about what the AI system
- 42:46can recreate, not just from a privacy perspective,
- 42:49but from an IP perspective. It may see, for instance,
- 42:51that you're interested in a customer market in
- 42:54this geography, at this price point, or this
- 42:56type of customer. And it can regenerate that
- 42:59kind of information, even though it's been anonymized.
- 43:01So watch out for these traps where vendors really
- 43:04want to grab your data. So here's an example
- 43:07of a good provision, in my opinion. It's very
- 43:10broad. It prohibits the vendor from training
- 43:13on your data. It prohibits vendors third parties
- 43:17or sub -processors from training on your data.
- 43:21So a lot of humans will say, vendor agrees not
- 43:24to allow a third party to train on your data.
- 43:26You really have to be careful and look at it.
- 43:28Is the vendor going to be training on my data?
- 43:30Are the vendor's contract partners going to be
- 43:33training on my data? And so there is a lot of
- 43:36tension in the marketplace. Because many AI providers
- 43:40are middlemen, right? They're not selling you
- 43:42the anthropic, the Claude or the chat GPT, but
- 43:45they're building their system on top of it. So
- 43:48when you're negotiating with the smaller AI providers
- 43:51rather than the larger developers, they may be
- 43:54contractually constrained as to what they can
- 43:56agree with you as to how they. can and cannot
- 43:59use your data. They may have certain obligations
- 44:02that flow upstream to their contract partner,
- 44:05right, who is most likely the AI developer that
- 44:07could tie their hands in the terms that they
- 44:09reach with you. So that's just a consideration.
- 44:12Although I have found that negotiating AI agreements
- 44:16with some of the smaller AI providers, you have
- 44:19better luck in. in reaching, you know, getting
- 44:23to yes and getting concessions from the smaller
- 44:25providers, whereas the large LLM providers more
- 44:28and more throughout 2026 has been a sort of take
- 44:31it or leave it kind of a story. They're generally
- 44:33unwilling and not very flexible on many terms.
- 44:36Unless you have a large account, compelling case,
- 44:40right, a specific risk and compliance obligation.
- 44:45So that's sort of, you know, what's playing out.
- 44:47But, you know, thinking about not just the middleman,
- 44:50But the people at the top of the food chain,
- 44:54so to speak, and the contractual restrictions
- 44:57may flow down to what you're able to accomplish
- 45:00through negotiation with an AI vendor. And so
- 45:04next slide, please. As I mentioned before, AI
- 45:08can learn from patterns in data, not just the
- 45:11information explicitly in the content. So AI
- 45:14vendors want as much data as possible because
- 45:16they find value in it to improve their services,
- 45:18to create new models. And so AI agreements have
- 45:21multiple data categories that you typically don't
- 45:24see or may not see in standard agreements. So
- 45:27there's AI usage of data, there's AI performance
- 45:29data. You may see terms like that in your standard
- 45:32SaaS agreements that are defined, metadata. And
- 45:36so when you're negotiating the agreement, you
- 45:39really want to make sure that customer information
- 45:42doesn't appear, your proprietary information
- 45:45isn't going to show up in usage data, performance
- 45:48data, or metadata that's the output, the byproduct
- 45:51of use. Because if it is, you don't want to give
- 45:54that away to the vendor to claim ownership of.
- 45:56You want to minimize however they can use it.
- 45:59You certainly don't want them training on it.
- 46:03So you really have to understand the data flow.
- 46:08If the usage data and performance data doesn't
- 46:10present an IP risk, where it's a very low risk,
- 46:14then that's a different story. But again, I will
- 46:17emphasize, thinking about your IP strategy and
- 46:21understanding the technology and the use case
- 46:23is going to really influence how you negotiate
- 46:25these agreements, how you define certain terms,
- 46:29and what rights you're willing to give your AI
- 46:31vendor. Next slide, please. And so let's look
- 46:37now for a few slides at the development contract
- 46:41and how that adds a level of complexity and requires
- 46:46some additional definitions. And so, right, so
- 46:49you're customizing or creating a new model for
- 46:53you in conjunction with the vendor. And so customer
- 46:57training the model, you want to define that term
- 47:00in the AI agreement. and distinguish it from
- 47:04the LLM, the services that you're going to be
- 47:07using. It may be built on top of it. It may work
- 47:11in conjunction with it. What's going to typically
- 47:15happen is your custom trained model or your database
- 47:17is going to rely on the LLM for the language
- 47:20fluency to generate the output and your contractors
- 47:23to really distinguish between that model and
- 47:26any customer trained model from a data provision
- 47:30standpoint, what data are you providing, but
- 47:32also ownership of that customer trained model.
- 47:35And it can take three paths. Customer ownership,
- 47:37customer owns a customized model, the data set
- 47:40and the weights, right? If you're developing
- 47:44a RAG database, you're developing some kind of
- 47:46fine -tuned model, that's really what you're
- 47:48shooting for. Ownership of the model, I've seen
- 47:52it happen, vendors are willing to do that. It
- 47:55just depends on your vendor. So it really is
- 48:00a case -by -case assessment. I mean, I said earlier
- 48:04that it's a take it or leave it approach with
- 48:06a lot of AI vendors. By contrast, there are so
- 48:09many companies out there now offering AI products.
- 48:13If you don't have a good feeling or if you're
- 48:14not getting any traction with your counterparty
- 48:17that you're negotiating with on provisions that
- 48:19make sense, that advance your organization's
- 48:22business case, find somebody else. The second
- 48:27path is that the developer owns the modifications
- 48:30to the model but gives the customer an exclusive
- 48:32license to use the custom trained model. That
- 48:36presents a little bit more of a risk to the customer.
- 48:40And then of course there's joint ownership of
- 48:42the custom trained model, which I do not recommend
- 48:44because you want to have a clear understanding
- 48:47of who owns the data, how it can be commercialized
- 48:49and what happens to it after the relationship
- 48:52ends and joint ownership. can muddy those waters
- 48:58quite a bit and give a third party control of
- 49:01your IP that you've never intended to when you
- 49:06are considering AI use within your organization.
- 49:10So now let's go to the next slide. We're going
- 49:12to look at some additional terms to include in
- 49:16your AI contract for when you're developing an
- 49:20AI system. AI systems are typically a multi -tenant
- 49:24architecture. All the enterprise versions that
- 49:27you're using with the LMs, it's not just your
- 49:29company that's in that data environment. There's
- 49:32hundreds, if not thousands, of other users using
- 49:35it at the same time. And so when you're developing
- 49:38an AI system for your organization, you want
- 49:42to mandate a siloed architecture that has your
- 49:46data, your customer data, your inputs, and any
- 49:49outputs that are derived exclusively from customer
- 49:52data, right? Siloed, segregated, so that third
- 49:55parties can't access them. They can't see them.
- 50:00They can't... the output remains yours, and it's
- 50:04only accessible to your team, your employees,
- 50:07or your contract partners, or your consultants,
- 50:10if you're hiring third parties for some kind
- 50:13of R &D project. And of course, you want to secure
- 50:16licenses for the use of third party or developer
- 50:18-owned data sets that might be commingled during
- 50:21output generation. So one example is if you're
- 50:24developing a RAG database, you may include in
- 50:30it your own proprietary information. You may
- 50:33include publicly available information that you
- 50:35pull off the internet, maybe patents, maybe statutes,
- 50:39maybe regulations. And then you may also be licensing
- 50:43third -party data for use in your RAC database.
- 50:46And it may be commingled when output is generated.
- 50:49So you really want to lay out and have your contract
- 50:51speak to who owns all the different categories
- 50:54of data. And if it is third -party data, you're
- 50:59going to license it yourself or the vendor or
- 51:01the developer is going to license it for you
- 51:03and have rights to allow you to use it. You're
- 51:05really going to want to get some reps and warranties
- 51:08in the agreement that the vendor has those rights.
- 51:15Deployment agreements warn customers that their
- 51:17output may not be unique or may contain information
- 51:20from the internet. That's a little bit different.
- 51:22from a development agreement where you can technologically
- 51:26actually set it up so that the output only has
- 51:29your information or information that you curated
- 51:32or have some right to as opposed to some unknown
- 51:34third party which may may influence or may negatively
- 51:38impact your ability to copyright that information
- 51:40or you know or whatever patent or protect it
- 51:43and you know basically in majority of disclaimers
- 51:46that I see in deployment agreements, the AI vendor
- 51:50tells you this, that your output may not be unique,
- 51:55may contain third party information, may resemble
- 51:57stuff on the internet, and it's basically telling
- 52:00you like, hey, you are not going to be able to
- 52:02copyright this stuff most likely. So there are
- 52:06definitely some keys to a development agreement
- 52:10because you're really going to... own the process,
- 52:12own the inputs, own the outputs, own the model.
- 52:16It's just a much more labor intensive and resource
- 52:19intensive endeavor. And so let's go to the next
- 52:25slide, please. And so here are some other terms
- 52:30that relate to the IP, implications of the agreement,
- 52:35your reps. your warranties and your indemnity.
- 52:39And I will say this, AI vendors are disclaiming
- 52:41everything. They're not standing behind a whole
- 52:44heck of a lot. They're not standing behind necessarily
- 52:47the quality of their output. So you really have
- 52:50to fight hard to get the protections you need
- 52:53to limit your risk and to protect your IP. And
- 52:58so you want to get some reps and warranties,
- 53:00right? The first two are pretty important, you
- 53:02know, get a warranty from the vendor that the
- 53:03model was not trained on sensitive or unauthorized
- 53:06data set, data sets. So that could be information
- 53:10that has third party copyrighted information
- 53:12in it, third party IP, you know, of course, data
- 53:16privacy issues. That's a, that's a different,
- 53:18a different webinar. So you really want to make
- 53:20sure that the vendor is telling you that they
- 53:22have the rights and the licenses to get to use
- 53:25the information in the model. and for you to
- 53:28use the services to give you some comfort that
- 53:30you're not necessarily going to be sued for some
- 53:32kind of infringement or IP violation, misappropriation
- 53:36of trade secrets, et cetera. In conjunction with
- 53:40that, it is pretty common to get a IP infringement
- 53:43indemnity, mostly for the training data. Output's
- 53:48a little harder. However, There are a number
- 53:52of carve -outs that really do limit the value
- 53:55of the indemnity provision. Of course, if the
- 53:59customer intentionally infringes a third party's
- 54:02output, a third party IP, the vendor is going
- 54:05to exclude you from the indemnity provision.
- 54:08But similar to what we see in traditional SaaS
- 54:10agreements, there are carve -outs or exceptions
- 54:13to indemnity. If the customer modifies a service,
- 54:17combines it with third party software, third
- 54:19party API feeds, And that's sort of the point
- 54:22of an AI system is the modifiability of it, developing
- 54:29custom workflows, integrating other data into
- 54:34it, combining it with other systems. And so by
- 54:37doing that, reading the agreement, you are probably
- 54:39losing your IP indemnity protection. Again, the
- 54:43point is, as you're thinking through your IP
- 54:46slash AI strategy, you really have to understand
- 54:50what are the consequences down the road if we
- 54:52are combining our system with a third party software
- 54:55platform, et cetera. Is that going to cause us
- 54:57to lose the limited protections we have? It seems
- 55:01to be that indemnity is taking a shared responsibility
- 55:05or a shared risk model. Were the vendors responsible
- 55:08for the AM model maturing data from a liability
- 55:11perspective? And the customer is responsible
- 55:13for the output. Vendors will tell you that they're
- 55:17displaying all the output. They take no responsibility,
- 55:20no reliability for it. And in fact, many AI agreements
- 55:24require the customer to have humans review the
- 55:29output before using it. And if you don't, if
- 55:32there's a finding in a court case, for instance,
- 55:34that you did not comply with your human in the
- 55:35loop obligations, you're going to lose your protection
- 55:40and you're going to be, you know, there's a risk
- 55:42of infringement claim going to judgment against
- 55:45you. The next slide, please. And so the last
- 55:50two slides are going to focus on what to do,
- 55:53how to set up a good termination or a good walk
- 55:55away from your contract partner. Now keep in
- 55:58mind, right, the goal of an AI system is integration
- 56:01through workflows across systems. I would not
- 56:04take that relationship lightly. You want to have
- 56:06a long -term relationship with your vendor because
- 56:09it is a real challenge to switch vendors and
- 56:12retrain your people and do all the things you
- 56:14did with data once again. But you also want to
- 56:16make sure that you are protecting your information
- 56:20on the agreement if and when it does end. So
- 56:24the typical customer expectation is that all
- 56:26my data will be deleted or returned by the vendor.
- 56:30If you allowed the vendor to train on your data,
- 56:34that's not so simple. AI models cannot simply
- 56:37unlearn data. It's a big, expensive, and technologically
- 56:40very difficult process and something a vendor
- 56:43is never going to agree to to pull some information
- 56:46out of an AI systems training set. That's why
- 56:49it's really key to have no data training clauses
- 56:52that are very broad and protect you. when you
- 56:56are doing an AI development arrangement, you
- 56:58want to make sure you've defined the custom aspects
- 57:02adequately in the agreement. What is the customer
- 57:04training data, the customer training model, so
- 57:06that way the obligations to destroy, delete,
- 57:11or continue use of these custom systems are clearly
- 57:14spelled out when the contract ends. You may want
- 57:18to push for deletion of the model with a certificate
- 57:21of instruction by the vendor. I think that's
- 57:24not controversial. And of course, you want to
- 57:28negotiate for automatic exit rights if the model
- 57:32that you're using is found to be infringing or
- 57:35if it fails an audit. And one of the reasons
- 57:37it may fail an audit is violating IP. It can
- 57:43be very challenging, of course, to audit an AI
- 57:46system. I mean, be careful what you wish for.
- 57:49500 megs of data from an AI system. I probably
- 57:52wouldn't know what to do with it. I would need
- 57:53to get an outside expert. But you still want
- 57:55to push for the protections if something does
- 57:58go wrong. So that way, if there is going to be
- 58:01the vendors found to be infringing a plaintiff's
- 58:04copyright, that risk does flow down to you. So
- 58:07you want to get out of that arrangement as quickly
- 58:09as possible or alternatively see if the vendor
- 58:13can supply aspects of the model that are not
- 58:15infringing. That is a case by case thesis. And
- 58:19then we'll go to the next slide, which is sort
- 58:22of some things to think about for information
- 58:24destruction at termination. I don't know how
- 58:27many more times I can say this, but it's really
- 58:29key that vendors want as much of your data for
- 58:32as long as they can keep it for as many different
- 58:34uses as they can. They want to retain embedded
- 58:37data in their systems, and that can lead to risks
- 58:41with trace secret protection. Sometimes vendors
- 58:45will impose an obligation on the customer to
- 58:48send a written deletion request as opposed to
- 58:51automatic deletion at the end of the agreement.
- 58:54So be wary of those and you want to have clear
- 58:57destruction protocols that destruction of servers,
- 59:01erasure of information. So it's very clear as
- 59:04to what the vendor must do when the agreement
- 59:06ends. And of course, verifiable compliance. Now
- 59:11let's go to the next slide. And we have a minute
- 59:16for one question. They are really focused on
- 59:23the. Really focused on the IP implications. And
- 59:31I will say this, we're talking about some of
- 59:32the copyright cases that came in and questions
- 59:35about human involvement and demonstration of
- 59:39substantial human contributions. And like with
- 59:42anything else in AI governance, AI use cases,
- 59:46documentation is key. You really want to show
- 59:49control of the creative or inventive process.
- 59:54from start to finish, especially if you're using
- 59:56a custom AI model to do that. You really want
- 59:59to demonstrate the problem that you identify
- 1:00:02and how you manage, control the AI system to
- 1:00:06get to the solution. And so next slide. And so
- 1:00:12that's my presentation. It's been an hour. I
- 1:00:14want to thank you very much. And if you have
- 1:00:17any questions after today, feel free to reach
- 1:00:19out. This concludes today's presentation. Thank
- 1:00:23you so much, Matt Cole, for an excellent presentation
- 1:00:26today and thank you so much to our audience for
- 1:00:28joining us. Have a great day, everybody.