Getting copyright right for Generative AI

[WHOOSH] GORDON HARRIS: OK, let's get started. It has now going to 2:00 PM here in the United Kingdom, so good day to you wherever in the world you may be watching or listening to this webinar. Today we aim to take a very practical look at copyright implications of generative AI across a range of jurisdictions. My name is Gordon Harris. And I will be your moderator today.

Since the launch of OpenAI's ChatGPT chatbot in November 2022, generative AI has become an unavoidable topic at panels addressing both policy and practical aspects of copyright law. So while AI technology and AI-based solutions have been developing for many years, impressive recent developments in generative AI are at once opening up new possibilities and, of course, raising deep concerns amongst creators.

Courts, legislators, and other public bodies have started to consult, examine, and in some instances take action. Consensus is building on the application of copyright principles on certain issues. The implications and related risks have also become clearer. So it's time to move on from theory and take a really good look at the practical steps.

So what does that mean to you? And as a matter of practice, what should you be doing about it? So to help us through our review of generative AI, we have a panel of experts from around the world, all technically experienced in the field of AI and its implications for IP and the wider business world. So let me introduce them.

First up we have a guest speaker from the USA. Franklin Graves is an experienced in-house counsel currently serving as a member of the technology law group at HCA Healthcare Inc. He's also an affiliated faculty member with Emerson College's Business of Creative Enterprises MA program where he teaches business and IP law.

Stephane Caron is a partner in our Ottawa office, with extensive experience in advising clients on copyright policy and related regulatory issues. He is co-chair of the International Copyright Laws Committee and the American Bar Association and co-chair of the Copyright Society Ottawa Chapter.

Matt Harvey is head of artificial intelligence for the UK in this firm, and based in our London office. He is the Joint General Editor of the book The Law of Artificial Intelligence by Sweden Maxwell, and was made a fellow of the Royal Society of Arts in 2022 in recognition of his leadership in the field of AI.

Elliot Papageorgiou is head of IP strategy in China based in our Shanghai office, and counseling international clients on contentious and non-contentious IP matters and cross border IP enforcement in China and across the APAC region. Elliot supports a number of embassies in China with current IP developments, and is regularly consulted by the European delegation and European intellectual property office in China.

And last but by no means least Samantha Yung is a Senior Associate in our Dubai office. Her practice covers a wide range of IP matters, including brand protection, advising on online and offline anti-counterfeiting, as well as on issues surrounding domain names, copyright, and IP enforcement.

Now just before we get going, it's quite difficult to take questions as we go along in this format. So could I ask you, please, to post any questions you have in the chat box so we can either get to them at the end if we have time or we will answer them in a follow up communication after the webinar. Time to get started. And to set the scene, I want to hand over to Matt Harvey. Matt, what is the big issue here?

MATT HARVEY: Thanks, Gordon. So, well, let's start with technology. So for most of the time I've been advising on AI, the advances have all been around machine learning. And typically, that's where the computer learns for itself how to replicate a human task or some other task from training examples.

Now until 12 months ago, the excitement was all about classifiers. So to give you the simplest example, you train an AI on images of cats and images of dogs, ask which is which, and from that it learns how to distinguish the two. So you show it a new picture it's never seen before of a cat or a dog, and it can classify it for you. It can tell you if a subject matter is a cat or if it's a dog.

Now, generative AI is the opposite. So to continue that example, it's also been trained on pictures of cats and dogs, but now you give it the class I want to see a new cat or I want to see a new dog. And it can generate a new image that didn't exist before fitting that class. And that class is also often in the form of a prompt. It's a text input that tells you what you want in your output.

Now we're seeing generative AI able to produce text, images, computer code, music videos and more, but the really interesting thing is it's now good enough to be commercially valuable. And to illustrate that, these are five images generated by Midjourney using the same prompt, but using five different iterations of a technology. The amazing thing is this progress has taken place in a single year, so we remain to see what will happen next.

Now let's turn from the technology to the copyright issues. The big issue here is a generative AI models are trained on vast amounts of copyright materials, especially text and images, these are often scraped from the internet. And to give you a sense of a scale, cutting edge image generators are being trained on around five billion images.

So the sorts of topics we're advising clients on are-- is scraping processing and training the model, scraping that data and training the model on it. Is that an infringement? Is a trained model a copy of a training examples used in that training? Are the outputs of those models infringements of the training materials? And also, are there rights in those outputs that you can enforce?

GORDON HARRIS: Thanks very much, Matt. So generative AI has the ability to create content in the nature of essays, as we know, pictures, software, even music, with only a single prompt. This new content is in most ways indistinguishable from works created by humans. So is it protected by copyright, and in what circumstances? Now, this is a question which may have a different answer depending on where you are in the world. So let's start in Canada with Stephane.

STEPHANE CARON: Thank you, Gordon. Perhaps before tackling the state of the law in Canada, I'd like to digress to ask more broadly why we should care whether AI generated content is protected by copyright or not. This matter Matt observed the generative AI's ability to create content worthy of commercial use is really the game changer here.

And we're turning to tools like DeepL to translate text with only post-facto human intervention or to Midjourney, DALL -E, Drip Studio and others to generate images. So the question is, does your business model rely on content being protected by copyright?

Let's start with developers of generative tools. Are they claiming any copyright over the output? So generally conditions of use of the tools are often framed in terms of ownership of the content without specifically referencing copyright. ChatGPT terms of use assign rights title and interest in the output to the user. Relinquishing rights to the AI generated content appears consistent with the notion that AI chatbots are tools, not creators. And that the content that is generated should be available to be used by their human operators. You can draw a parallel.

Hemingway's favorite typewriter might have been essential to the creation of his novels, but Corona and Royal Quiet never laid claim to authorship. Of course, the service providers, the developers are free to place contractual conditions on the use of the services, and particularly where the services are being used for commercial applications.

So, obviously, terms and conditions of generative AI tools used by businesses need to be reviewed carefully. And maybe that's a first takeaway. You consider those terms, and you may want to look at what service providers are being used by employees to generate AI works.

However, as we look at this issue through the lens of copyright, it then becomes-- the question then becomes what exactly is being assigned by way of property. The real value of copyright as we know is the protection of owner's ability to monetize content by assigning and licensing it. And, of course, by being able to police the rights, vis a vis their competitors, whether they be-- and their pirates-- the pirates and counterfeiters. So these parties, they have no contractual relationships with. And that's where copyright proves its value.

GORDON HARRIS: OK, so, well, thanks for that background Stephane. Now can we move on to the actual law in Canada--

STEPHANE CARON: Yes, of course. Sorry, I lost my way there. Canada is on-- is one of a handful of countries where wholly created AI pictures have been successfully registered as copyright protected works. Anybody who's familiar with our system, though, knows that there is no substantive examination of AI copyright application. So the fact that some works may have been successfully registered does not really get us very much from a Canadian perspective.

Let's look instead at some of the conditions of the existence of copyright in Canada as it will echo requirements from other Berne treaty countries. Candidates jurisprudence holds that copyright protects original expression. To be original, a work must be the result of the exercise of author's skill and judgment in the expression of their ideas.

The key conclusion then is that this exercise must involve some human expression, and cannot be merely mechanical. So the general consensus under current Canadian law is that the author of an original-- of original works cannot be a machine. Does this mean that AI generated content is never protected? Well, let's see here-- it's helpful here to consider the distinction between AI generated and AI assisted content.

Is the content entirely generated by the chat box tool based on a prompt, or has the human author contributed skill and judgment in the expression-- to the expression? So the nature and the extent of human participation in the creative process remains, therefore, essential to being able to claim copyright protection at the end of the day in Canada.

Now, the takeaway really is that we have to consider and keep track of what that human contribution is. The US Copyright Office and the US courts have discussed this point, and so I suspect Franklin will be able to provide more information on that subject

GORDON HARRIS: Thanks very much indeed, Stephane. So, indeed, now over to Franklin to give us a view on this point from the USA. Franklin.

FRANKLIN GRAVES: Yeah, thanks Gordon. I'm excited to be here today and a part of this incredible panel. So I'll start by quickly diving into three key decisions that were made by the US Copyright Office. And these decisions illustrate their continuing decision to deny registration, and, therefore, protection of AI generated content.

So as we explore these three cases, though, or decisions, keep in mind what Stephan just said about AI assisted versus AI generated content, because it is important to understand what is actually happening during the creative or the generating process.

So, first, one of the more infamous works is the graphic novel Zarya of the Dawn by Kris Kashtanova. That's the one there with a very Zendaya looking main figure on the cover. And it gained-- so Kris actually obtained a registration for the comic book from the Copyright Office here in the US. But following some publicity around it being one of the first partially generative AI works that were successfully registered in the fall of 2022, the Copyright Office was like, wait, hold on. What's going on here?

And so the images in the graphic novel, come to find out were generated using Midjourney and then edited in Photoshop. So the Copyright Office partially-- ended up partially canceling the registration due to this lack of human authorship, which is what we're always-- which is going to be the underlying theme for everything here in the US.

So the only part of the registration that remained covered and therefore protected by copyright is the text of the story which was original to Kris. They came up with the story by themselves. And then also the arrangement of the images and the text as a compilation, which is what you would do for a graphic novel.

And so the second key decision involved-- I think everyone around the world is probably familiar at this point with Dr. Stephen Taylor and the work titled A Recent Entrance to Paradise, and his other attempts on the patent side as well. But his attempts at registration here in the copyright realm were unsuccessful in the US from the start because his application indicated that the software program creativity machine was the author of the digital painting.

So the denial of the registration was upheld after a series of challenges at the Copyright Office review bot process, and then it went outside the scope of the Copyright Office and went into federal court systems here in the US. And, ultimately, there was a recent decision from US District Court Judge Howell that stopped short of exploring the full analysis I think everyone wants regarding this human authorship requirement because the review in that case was limited to what the US Copyright Office's actions were, which were administrative action to refuse the registration.

Again, going back to that original application and the original assertions by Dr. Taylor that there was no human involvement in the generative work trying to be registered. So it's a rather limiting decision at this point from a federal court case here in the US, so it doesn't offer much beyond what the office-- the Copyright Office has already identified.

And then, luckily, earlier this month-- you're going to hear a lot throughout this presentation. Just this month a new thing happened, because that seems to be the case right now just across the board around the globe. But, again, earlier this month in the US, the office upheld a refusal to register the work space opera theater. And it's a digital painting most well-known for winning an award at an art show in the US. And so there's a picture of it right there in the middle.

And I think if we go to the next slide, you'll be able to see the output from Midjourney, and then some fine tuning and some additional work that was done to end with the output, the ending result there, the work that was actually registered and that actually ended up winning the art show. So it was generated in Mid-journey after a series of over 600 prompts. So that's the user inputting and fine tuning how they are engaging with the chat bot to generate the output image.

And this was also despite it-- the registration was denied despite it being edited in Photoshop and run through another AI tool for upscaling. So, again, the Copyright Office here is looking at the lack of-- what they deem to be a lack of control over software like Midjourney that is generative AI-- that's a generative AI system that really outputs random images to a degree based on the prompt.

And so the issue here, again, for this particular artist and creator was that they refused to disclaim the generative outputs of Midjourney and the upscaling done with Gigapixel AI as part of the protected work. So when exploring registration and protectability of generative outputs in the US, the most current guidance is available at copyright.gov/ai. And presently the arguably-- and I think a lot of people agree-- unclear on exactly how much human authorship is required for copyright protection.

So the only rule we have from the Copyright Office is this de minimis standard of analysis. So what that means in short is you should ask yourself or your clients or your people you work with should ask themselves is the generative material more than a de minimis amount?

And if the answer is yes, then it must be disclaimed on the registration application. And so as of right now in the US if there is generative output that has no human authorship or has-- it's the general material meets that de minimis standard amount, then it would be considered public domain and ineligible for any type of copyright protection.

So these issues remain a high level of risk when it comes to licensing or assigning the rights to generative works or synthetic materials or whatever you're dealing with. So whether it's partially generated or entirely generated using an AI system really is hard to determine at this point where the protection falls. And so the best recommendation, I think, is to make sure you aren't warranting originality or transferring or assigning copyrights that don't exist. And a phrase I've actually seen in the wild as part of a license is to grant all rights, quote, "if any," so be on the lookout for that one.

GORDON HARRIS: I like that. I think that's a nice solution, if any rights, great. OK, thanks very much, Frank. And now we'll shuffle around the rest of the world. And let's start in Europe with Matt.

MATT HARVEY: Thanks, Gordon. So, yes, generally the European approach is very similar to the US and Canada. It's, again, a Berne convention region. So the test of whether there's human authorship is going to be the same, except that there's a harmonized phrase we now use in Europe, which is have you got the author's own intellectual creation? Now where the author has exercised that, where they've used AI merely as a tool to express that intellectual creation, then there'll be the human author.

But generally speaking, we expect that a prompt describing the subject matter of a picture isn't going to be enough because it's not expressive. It's the idea is you're expressing there, not how it should appear. But unusually in two European countries, the UK and Ireland, we've gone beyond that.

And in circumstances where there is no human author, so where an AI has made something essentially autonomously or it has made the expression in the cases we've just seen on screen, there we provide for computer generated works. A special class of copyright that does allow you to protect and enforce those rights.

The complexity there is that by default, the copyright will be owned by the person making the arrangements. So the practical tip there is to have a contract in place so you know who's actually going to own this, and you don't have some debate about who made the arrangements. There is another legal wrinkle, and that's this harmonized European test of your own author's-- the author's own intellectual creation. That came in after the computer generated works provisions, and no one really knows if that could be met by an AI and if that provision no longer really works.

So even in the UK and Ireland, which specifically protects such works, we're not really sure you get any copyright at all. And even then I'd say if you're an international company, having protection in one or two countries around the world may not really be useful.

So we're coming back to Stephane's point which is, does your business model allow you to use these computer generated works valuably? And I would say for internal use or maybe ephemeral social media posts and the like, great, you probably don't need copyright. But you do need to have an intelligent discussion to see if it is the right tool for the task.

GORDON HARRIS: Thanks very much Matt. This is a truly global topic, and to prove that we're going to move over to China now and hear from Elliot Papageorgiou. Elliot.

ELLIOT PAPAGEORGIOU: Well, unsurprisingly Chinese statutory law hasn't really kept up with jurisdictions such as the UK and the United States on this. Copyright law in China is quite complex because it's also administered by the ministry of culture. Having said that, as it currently stands, the definition of copyright work under China's copyright law requires it to be created by a natural person.

So in other words, it per se excludes anything that doesn't have any human input. Or what happens in practice is that any human input is emphasized and the relevant AI is considered to be a mere nuisance to that human author. Of course, you still have the requirement about originality and expressed in a material form, et cetera, et cetera, but this is as far as differentiating between potential authorship by an AI and whether that would exist in China.

At the moment the answer would be that Chinese law would not recognize copyright, even though China, of course, is a member of the Berne Convention as well. Now, as far as Chinese courts are concerned, there really only been two cases. And they've both been of two of the three big guns you would expect to be active in this area. The three big guns being Ali Group, being Tencent, and Baidu. And, in fact, in one of the leading cases from January 2021, which was a Tencent case, it was held that, again, confirmed that the copyright could only vest in a work that is generated by a human being.

Now in that particular case, there was sufficient human input that it really wasn't part of the ratio of the case in deciding that copyright did persist or did exist in that particular case. In an earlier case-- and that was, by the way, the famous Dream Writer case where the question was whether that particular program could be given authorship. In a somewhat earlier case which was actually between a law firm and Baidu, it was in fact determined that there was a need for human input. If there was no human input the copyright would not be recognized under Chinese law.

Now, incidentally, as far as copyright records are concerned, similar to Canada, China has a formality requirement, although copyright subsists because of the Berne Convention it has a requirement that it be formally recorded if you want to enforce it practically.

And at that stage, even if there is no formal or substantial examination as to subsistence of copyright, the Copyright Office will have a look at what do you input in the relevant field for author. And if it doesn't appear to be a human name, then they will pull you up on it and they'll ask you to extrapolate and to explain.

GORDON HARRIS: Thanks very much indeed, Elliott. And now to Samantha in Dubai.

SAMANTHA YUNG: Thank you, Gordon. So similarly in the UAE we run in a civil law jurisdiction. And currently there is no specific provisions in the UAE federal copyright law. Its implementing regulations or the intellectual property law of the DIFC that specifically addresses or clarifies the IP rights and protection of works involving generative AI.

But under the current UAE legal framework, some have argued that human authorship should be implicit in the copyright law because the law is drafted based on human involvement, such as provisions of moral rights, duration of copyright protection based on life of the author, similar to what Stephan has mentioned earlier on.

But there is also another line of thoughts that non-human expressions should also attract copyright protection. And this is also not prohibited specifically in the law. And arguably, it is also consistent with the economic justification to stimulate knowledge creation. And it also supports the overall innovation policy in the UAE.

But having said that, the law also further specifies that copyrightable works must be innovative, or in other words creative, in character. So this is similar to the requirement of the originality in other jurisdictions. The Court of Cassation of Dubai has clarified that innovative characters requirement bestows genuineness and distinctiveness upon the work. So just merely collecting, reviewing, or reproducing some information would not satisfy this requirement.

So based on the current UAE laws, seems like human involvement in the creation of work seems to play a part. And even if non-human creation, non-human expressions should be protected. I think the other question would be whether AI generated works present sufficient innovative character that can be protected. And as recently mentioned in an IP event by the Assistant Undersecretary of IP of the UAE Ministry of Economy, there are, in fact, looking closely to clarify the state of copyright law in view of the vast development of the AI industry in the country.

GORDON HARRIS: Yeah, thanks very much. I mean, even the UAE isn't really only just one jurisdiction. And, of course, there are other-- plenty of other jurisdictions in the GCC area. So are there any of those other jurisdictions where anything is standing out in the Middle East? Who is setting the standards here?

SAMANTHA YUNG: Right, Gordon. So amongst the GCC countries, actually, Saudi Arabia has issued a new draft intellectual property law in April this year which seeks to clarify various issues, including the relationship between IP and AI works. So this draft IP law proposes IP protection to AI generated works whenever natural person's contribution is prominent or substantial.

But on the other hand, if the IP rights of works purely generated of AI independently or the lack of human contribution, then it shall be remain in the public domain. So this draft law is still in consideration, and the question remains on actually the level of human contribution, and whether it would meet prominent or the substantial threshold.

So overall, it seems like there-- we might assume that there could be no copyright protection to AI generated works, so companies should really consider having an internal employee guideline to address the use of generative AI tools as one of a practical tip. And should always request employees to disclose such uses so the business and the legal teams are aware of this and could incorporate some of the appropriate adjustments, including maybe removing certain warranties as to the originality in any agreements or arrangements with third parties.

GORDON HARRIS: Thanks very much indeed, Samantha. That's great. OK, let's move on to another topic now and look at some of the key disputes over infringement for generative AI. It's fair to say that developments in this field are not without controversies. And when it comes to potential disputes, none of us is surprised that there has already been a fair bit of court activity in the USA. So let's start there. Franklin, over to you.

FRANKLIN GRAVES: Yes, I think it's fair to say we certainly have our share of ongoing controversies here in the US. So when we talk about copyright litigation in the United States, we're really looking at it how it explores and challenges the limits of infringement claims, including the popular fair use defense to these types of claims.

So infringement in these situations typically occurs when copies of a work are made to compile training data sets, or when a model is created if-- depending on whether or not that model would be considered a derivative work or not, and when a model can generate output that is substantially similar-- that's the key word here in the United States-- to the training materials.

So it's interesting to see how the bounds are pushed and tested in a lot of these litigation scenarios. So presently the legal battles resolve-- or revolve around various causes of action, mainly focusing on three-- the different theories of copyright infringement, like direct infringement, vicarious, and contributory infringement. There's also the often overlooked but critical matter of violating the Digital Millennium Copyright Act. And that has a provision on copyright management information, or CMI.

And so what this provision does is it prohibits the removal of copyright notices and author information, basically emphasizing the importance of preserving rights and metadata when copies of training materials are being made. Currently the spotlight in these lawsuits is largely focused on the developers of the models, such as OpenAI, Stability AI, Midjourney, and Meta. And if they are different entities, the associated service providers like Microsoft or DeviantArt.

So particularly Giants like Google and Microsoft are finding themselves in the crosshairs as they often act as both the developer of the model or the tool and also the service provider that is making that available to consumers or customers. So there is one exception to the class action litigation that we're discussing. The class action brought against Microsoft's owned GitHub for their code drafting assistant tool called Copilot.

That lawsuit is actually based not on copyright infringement claims, but rather the lawsuit is based in part on breach of contract for not following the license terms of the open source licenses that applied to the software code that was used to train the models. So, again, it's much more nuanced case that we don't have time for to fully cover, but it does include the usual DMCA claims for the removal of CMI.

There's also some-- it's worth exploring why these are being brought as class actions, but for the sake of time, we can circle back to that if people have questions or want to learn more about that. So it's worth noting that it's not just about copyright law, though. There's ongoing litigation exploring realms beyond copyright, which is my opinion, just as interesting.

But we are seeing claims that are based on violations of a federal statute here in the US called the Computer Fraud and Abuse Act or the CFAA. We're seeing stuff brought under privacy laws, although one of those was just recently dismissed. And we also see unfair competition claims and various state law claims and state laws coming into effect and being applicable to these activities. So the current listing includes over a dozen active lawsuits, mostly class actions, like I said, that are filed across the US and federal courts.

And as early as this month-- again, we're talking about current stuff because it happens frequently-- a group of authors in the Authors Guild-- which is an organization that represents authors and their interests-- like writers and stuff like that-- filed two lawsuits for copyright infringement. One against OpenAI and another against Meta for copyright infringement claims. Then one on screen-- yeah, the one on screen is from-- the one on screen is from Getty Images, but I think Matt will talk a little bit more about that.

GORDON HARRIS: OK, Matt, let's move on to Europe then. And what's the picture looking like there?

MATT HARVEY: So nothing like the states yet where there really does seem to be a new case every week. So far in the UK we have Getty versus Stability AI, so an exact parallel of one of the cases in the States. Getty is alleging that there is infringement at the training stage in the model itself and in the outputs.

Now we haven't seen the defense yet, so we don't know what exceptions Stability AI is going to rely on. But what we do know is whereas the US has this open ended fair use defense, which may or may not be found to cover these activities, in the UK, it's a narrow defense for text and data mining. It really requires that you're doing it for research for non-commercial purposes.

Now, they may come up with other defenses, but we'll see what happens. I'm not aware of copyright cases about the training and use of models elsewhere in Europe, but there are ongoing investigations about privacy and large language models. The biggest news is the EU's draft AI Act. And at the moment, it proposes that anyone deploying generative AI in the EU will need to disclose a list of copyright works used to train it. And I really think that could lead to a lot of litigation.

GORDON HARRIS: Matt, earlier on you said that there can be-- written down thousands or even millions, I've now heard you again today and you said billions of images, of items contained in the data set used for machine learning. So as a matter of practicality, how would you go about listing all the copyright works used to train an AI? I mean, surely that is a gargantuan task.

MATT HARVEY: It would be if you listed billions individually. I think in practice, most companies are using pre-baked data sets. So those five billion images that are being used for image generators it's typically people are using a specific database called the Leon Database. And I think that's what people will do. They'll say we use the Leon Database, we use the Common Crawl of the internet as at this date and see if that suffices.

GORDON HARRIS: All right, thanks for clearing that up. Now what about the rest of the world in terms of disputes? Stephane, do you want to lead off on the position in Canada? And then we'll head Eastwards.

STEPHANE CARON: Sure, Gordon. So there are no decided copyright cases in Canada dealing with generative AI. There's certainly been a mobilization of authors and creators, particularly in the cultural industries. What used to be cause for blanket exceptions to TDM have been met with demands for more nuanced approach.

The Writers Union of Canada, for example, has been pressing Canadian government to maintain copyright law protection against generative artificial intelligence services. The use of copyright protected content for machine learning is increasingly being scrutinized. We do have a-- legislative reform is also in the works.

Now, the government does have a proposed bill that is intended to deal with AI currently titled Artificial Intelligence and Data Act. It does not deal with copyright specifically, it deals more with ethical and privacy concerns. But what's interesting about that one is it does provide a framework for making businesses more accountable and AI services more transparent. And these are the beginnings of a government-- of a more involved government regulation in Canada.

GORDON HARRIS: Thanks very much indeed. China, Elliott.

ELLIOT PAPAGEORGIOU: Well, as mentioned earlier, there are really limited number of cases on point in China. In fact, there's no final copyright infringement unfair competition case. There would be the other head who would try to park this under in China. At the moment that has been decided and reported.

There are a couple of pending cases that we're aware of but as I said, they're not in the public domain yet. Having said that, really these cases are imminent. And there's really no legal obstacle to a litigation on the basis of copyright infringement or unfair competition. In fact, unfair competition might have better legs in this situation, China being a civil law jurisdiction and they're having a scope of expanding this head of action of unfair competition.

And unlike copyright, you don't have to necessarily prove human authorship. You do have to show that there have been a leveraging of somebody else's input, which is unfair, and that there has been some damage incurred. Having said that, these cases we believe or I believe are going to be seen where somebody using the generative AI appropriates, or China might say misappropriate something that is in the realm of Chinese cultural content quite by accident.

So you're trolling the internet, you pick up some material, It's considered by China to be something which is culturally sensitive, it's been integrated into something else. If that sort of action were brought in China, I think you'd have a pretty good position if you were to argue as the owner of that cultural content that it's been misappropriated.

GORDON HARRIS: Thanks very much. It is interesting, isn't it, to see unfair competition really flexing its muscles in China or across not just copyright, but other intellectual property rights as well. It's a good underlining and another option. OK, Samantha in the UAE.

SAMANTHA YUNG: Thanks, Gordon. Yeah, in the UAE we are also not aware of any court cases raised regarding generative AI in IP infringement yet. But I think as more generative AI products will be released and adopted for commercial use here, more issues may be raised by rights holders. And we will also see more discussions as the cases in the US and in the West develops.

On the policy side, the UAE government has actually set up the UAE council for artificial intelligence, which is led by the minister of state of AI. And is responsible to propose policies, to promote AI friendly ecosystems, to encourage research, promote collaboration, and to implement the UAE national strategy for AI in 2031.

So there are also non-binding guides that was introduced by the government, including introduction to AI that identifies key considerations to be reviewed, such as AI governance, data governance, cyber security, et cetera. And also AI ethics principles and guidelines that was issued in December 2022 which promotes transparency, fairness, accountability, and preserving privacy.

So at this point, there is no guidelines or regulations addressing challenges posed by gen AI in terms of potential IP issues, but under the patronage of Sheikh Hamdan, the Dubai Future Foundation will host the Dubai assembly for generative AI next month. So we will see that various roundtable discussions or working groups will be formed after the event to discuss on some of the law-- IP laws, including copyright law, we think.

GORDON HARRIS: Thanks very much indeed. OK, since it is likely that the act of training may involve effectively infringing someone's copyright-- we've already touched on this earlier, but let's look at it in more detail now-- what built-in exceptions are occurring in national laws to help with this process? Governments presumably want to encourage generative AI don't want to stamp on it, but it is inevitable, really, that in many cases rights will be infringed. So let's start this time in China with Elliott and work our way back.

ELLIOT PAPAGEORGIOU: It's quite a presumption that the government is going to encourage AI in China. I guess it would if it's with-- how do they call it? With a flavor of according to Chinese legal principles, maybe then it would encourage it. But getting down to the topic, as far as the current Chinese corporate legal framework is concerned, we, obviously, don't have the sort of broad exemptions that you might see under first Amendment rights in the US or in other jurisdictions.

In fact, article 24 of China's copyright law has really very limited exceptions to copyright infringements, though, such as individual study, research appreciation. Common ones such as appropriate quotation use, incidentally used as part of reporting news classroom teaching, things like that used by the organs of the state.

Now there are some additional categories which you might be able to use as a safe harbor. One of them, really, from a commercial perspective is not very interesting because it relies on the particular justified use being gratis for non-profit fearless performances. In that particular situation, you could probably claim that there should be the benefit under this particular protection. But, again, limits for commercial use.

Secondly, copying or painting photography, et cetera, of a work of art however that's defined by whoever it's created, which is in a public space-- and this is an interesting point in China-- provided that you attribute the art piece and provided you give an indication of acknowledging that this is not your work, that it's a photograph of somebody else's work, that is also something that could be helpful in a situation where the AI relies on public photography and part of its material that is processing public photographs or works that are visible to the public.

And then last but not least, there's an omnibus clause which is the usual in any other circumstances stipulated by laws, administrative regulations, et cetera, et cetera. So that leaves the back door to a further development of this area of protection. But generally speaking, there is no specific protection that I think would form a safe harbor for AI use. Interesting. Thanks very much. Samantha.

SAMANTHA YUNG: Thanks, Gordon. Yeah, so in the UAE as well there is no clear general text data mining machine learning exceptions in the UAE. But in the copyright law, it offers quite a narrow copyright protection exceptions. So similarly, like making a single copy of the work by a non-profit archive or library or authentication offices either for the purpose of preserving the original copy or to respond to a request for study or research. So you can see it's very limited to nonprofits and also for study or research.

So it seems that in the federal law, copyright law since it will be quite narrow. On the other hand, with the DIFC IP law, it seems to contain a slightly broader type of copyright exceptions on criticisms, discussion, or information. There might be some room but it is still unclear. So DIFC is a financial free zone that is an independent jurisdiction within the UAE.

And it has its own legal and regulatory framework. But, yeah, as overall, all of this is not tested or tried yet in the court, so it still remains a point to be clarified hopefully by specific legislative amendment or interpretation guidelines. So overall, another tip we suggest is those who intend to train AI in the region to keep an eye out on the development of the laws and also use licensed or clean data sets and models, that'll be best.

GORDON HARRIS: Thanks very much indeed. OK, we're going to move on to have a look at North-- position in North America now. And I think Stephane and Franklin maybe you can deal with this point between yourselves. I'd like you to deal also with what I call the Bart Simpson question, which is, what is the worst thing that could happen? going to do something bad now, what is the worst thing that could happen? What is the actual loss and damage that might be suffered through a copyright infringement in this area? So Stephane, I think you're going to kick us off.

STEPHANE CARON: Hey, Gordon, so maybe just to then answer your question is clearly in a scenario where a court would hold that there has been an unauthorized reproduction of a work and, therefore, copyright infringement, there would be liability for copyright infringement. What might that look like in Canada? Well, it would be on the one hand any profits that have been derived from that activity. And that might be downstream profits from licensing or selling AI generated content. Again, here I'm going on basic principles.

The other aspect as well is to the extent that conceivably the generative AI results in a competing problem, a product that causes loss of sales. That would also be considered. And, actually, that's a very good point to lead me into the answer to the broader question that we're asking ourselves in this section is, what are the exceptions that are applicable?

So very much like the other jurisdictions, we don't have a specific exception right now in Canada for text and data mining or for machine learning. And maybe what-- but we do have general exceptions that can find an application in the TDM context.

I'll just say this as well very quickly is that there's a tendency to default to statutory exceptions as a means of addressing copyright issues in the TDM and machine learning environment. And it's interesting to keep in mind that there are also licensing models that can permit cooperation with rights holders that can be quite beneficial.

Although I don't know that she's-- anybody has taken her up on her offer yet, the Canadian pop star Grimes is offering the use of her voice in the creation of AI generated music without penalty. Offering to split 50/50 the royalties of any successful AI generated song that is created using her voice.

There's also an other example of the estate of the painter Magritte whose licensed scans of his artwork in order to generate a further Magritte works that can be licensed and sold. So these are interesting avenues, but let's get back to the exceptions quickly. We have two general exceptions in Canada that are often referred to in the TDM context. One of them is fair dealing and the other one is a narrower exception for temporary reproduction of in technological processes Section 30.71 of our Copyright Act.

Let me focus here for our purposes, though, on the fair dealing. And that is that in Canada, the fair dealing exception and probably for the purpose of research more specifically in Canada would be-- could be applicable. It's important to note that it has received a purposive interpretation from Canadian courts. And it could apply equally to commercial and non-commercial, depending on the circumstances.

So there's potential here for application even in commercial context. And Franklin can discuss it further. I know it shares a number of similarities with fair use exception in the US, but it is-- it requires an analysis of various factors to be weighted in coming to a conclusion as to whether this reproduction is fair or not.

And it is very much a case by case exercise. So while it has-- it is flexible in the scope of its application, it is not always as predictable in terms of the scope of the exception provided. And just to end on this point and to tie it back to our comments initially is that one of the aspects, of course, is the impact of the work on the-- of the AI generated work on the original works that have been copied.

And to the extent that the work product is competing with the original works, to the extent that AI generated content is competing with the original human created works, then there's going to be, I think, a higher concern, a higher likelihood that these exceptions may not apply, or they should certainly be considered from a policy perspective as well going forward.

GORDON HARRIS: Thanks very much indeed. OK, so, Franklin, could you put the US gloss on that?

FRANKLIN GRAVES: Yeah. So I think a lot of what Stephane just said is very applicable, especially that distinction. So here in the US we have four factors under a fair use analysis. You can Google it and find out what those are. They're pretty easy to find and understand, but applying them is the issue.

And so we're actually seeing that right now with a court case that's been going on since 2020 between legal research company Thomson Reuters and Ross Intelligence, which was-- Ross was a-- they got put out of business because of this litigation. But they were a tech startup that offered an alternative to the world of legal research. And so they are now going to jury trial. So that's the latest update literally as of this week, as of, I think, two days ago.

So, again, as Stephane said, it's highly unpredictable. And I think that's the risk. And so looking at it, taking a step back I think looking at it, you have to look at it in three different buckets. So the training materials, the training data sets that are used have risk associated with them. The models that are the outcomes of those data sets and the systems, AI systems that are built around that have their own set of risks. And then also the outputs-- that's the third bucket-- have their own set of risks.

And so the fair use analysis can apply all the way along, but whether or not it will stick is the issue and it's yet to be determined. So I think going back to what Matt said, actually, I wanted to touch back on that. And I'm going off my plan to talk here. But going back to what Matt said, I want to point out there's a great website called huggingface.co that if you're working with engineers, working with developers in this space-- they know about it and you should too if you don't-- it's where people can upload models and also data sets.

Go and look at what are called data cards or model cards because that outlines what Matt was alluding to the EU AI Act is already getting at, which is disclosure of what is being used to develop and train a model. And the really good ones or the-- it's kind of like if you've ever used open source software, you're going to get ones that are really high end and people know what they're doing and the developers are doing great documentation and explaining how they got to where they got.

But you're also going to have ones that fall off the rails there and don't disclaim or disclose biases, risks associated with the model or the data set, or even the sources of the data. And so that's where the model cards or the data cards come into play.

So I wanted to highlight that really quickly. But, yeah, really, I think the analysis of fair use is up in the air. So unless Congress acts here in the US-- which is highly unlikely at this stage-- the Copyright Office is going to continue what they're doing, which is trying to support with their interpretation of US law, which is a human authorship requirement. And so, really, we're just going to sit and wait until these court cases come out, or as a result of the court cases, Congress acts due to lobbying.

GORDON HARRIS: OK, thanks very much indeed. Right, finally then we'll come back to Europe. And it's a good principle in life, isn't it? If you want to know the answer, ask the guy who wrote the book. Matt, over to you for the last one.

MATT HARVEY: I'm conscious of the time so I'll be brief. In Europe there are harmonized exceptions. The most promising is temporary copies. So if you gather the data, you process it, you train the model but then you dump it and delete it, could that be an exception? There has been speculation that could work.

The EU copyright directive introduced text and data mine exceptions across Europe, and that covers both non-commercial and commercial purposes. For commercial purposes, that's subject to an opt out. We're not quite sure how that should be done in practice. There are some websites that purport to allow you to do it. We'll see what happens there.

Now, the UK illustrates the challenges of changing the law in this area, the currents of politics and international trade. So we at the moment decided because of Brexit not to implement that commercial exception for text and data mining, so we can only do it for non-commercial research. We then proposed to have a widest in the world to allow you to do text and data mining for any purpose, but this caused a backlash from media companies and other rights holders, but also international trade partners.

We're back to the drawing board, we have a consultation, we still only have non-commercial purposes as an exception. We're expecting a voluntary code as the next step, but we are a few years off the mark. It's not very pro-innovation I'm afraid here.

So what I would say is we're competing with jurisdictions like Japan and Singapore, which have very wide exceptions for text and data mining. And I think there is scope for forum shopping when you're deciding where to train stuff. That does rely on an analysis of whether the model itself is a copy.

If it's not a copy, train it's somewhere else. Bring it to the UK or where you're based. If it's a copy, maybe an API to another country? We'll have to see how you get around that. The other thing I'd want to mention is technical measures. So a lot of this depends on the model size versus the training examples. There are technical measures which may mitigate the risks of infringement.

GORDON HARRIS: Thanks very much, Matt. Well, we said at the start that this was going to be a highly practical session, so let's be as good as our word. We've got a couple of minutes left, so it's a few seconds, guys. That's it, one line. Can each panelist tell us what is your main takeaway message for the audience? What is the practical tip you would like people to remember at the end of this webinar? Start with our guest speaker Franklin.

FRANKLIN GRAVES: Thank you. I alluded to it earlier, but open source software has an existing framework that a lot of organizations or a lot of people are already familiar with and can be followed in these instances of how to develop around AI or have third parties develop around it and how to draft for it in your contracts and track and monitor the use within your organization.

GORDON HARRIS: Thanks very much. Samantha.

SAMANTHA YUNG: Yes, so in terms of limiting risk, I've also mentioned previously in terms of the uncertainty around tax data mining and machine learning exceptions and other copyright law exceptions might not apply. So it's recommended to use clean licensed data and also be aware that there are generative AI tools out there that are trained by licensed clean data as well. And those are also good tools to look at.

GORDON HARRIS: Thanks very much indeed. Stephane.

STEPHANE CARON: I think it's going to be that, really, we have to assume that AI generated output will not be protected by copyright on its own. So if your business model leverages copyright, it's important that you still need human authors involved in the creative process and to document their participation.

GORDON HARRIS: Thanks very much indeed. And Matt.

MATT HARVEY: Have a company policy for the use of generative AI. There are lots of practical measures and technical safeguards to avoid infringing outputs. And you can also deal with trade secrets and privacy, hallucinations, and other risks at the same time.

GORDON HARRIS: Thanks very much. And Elliot.

ELLIOT PAPAGEORGIOU: But I do have to say think and borrowing here-- claiming very much maybe this is coming from China, I feel a bit safer about this Pinterest image. I think the key thing is to really keep a sense of common sense. I think Matt, Franklin, I think the entire team have basically said the important element is to look at what your business is made of and to keep a good common sense about what your business depends on. This is especially important in China.

So if one could capture it in two pithy sentences, we should really make sure that we concern ourselves with keeping our real intelligence and not get too wound up in artificial intelligence. Because, ultimately, it's the business case that still matters. And think Matt, I think Stephane, Franklin, Samantha, and yourself Gordon, will all agree that ultimately we're here to support business so hopefully we'll stay on the ground with this.

GORDON HARRIS: I should have guessed that if anyone was going to finish with a joke, it would be Elliott. But thanks very much indeed. We've had a couple of questions, so what we're going to do, we're out of time so we'll answer those questions separately. I saw one I think Franklin answered somebody direct earlier on, and I've seen another couple of questions in. So we will come back to you with a written answer, which we will broadcast to everybody after this event.

FRANKLIN GRAVES: Real fast, I think the one that came in is an anonymous attendee, and I think Matt touched on it. I think Matt touched on this question, which I think another way to phrase your question is, the UK allows for generative works to be registered. So how does that then transfer to a jurisdiction that doesn't like the US or Canada? And has that happened before and been upheld?

MATT HARVEY: No registration here, it's automatic. We will respect generative AI wherever it's made, but it's a question of national law elsewhere.

GORDON HARRIS: Oh, and I think there is going to be an issue of what we would loosely as lawyers call forum shopping going on here where maybe there are some places where it's a very good place to be teaching your machines more so than others. So we will put out some more output on that.

Thank you to all of the panel for their insightful contributions today. This is a topic, which is going to go on growing and developing. As we have heard, a lot of countries are addressing how best to deal with this from an IP perspective. And we can expect many more significant developments over the next year as we've indicated. Some of the countries may do different things which may lead people to take different routes as how to go about this.

We will come back to you, probably in a year's time or so, with a follow up to see what has happened and to see if our panelists' predictions are coming true. Thank you to all of you our guests today for your time and attention. Please look out for forthcoming events in this series, the life cycle of a smart idea, which is looking at all aspects of intellectual property from a practical user's perspective.

This event will appear on our website soon as a recorded event. So if you want to encourage other people in your organizations to watch, you will be able to do so. So with, once again, many thanks to the panel for their hard work and preparation, thank you and goodbye from all of us. Thanks very much indeed.

ELLIOT PAPAGEORGIOU: Thank you.

FRANKLIN GRAVES: Thank you.

SAMANTHA YUNG: Thank you.

[WHOOSH]