Gordon: Ladies and gentlemen, I won't keep you waiting any longer. The numbers are going up nicely and since the opening item is an introduction hopefully the majority of our guests will have arrived by the time we get down to the main body of the content. Good day to you, wherever in the world you may be watching or listening to this webinar, and welcome to the latest installment of our long running IP focus webinar series, The Life Cycle of a Smart Idea. We followed our idea from its inception through choices about the best form of protection, branding and advertising the invention, enforcing rights of different types and even pursuing trade secret thieves around the world. Today we turn to the increasingly relevant and important question of artificial intelligence and the data which lies behind it. What if the bright idea is in the area of AI? How is that best handled in terms of protection, commercialization and delivery? Well, to help us through this new and complex area we have a panel of experts from around the world. All technically experienced in the field of AI and its implications for intellectual property and the wider business world.
So let me introduce the panel to you. First up, from the UK, Matt Hervey. Matt is our UK head of artificial intelligence and a partner in our intellectual property team in London. He is also the general editor of the Law of Artificial Intelligence, a practitioner's book published at the end of December, 2020 by Sweet & Maxwell. Matt also co-wrote the chapter on AI and Intellectual Property and he advised companies of all sizes from FTSE 100 to startups on their IP strategy for AI. He is in AI related working groups for AIPPI, the International Chamber of Commerce and the IP Federation.
Moving across to Canada, welcome to Shahrzad Esmaili. Shahrzad is a registered Canadian and US patent agent and a partner in the IP group of the Toronto office. She holds Masters degrees in both computer and electrical engineering. In fact her graduate thesis, which focused on machine learning methods for classifying music genre, sparked her curiosity and interest in AI. She's a frequent speaker on the unique challenges of global AI protection including for Lexpert and Landslide. She's passionate about gender diversity and stem subjects and has co-founded the Canadian Chapter of ChIPS, the non-profit organization aimed at advancing diversity in tech and IP law.
Further West to Vancouver, to Roch Ripley, Roch is a partner and head of the intellectual property department in our Vancouver office. He's a lawyer and patent agent with a degree in electronics engineering who provides strategic intellectual property advice to companies of all sizes and works with clients to protect AI based technology, contractually, and with patents and trade secrets. He has particular expertise in AI vision processing technologies.
Now whether we keep going West, or whether we go back East, one way or the other we arrive in Beijing and meet Jason Yang. Jason is a principal patent attorney in our Beijing office, responsible for the prosecution and invalidity of high tech patents in China. Jason used to be an R&D engineer in the semi-conductor industry and he's now highly experienced in various patent related works with a primary focus on prosecution, drafting and invalidity proceedings. He's worked with clients and attorneys around the world and he's especially interested and experienced in AI audio processing.
Today I'm going to lead our team of experts through this area with a series of questions to bring out the issues. Now, you can add to those questions using the Q&A button at the bottom of your screen, and if we don't get through all the questions you raise during this session, we will copy and note them and publish a follow up paper dealing with the issues raised. We're very keen to ensure that you find out all you want to know from today's session. At the end of our session our four speakers are bravely heading into breakout rooms so you can join them to ask follow up questions of particular interest. Now many of you have already registered for that but anyone can do so on the day and we'll come back to that later. We're also going to ask you, our audience, a few questions as we go along using a polling system. That's how I'd like to start. Let's establish just how much experience our audience already has in the field of AI. So a quick audience poll, please. Is your company or firm investing in AI? And there are your options. I get to see the results coming in real time which is fascinating. With a pretty good proportion of votes already in there, interesting, the largest component is 'just starting', ahead of 'no'. Quite a few planning to and some say they've made very significant business. Even a few where it's the core business. So, probably in line with industry as a whole. Let's move on now, shall we?
To set the scene for us I want to start by asking Matt to describe the current landscape. You know the old adage. If you want to know the answer ask the guy who wrote the book, so Matt, how has the technology underpinning artificial intelligence changed to the point where real viable solutions are emerging? And what's the significance of the role of data in enabling the current advances? Over to you.
Matt: Thanks, Gordon. Well it's interesting, isn't it? That poll did show that even among the attendees here we've got a wide range and an increasing number of companies that are actually making significant investment or where it's the core investment. So let me sketch out why I think AI matters to more and more companies. I think the counter-intuitive point is that actually AI research is really old. It dates from at least the 1940's, but there's been a dramatic leap in performance really only in the last 10 years, and that's all about a subset of artificial intelligence known as machine learning. Which in essence is computers programing computers. That method has cracked previously intractable problems, and especially language and vision, and why does that matter? Well, there are two key points. First, AI opens up opportunities across all industries and it's the opportunity to automate previously labour intensive tasks. So vision unlocks self-driving cars and medical diagnosis. One of our clients is involved in using machine learning for rapid diagnosis of COVID, analyzing samples visually. Language allows you to assess an insurance claim by reading bio-machine. Thousands of documents in the medical history. So that sort of automation may transform the speed and costs of any kind of business. Second, AI may disrupt traditional business models. Automotive clients, for example, are using machine learning for vision to enable autonomous vehicles, but then they start to sell mobility as a service, and then they establish this service of mobility as the platform for ancillary services. Services oft enroute. Things like search, content recommendations or virtual personal assistants and that's a chance to take a slice of a customer's purchasing decisions in their general life.
So now let me explain why companies increasingly want IP strategies for data in particular. That's because this revolution in machine learning has been enabled by three changes. Those are faster computers, increased memory and, this the key point, vast data sets. Now data isn't essential to all machine learning but it is key to what's known as supervised learning. That relies on very large labelled data sets. Now that could be a pre-existing data set so it could be a whole series of x-rays which are already labelled with the doctor's diagnosis. If you get a set like that you can train an AI to make the same diagnosis that the doctor makes. Or it could be a new data set. You take dash cam footage from cars and you actually pay people to l laboriously label everything in the image. Every pedestrian, every cyclist and, again, if you take that training data you can train up the AI in a self-driving car to identify road users. So access to the right kind of data could be key to unlocking an entire market. The cost of an appropriate data set could easily run into millions of dollars. But, fundamentally, data is hard to value. It's not used up by one party using it to train an AI. The value might be radically different to different parties. Let me give an example of CCT, the footage of a street. That could be used by one party to do geography or street planning or mapping. Another person would want to derive driver behaviour from footage. Another one would want to know about traffic flows. A fourth one might want to know about weather. Each of those parties might find that data set more or less valuable, depending on what other data they already have access to, to combine it with. We also have an evolving market here. So the EU, in particular, has measures to create public data pools and to create data markets. Then we've got the possibility of competition law coming into play to enforce access to third party data, to breakdown the whole that the data encompass, and then there are concerns about privacy, data bias. So all of that is making this a complex field and what's the key takeaway? It's this. Given the uncertainties in value of such data we help clients to leverage IP, in all its form, and trade secrets and contract, to be effect to protect and retain control of their data.
Gordon: Thanks very much, Matt, and obviously we'll get a return to some of the IP implications of that new technology a little later. Next can I turn to Shahrzad in Canada. What is the current landscape of AI in Canada? What are the trends? How do you think the industry can improve itself in Canada and is the government participating in this to support the booming R&D?
Shahrzad: Thanks very much, Gordon. These are great questions, and seeing the audience response, clearly there is interest in AI and many are beginning to dabble in AI. According to a CEPO report measuring AI activity in Canada, Canada's one of the top countries in the world, in terms of AI patents assigned to Canadian researchers and institutions. In fact, we are 6th globally behind China and the US. Canadian researchers and institutions have accounted for nearly 2% of the 85,000 AI inventions patented world wide, in the chart that you see up there, and AI patent filing activity for Canadian has grown significantly in recent years. In the 5 year period between 2012 and 2017 the number of Canadian AI patents filed increased at an average rate of 31% each year. The numbers have continued to grow since then and what is fascinating, to me, is in terms of the AI specialization fields, and I know Matt touched on some of these, Canadian AI researchers are focused on AI technologies relating to natural language processing, knowledge representation and reasoning and computer vision. To answer your question on how can the industry improve in Canada, this is something that's near and dear to my heart, the industry can improve in its inclusion and gender diversity. Currently there's one female AI researcher involved in AI patents for every 6 male researchers in Canada. Internationally, the scene is a bit better. There's one female AI researcher for every 3 male AI researchers. So we do need to improve on that front in Canada. To answer your question on what is the Government of Canada doing, there are many different initiatives used by Canada to support AI research and development, and one of them is in 2017 the Government of Canada contributed 125 million dollars to set up the Pan-Canadian AI strategy. Along with this being the world's first national AI strategy the goal is to, unsurprisingly, retain, attract talent as well as supporting business growth and collaboration. Some positive results that have come out of these types of investments, a report from U of T showed that Canadians have filed the most AI patents per capita within G7 Nations and China, between 2015 to 2018. So it does appear that these initiatives to increase AI activities are working. My takeaways are that although there's certainly a trend towards increased patent filing activity relating to AI, and clearly an incredible knowledge base of researchers, Canadians still need to increase their AI patent filings to keep up on a global stage.
Gordon: Thanks very much. That's great. Roch, obviously this is attracting the attention of governments, not just in Canada but around the world, in terms of potential major disruption and changes. So where we might find those disruption and changes? What are the sectors where AI seems to be developing fastest? What's happening just over your Southern border in the USA?
Roch: Thanks, Gordon. That's an interesting question because there's broad based investment in AI, as we've heard Matt and Shahrzad say, and not just by traditional tech companies. A lot of businesses that generate or have to manage a lot of data, or the data can't be processed using traditional what's called a deterministic algorithm, so an algorithm says if X happens then do Y, they're investing heavily in AI. So if they're technology companies they maybe doing it themselves. If they're not technology companies, like banks and law firms, then a lot of them are investing in AI tech companies and having those tech companies do the work. A couple high profile acquisitions, DeepMind was sold to Google for 500 hundred million, roughly. SwiftKey was sold to Microsoft for about half that, 250 million. So big numbers and high profile transactions. Few sectors that we've seen a lot of activity, just building on what Matt said earlier, automotive, of course. You've got the image processing. The visual images, the radar images that are coming in, and perhaps the most difficult problem, predicting driving behaviour from the data in those images. Matt also talked about health. So pharmaceutical drug development, personalized medicine. Financial services, banking. Financial transactions. Insurance, we heard Matt talk about that too. Social networks, there's an incredible amount of work being done here. Facial recognition with all the images we upload to the internet. Image classification. If you're on Instagram and you upload anything related to COVID you may have noticed that it will automatically get flagged as a COVID related post. Natural language processing. In the news recently, of course Twitter and content there, that's automatically flagged as being generated by a bot. The most important use of AI on the internet, video filters and making me look like a bear, right now, on the presentation. I'm going to have to shut this off before I talk about patents to remain credible.
In the US, the development most people talk about is the 2014 Supreme Court decision, CLS Bank v. Alice. That injected a lot of uncertainty into computer implemented inventions, generally. So the issue there was subject matter eligibility. Not whether the substance or the intention was new, or sufficiently different from what came before, but whether it was the right stuff in which you could get a patent. Was it too abstract an idea in which to get a patent or not? So applied to AI, the concern would be is the underlying AI model, like the kind of neural network you're using, is that too abstract, regardless of how you implement it, to get a patent? For a few years that uncertainty made it a lot harder to get patents at the US PTO and to enforce them in court. A couple of years ago in 2019, the beginning of 2019, the US Patent Office issued new practice guidance which, to a very large degree, has it made it easier to get patents for AI based inventions. Generally speaking, if you can show that your AI based inventions integrated into a practical application, or some kind of technical benefit, I haven't found the eligibility issue to be a problem before the office right now. In fact one of the examples the US PTO issued in 2019 for subject matter eligibility was of how using a neural network for facial recognition was eligible for patenting. But there is still uncertainty before the courts. There's the guidance, is it binding on the courts? But I do think we're starting to see the pendulum swing a bit back in favour of the patentee.
Gordon: Thanks very much, Roch, and well done for being able to switch that technology off. Obviously there's a famous example of someone not doing that. I will certainly not try it. If I suddenly look like a cat it's because my cat has actually arrived here, which is not out of the question. It's not just in the so called developed economies of Europe and North America where the rapid development of AI technology is visible. No surprises here. The fastest growing large economy in the world over the last decade has been China. So, Jason, what is there to report in terms of the development of AI, ongoing innovation and government support in China? Which sectors are you seeing most involved in the development and growth of AI technology in China?
Jason: Thanks, Gordon. I'll start from the general point of view before getting to the Chinese practice later. As Shahrzad showed in the graph you can see that China has become the biggest country in terms of AI filing and it has developed pretty much every sector. But compare to the US, as Chinese companies tend to invest more at application level. The reason is so obvious. China has a large market. The return from the application's much quicker. Some of the strongest AI fields invested by Chinese companies natural language processing and machine learning, I think. Natural language processing has brought up many applications in China. Basically it's technology for computers to perceive human languages so they can understand the contents from a document or speech. Some of examples of natural language processing in China, intelligence customer surveys, virtual assistant and machine translations. These applications have played an important role for e-commerce and that's part of the reason why the biggest AI filer in China, up by two, Alibaba, Transcend and Huawei, all of them have been very successful in e-commerce and consumer products. Speaking of government support, there have been a dramatic financial supports for AI companies. Be it cash or tax returns. I believe that happens in many other countries as well but in China they're also discussing about policies for providing free data sets for companies so they can better develop AI technologies. Of course, just like Matt mentioned, data is very important for AI companies because they need sufficient data to optimize models or algorithms. Although there are many free data sets available on the internet now, it's still more useful to get data from local government, especially like data for traffic, data for health care, data for education, etcetera. Considering the privacy and security issues in this area, this kind of data is usually hard to get from open source. Speaking of ..., my take is that I think self-driving and health care are promising to apply to for AI to develop in China in the next few years.
Gordon: Thanks very much, Jason. Well, this is an IP focus series of webinars so it's only appropriate to move on to the issue of IP. Before we get started, another question for you in the audience. The question this time is this, what barriers might there be which would prevent you seeking to patent AI related inventions? Is this patentability? Budget? The possible enforceability? The fact that you have to disclose for patenting as opposed to keeping secretive or anything else? Let's see those answers come in. Interesting. Very interesting to me that the patentability and the disclosure secrecy point are winning out here. Obviously budget and enforceability are still there but patentability and disclosure secrecy seem to be the main issues. We will reflect on that as we move forward. Thanks very much, indeed, for participating. It's not difficult to see how the whole range of IP rights are involved in the possible protection of AI technology and its underlying data. It's equally obvious that if there's going to be large scale commercial exploitation of AI then businesses will want to understand how best to protect their investments. So back to Matt for the scene setting. Matt, can you do just that? Set the scene for us with an overview of the core issues underlying the IP protection of AI across the board and then we'll return to the individual rights in more detail.
Matt: Thanks, Gordon. The sad truth is that most of IP was not designed with AI in mind so it's really important to understand the gaps and to plug those, particularly of trade secrets, with contract and with practical and technological measures. So to give an example, the fundamentals of patent eligibility in Europe were developed in the 1960's for the EPC. That was a time when we had mainframe computers operated by punch cards. Some very basic graphics and no internet at all. Now AI was an area of research, but it was focused on coding by programmers, not machine learning. So it's really 50 years before the current advances we're seeing. Most of IP law never was designed to address two key developments. One, the importance of data which I think we've now all spoken about, and secondly, machines able to produce human like outputs. So creative works for purposes of copyright and inventions for the perhaps of patents. So here are the key strategic points. You need an effective patent strategy and to do that it needs to navigate excluded subject matter sufficiency and enforcement. You need to use copyright to protect code. But always be aware it doesn't protect the actual functions of a program. Where IP, or traditional IP, doesn't work you have to use trade secrets and contract because both of those are apt to cover data, algorithms, models in sight and the like. Finally, exploit jurisdictional advance which we can advise. For example, in the UK we expressly allow for computer generated copyright works and designs. Indeed our firm successfully litigated the leading case in that field.
Gordon: Thanks very much, Matt. That's really helpful. Now in the first of these webinars, which unbelievably is nearly a year ago now, we considered the selection process of whether you should seek a patent or trade secret protection for your bright idea. Roch, could you talk us through that process in relation to the protection of AI technology? What is it? Patents or trade secrets? Or a little bit of both?
Roch: Sure, I'd be happy to. It's an age old question and I don't think anyone is going to be surprised to hear me say there isn't a single answer that applies to all AI based technology, given how many sectors we've already heard it covers. So like the poll question that you just saw on your screens, reference the legal downside to patenting is disclosure. Generally speaking, whatever you put in an application is published a year and half after filing, regardless of whether you ever get a patent. I'll add here there is one exception that the US PTO offers. If you commit to filing only in the US you can request non-publication, which means your patent won't publish until granted. So you can avoid the situation where your technology is published without getting a patent. The first question is, with that in mind, can you exploit your AI innovation without disclosing how it works and will it take competitors more than that year and a half to reverse engineer? Or otherwise, recreate what you've done once you go public with your innovation, with your product. Will you be able to recruit talented employees if you don't let them publish in this field? Because this is a field where a lot of talent, they would make a condition of working the right to publish their research to further the field and for their own reputations. So if the answer to any of these questions is no you might as well patent, budget permitting from a legal perspective, because you'll be disclosing anyway. If it's going to take more than that year and a half it gets trickier. How much longer than that year and a half? Is the technology still going to be valuable then? Enforceability, also mentioned in the poll, can you detect infringement? Because if you want to enforce, even in jurisdictions where you're allowed discovery, you'll need some kind of reasonable basis to commence a suit. For example, if your AI is going to be used entirely behind the scenes, like in an assembly line, that's a factor against patenting. Technical merit is important to consider too. If you've got a very marginal invention, technically, it usually means it will be harder to get a patent so there may be no point in spending the time and money on patenting. One thing that I want to add that I think is particular to AI is that, respective disclosure, you can kind of have your cake and eat it too right now, so to speak. There's not a requirement to provide an actual copy of the training data you use in a patent application. So you may be able to patent a trained network, for example, and while you have to describe the kind of training data you've used, it would be up to your competitors to source that training data independently. So practically, given how expensive like Matt said it can be to acquire that data, this can give you the benefit of patent protection while still making it practically difficult for competitors to copy your innovation, from what you have to disclose.
Gordon: That's a great point and very interesting in the light of our poll, isn't it? Where people were concerned about patent eligibility and secrecy and you're saying that you might be able to have your cake and eat it. That's really interesting. So moving on to some of the more specific issues of patent protection, Shahrzad, are the patent examination rules friendly to AI in Canada and have standards changed at all in recent years?
Shahrzad: Thanks, Gordon. Clearly, looking at the poll results, patentability is something that's top of mind for our viewers here today. In Canada, traditionally examination rules have not been friendly to software and AI. AI inventions are examined through a lens of computer implemented inventions and traditionally have been tough to patent because of CIPOs problem solution approach to patent examination. I'll get into that a bit. This approach to claim construction and determining the essential elements of a claim, basically required that the examiner first identifies the problem proposed in the patent application, and then determines which elements of the claim are essential based on which elements are necessary to providing a solution to that problem. The issue is that the patent examiners would often characterize the problem in such a way that the computer, be it hardware, software components of the invention, were somewhat conveniently found non-essential. Then all that remains in the claim is unsurprisingly abstract and near principles or abstract the ... are excluded from patentability in Canada. So now the good news. Claims construction and what forms patentable subject matter, especially in relation to computer software and AI inventions, has undergone a seismic shift in the past year with the release of the Federal Court of Canada's Choueifaty decision. While it may be too soon to call, patent examination is likely to be friendlier towards AI inventions, post this decision, especially if the AI inventions are adequately described. Now, what did Choueifaty refer to? Choueifaty involved a patent for an anti-benchmark portfolio, a financial portfolio, and described a computer method for selecting and managing investment portfolio assets that minimized risk without impacting returns. The Patent Office applied a problem solution approach such that the computer was found to be non-essential and the invention considered to be a scheme or rules that amount to mere calculations. Now, at the Federal Court level, the Federal Court rejected CIPO's problem solution approach to claim construction. CIPO then put out guidelines in response to this decision to clarify. Use purposive construction to determine claim elements as essential or non-essential. What does this mean? All elements of the claim are presumed to be essential unless established otherwise. Also there's a requirement, as Roch even mentioned, that the invention has a physical existence or causes a physical effect or change. According to the guidelines, in computer inventions, including AI inventions, the computer invention will be considered patentable subject matter if running the algorithm on the computer improves how the computer functions, thereby solving a problem related to the manual or productive arts, as required by CIPO. In the end, CIPO found the Choueifaty patent to be allowable because the invention provided specific computer advantages. It was said that the computer calculations were considered not to be merely for yielding information but for permitting the computer to perform portfolio optimization with greater speed and less processing. That was important. The speed and less processing. There were technical advantages here. What are the takeaways for AI inventions then? Some of the things for you to think about is does the AI algorithm, running on the computer, improve the functionality of the computers involved? Does it improve processing time or reduce computer resources utilized? Can you link the AI algorithm to causing physical effects or change? For example, does the invention result in control or affect operation of other physical electronic systems? Does it control a robot camera sensor? These are things that would make it more likely to be patent eligible.
Gordon: Thanks very much, indeed. That was really interesting and I think in a new area like this, where new issues are emerging all the time, the world's corps are going to be looking sideways at each other for guidance. So a comprehensive decision like that may well find itself being influential beyond the borders of Canada as well. Of course those are not the only options when it comes to IP protection. Much of what underpins AI technology is found in software and that's usually protected by copyright. So, Roch, can you talk us through the options for using copyright to protect the actual technology and the underlying data sets?
Roch: Sure. So copyright's inherently a limited type of protection. It's kind of a silly thing to say but it bears repeating. Copyright only prevents actual copying of copyrighted works. So in the AI context what's that mean? If you've implemented a trained AI model in computer code, and someone gets and copies that code, copyright can be useful. If an employee absconds with your source code you can enforce copyright against them. You can also use copyright to protect against more traditional piracy. Although I'd say that's less common now then it was some years ago with digital rights management and server side implementations. But copyright does not protect against independent creations. So if a competitor sees what you've done and tells their own engineers, go and create and me a competing product, and they do that independently of your product, copyright's useless for you regardless of what jurisdiction your in. If your employees publish papers on your AI innovation, and competitors read those papers and replicate your product from those papers, copyright's not going to help you anywhere either. You would need patents for that. In respect of training data in particular I wouldn't count on copyright to protect you. One, depending on where you are in the vertical you may not own any copyright. You may not own the data. Two, copyright may not even subsist in training data. To get copyright in data, for copyright to subsist, you generally need to show that the data was the result of skill and judgment or, depending on the jurisdiction, borderline original even and if it's mechanically created work product, which a lot of this training data is, copyright wouldn't subsist, arguably, and wouldn't help.
Gordon: Thanks very much, Roch. Very briefly, one point which is unique to the EU and, indeed post-Brexit UK, is the availability of a data base right. So, Matt, can you just briefly take us through the applicability and relevance of the data base right?
Matt: Absolutely. This is a potential opportunity for a bit of jurisdiction shopping and that's because EU law provides for a so called sui generis data base right and, actually following Brexit, there's now an equivalent right for UK companies if they develop a data base from the beginning of this year onwards. It is a neglected right, but might be worthy of re-evaluation because data base rights very quickly fell out of favour after a series of early CJ EU decisions in 2004, which cast out of what have been come to be known as spinoff data bases. That's because this is an investment protection. It applies where you have invested in either obtaining, verifying or presenting the contents of a data base. The theory was that it may not apply to businesses whose investment was not in any of those but in generating the data in the first place. The examples are about football leagues generating football listings which they were going to do in any case. Indeed, theoretically, that means it may be a life sciences company that's invested millions in a clinical trial, has generated all sorts of valuable data, wouldn't get any data base right because it's the wrong form of investment. But in the era of machine learning pre-processing of data before it's ingested into the machine learning algorithm is now critical. You need to, for example, de-SKU, de-crop, remove noise, down sample and so on and you may have to do some very clever stuff to avoid bias as well. It's specialist work; part science, part art. That really might count as a investment and verification and/or presentation. So the key takeaway, if you're operating in the UK or the EU, it's probably a good idea to record the sort of pre-processing you're doing and the costs you're expending to maximize the chance of attracting a data base right.
Gordon: Thanks, Matt. That's really interesting. Finally, in this area we've already identified China as the powerhouse of AI, so Jason, China's always been a pretty difficult place to obtain software related patents. What is different in the approach to AI? Is the examination process maybe a little more friendly now?
Jason: Sure. That was true. It used to be hard to obtain software patents here because you could only draft in method claims and means-plus-function claims for software. Assuming now that method claims are pretty hard to enforce in China, because we don't have the process of discovery and the burden of proof is on the plaintiff's side, that means that in recent years China has been adjusting the requirements. Now it is much easier to obtain a software related patent than before. Basically, except for some computer program per se, almost all subject matters have a chance to be granted in China. Speaking of AI, the authorities have also pushed the Patent Office to be more friendly for AI patents because they would like encourage innovation in this field, but the examination ... used to be very vague and in the last year the Patent Office officially made some changes to the guidelines. In the new guidelines regulations were set forth for innovations including algorithms or business rules. Now a clear process, including two steps, helps you to decide if a subject matter is eligible in China. Step one is to see if a claim is purely an abstract concept or business method without any technical feature. In other words, if there's a single technique or feature residing in the claim it should not be classified as a pure mental activity and we can move on to the next step. Step two is to see if technical features form a technical solution. That is to say, if technical features are closely related to a technical problem and result in technical effects, it's generally okay in China. For an AI invention that relies on algorithms it's very important for the data process, whether algorithm or model, to have an exact technical meaning. For example, in each classification it's typically considered as a technical solution because it's all about visual, the presentation, but if it's just data classification from pure mathematical perspective and there's no applied situation for that, it's generally considered too generic to be a technical solution. So my highlight here is to include a specific applied field when working on scope of protection. In summary, there have been several changes in China definitely in favour of AI filing and the other good news is that the Patent Office has decided to speed up examination in the near future. So this also benefits the AI protection here.
Gordon: That's great, Jason. We're going to come back to that speeding up of the process a little later on but thank you for that. So, with all that in mind, let's move from the possible to the practical. What are the best practices for business to adopt in relation to AI protection and exploitation. I'm going to turn to Matt, once again, for some scene setting. What are the issues to be considered under this heading?
Matt: Okay. So here are the headings. First, and center, trade secrets. So internally, staff policies, practical and technological measures. Externally, requirements for collaborators under contract to keep your secrets safe. Second heading is those contracts and more details. So you want to protect assets as between the parties regardless of whether they're actually recognized in traditional IP. So really it's about clarifying control of key assets such as data, derived data, models and insights. The definitions and working practices are being developed as we speak and there's a bit of bum fight every time. One solution considered de-marketing fields of use as between the parties as a solution. Third, pursue patents where appropriate and especially where trade secrets may be lost. Finally, copyright and data base rights arise automatically but make sure the ownership is clear, particularly when you're using contractors, and consider registering US copyright in source code because that gives you entitlement to costs and statutory damages without proof of loss. Back to you.
Gordon: Thanks very much, indeed. So taking all those issues in turn, can we walk through the most practical recommendations which we can give at this stage? So, Roch, what are the common roadblocks and what do businesses need to do to ensure that they maintain control of the data, which is so crucial and underpin the AI technology?
Roch: So to build a bit on what Matt said, I'd say you need legal, physical and cybersecurity measures to ensure confidentiality and protection of data. They compliment each other and lacking one can obviate the other. For example, you may have great legal protection in form of non-disclosure agreements but if people are allowed to walk in your premises unsupervised, and see what's going on and leave, that can render ineffective what you have written on paper because it shows in fact you haven't executed on those obligations. Practically, you have some kind of internal system to track innovation and disclosures. I often see it in the form of corporate intranet or Wiki and on the intranet you can have, for example, innovation disclosure forms and chat rooms to facilitate collaboration. Effectively you end up building a repository that can be periodically reviewed by, for example, an innovation committee to determine whether to patent something, whether to keep it secret. It can be useful for onboarding new employees and for retaining institutional knowledge when employees depart. It's also useful when you're trying to raise money because you can collect documents pretty quickly, generate a listing of your trade secrets for due diligence. I think it's also very useful to keep track, carefully, of who's going to disclose when. Particularly when you get near a release date and an innovation starts to be disseminated more widely, for example to sales, which very often naturally leads to interaction with the customers. Because if you want to patent you'll want to keep your technology secret until some kind of planned, in control disclosure event such as a paper publishing a conference or what have you. Regardless of whether you patent you'll probably want to keep a fair amount of your information secret. So for example, source code, you'll probably want to keep secret unless you're contributing to some kind of open source repository. The training data, we've heard how valuable that is. You want to keep that secret. I'll note here, from a practical perspective, it's also a difference in patent law in terms of what is a public disclosure, what constitutes a public disclosure, between the US and generally speaking the rest of the world. Outside the US the test is usually phrased as what's made available to the public. So for example, if you sell a product and the invention can be reversed engineered from what you publicly sold, that constitutes a public disclosure. But in the US the net is cast wider. For example, an offer to sell can constitute a disclosure. Generally speaking, commercialization, even secret sales, can count as disclosure. There's a one year grace period in the US that can be helpful here if you do self-disclose but that grace period doesn't exist in a lot of other countries, including Europe and China, so if disclosure is a concern that's something I'd certainly recommend seeking advice on.
Gordon: Thanks very much, Roch, and yes that can be very dangerous, as well as helpful, that grace period. Now, Shahrzad, from your experience what recommendations would you make to a new business starting up in the field of AI?
Shahrzad: Thanks, Gordon. I'd like to build upon where Roch and Matt left off. I think they both gave some great advice already. I would say to start off with decide how you want to protect your AI invention with considerations to trade secrets, patents, or both. For example, patents gives the patent holder right to exclude others from making, selling, using or importing the invention. Here when thinking about where to file you may want to consider countries who have interest in AI. The chart I showed at the beginning showed some of those jurisdictions. China, US, Japan, UK, Canada, those may be some options. The second question that I'm often asked is can I prepare the patent application myself? My advice is a strong no. The patent agents know the intricacies of the various legislative and judicial systems and will know how you need to frame your invention for sufficient disclosure and patent eligibility considerations, among other things. The third is to think about patent eligibility. Clearly this is top of mind for many folks here. Think about does your AI invention tie to physical and tangible existence? Jason gave a great example. Think about does it have a technical application? For example, this may be how does your AI invention improve the functionality of an existing computer system? Does it improve speed or reduce resources? In most jurisdictions AI systems used to control other physical devices are looked at favourably. For example, an AI algorithm which recognizes audio and activates car components in return will be more likely to be patent eligible. The fourth is to know your competitive landscape and, as I showed at the beginning, Canadians, for example, are filing applications relating to natural language processing, computer vision, robotics, knowledge representation and reasoning, prediction, speech processing. Know what your competitors are doing. Finally, I would say plan an IP strategy. As Roch mentioned, for example, if you're going to patent you don't want to disclose. You need to be planning in advance as to what you're going to do.
Gordon: Thanks very much, indeed. Now, back to Jason. I said we'd return to the question of expedited examination. You mentioned that AI patents may be a possibility for expedited processing in China. Will that be available for foreign applicants as well and what should they do? And while we're on it, what are your recommendations for businesses outside China which may want to ensure their protection for AI technology within China?
Jason: Sure, Gordon, like I said the Chinese Patent Office decide to substantially speed up the prosecution in an official document released just earlier this month to set a goal for improving examination quality and efficiency and ask that the whole process of examination should be lowered to 18.5 months, and even to 14 months for those of high value, and examination for trademarks should be lowered to 4 months. Rather than ask what kind of patents are considered as of high value, I agree that term is quite subjective, but from my experience a patent for newer technology like AI, or other software related patents, are very likely to be considered as valuable by the Chinese authorities. In addition, the change is universal so it's not only benefit local companies but also applications from abroad. On this last thing this is very ambitious because currently the average time needed for patent examination is about 2 to 3 years. I've even talked to some Chinese examiners. Although they're apparently not very happy about extra work load, they did say the Office is very determined to push it forward. So let's just wait and see. As the Chinese authorities are now welcoming patents, and they're determined to speed up the prosecution, foreign companies may even take advantage of such a trend. For example, if it's your plan to file patent in China anyway, you may consider doing that earlier rather than later because if the Chinese patent's quickly allowed you're in a very good position to speed up prosecution in other jurisdictions. For example, Patent Prosecution Highway, or PPH. As a patent agent I can say that if you have similar technologies and want to file patent in China, generally you don't need to worry too much about the eligibility or patentability of your tech. There must be a way to get over the examination here but there's a big general concern you'll want to avoid in China which is grace period for novelty. Like Roch emphasized, grace period exists in US and also in Canada. So before filing a patent some may prefer to publish their AI research while still able to obtain a patent later. But unfortunately that's not the case in China. The grace period for novelty is practically very hard to obtain in China because it limits the publication to few domestic journals in just 6 months. So if you plan to file patent in China, at some point, please do not disclose your technology before first filing or it will be very, very hard to get a patent here.
Gordon: Thanks very much, Jason, and that's the bear trap I referred to earlier about the good news and the bad news about the grace period. What I want to do now is I'm going to ask the panel to really speed up on this one because we're running low on time but I just want to throw one question at my own whim up and all ready and see how we're going. So, Matt, what's the most important IP lesson you learned from editing the Law of Artificial Intelligence?
Matt: Right. Editing a 600 page book, one lesson. Okay, I think it is IP strategy has to be holistic. You need to have an understanding of the technology, the market, interactions of other areas of law. So privacy, competition law and you need to know when it's of direction via involving policy and regulatory landscape.
Gordon: Thanks very much, indeed. That's great. Shahrzad, how does the Canadian legislation and jurisprudence deal with the question of AI inventorship? Bear in mind here that this may have application way outside Canada. This is being looked at all over the place.
Shahrzad: Yeah, absolutely. I think that there are two view points here. The Patent Act and the legislation in Canada. The Patent Act has wording to the effect of the commissioner shall grant a patent for an invention to the inventor or the inventor's legal representative. Based on this language the AI is not expressly prohibited as an inventor as a patent can be granted to an inventor or to an inventor's legal representative. There's no mention that either the inventor or the representative must be a human being. Now, turning to the jurisprudence, I believe that's less favourable to AI as an inventor. The Supreme Court of Canada in Apotex tells us that the inventor is the person or persons who conceived of the invention. That being said the Court also stated that the ultimate question is who is responsible for the inventive concept. This might still leave the door open for an argument that where an AI program is responsible for the inventive, the AI might be considered the inventor for the purposes of the Patent Act. So, it's an open debate.
Gordon: Thanks very much and obviously we've seen a little bit of that in the UK and the Davis case and indeed in the rest of Europe as well. So it's an interesting topic of the moment. Roch, what do you tell clients when they ask how to deal with the uncertainty in the patent systems as it relates to changing patent eligibility rules?
Roch: Well, this goes back to the poll question when patentability was ranked the primary concern of the attendees. I just say not to treat or overreact to the most recent decision of practice guidance, or treat it as gospel, and take a bit of a longer run view. In the US I mentioned we had the Alice decision in 2014. Then we saw the Federal circuit start to carve out exceptions. Then we saw the US PTO issue new guidance in 2019. Shahrzad's already mentioned the Choueifaty case in Canada that came out just last year. Maybe 2 years ago? I've lost all track of time. That's already being challenged. The practice guidance the Office issues is already being challenged in court. Various jurisdictions and courts, they have trouble dealing with this issue. You get something akin to a pendulum swinging back and forth. I happen to think the pendulum is swinging in favour of patentees right now but I don't overreact to that. I mean we can talk about that more in the breakout room if anybody would like.
Gordon: So take the long game is your advice, generally. Jason, aside from China, what trends are we seeing in relation to AI protection in other countries in South and East Asia?
Jason: Sure, Gordon. For East Asia, Japan is the sort of biggest country in terms of AI filing and it is no secret that they are doing very well in harbouring development relating to AI. Japan is very forgiving for software related patents. Many subject matters are permitted in Japan such as medium computer program, data structure, even machine learning model, etcetera. The only requirement seems to be that the information to be processed has to be done by a computer. Speaking of South Asia, ASEAN has to be mentioned. ASEAN stands for Association for Southeast Asian Nations. That includes ten States like Singapore, Malaysia, Vietnam and Thailand. As well as the most ambitious State in high tech, Singapore has applied technologies from many sectors, especially for Telecom and FinTech. Singapore Patent Office has also supplied the patent prosecution specifically for AI technologies. If I remember correctly the Singapore Patent Office plans to shorten the prosecution for AI patents from over 2 years to less than 1 year. Since the ASEAN patent examination incorporation a quick patent prosecution process in Singapore can be very helpful if you want to file IP in other ASEAN State. The corporation is basically a share patent search in the examination scheme for all the member States, so if you have a positive examination results in one State, it will help the corresponding application in another State. Of course, I'm not an expert in Singapore laws but we do have a Singapore office that handles various IP registrations. If you're interested in that jurisdiction you can contact me or any of our colleagues so you'll find the right person to start with.
Gordon: Thanks very much, Jason, and you took the words out of my mouth about Singapore which was always, I suppose, going to be at the forefront of this kind of thing given that the intention, the plan of the Singapore Government to create a knowledge based economy in the long run. Let's aim for the final poll. Other than intellectual property what is your biggest legal concern for AI? Is it privacy? Is it potential liability? Is it competition law? It is that there's too much regulation coming too soon? Or is that there's not enough regulation and it's coming too late? Or anything else. Very interesting. Nobody seems to think we are over-regulated too soon but quite a few people think we're under-regulated and it's coming too late. Privacy I can see there is right at the top of the list and that really reflects on a lot of what we've been hearing. Of course liability is always going to be a concern particularly when you are relying on what a machine has done. So, interesting and possibly predictable but nonetheless interesting answers and thank you very much to the audience for participating in the polling today.
Now, we are just at the end of our time. So there have been some questions asked but I can suggest that people take those questions offline and move into the chat rooms now where they can ask those questions. Also, we will make a note of them and we will produce something that deals with those questions. I think one or two speakers may have been typing answers while we've been going along anyway which is extremely helpful. Thank you very much, indeed.
Thank you to everyone who's attended today and I emphasize everyone. I think we peaked at 275 attendees at one point. I hope that you've enjoyed this session and learned some very useful things along the way. We will, of course, be returning to this topic repeatedly in this webinar series in the years to come. The full webinar will be posted as a recording on our website in the coming days, if you would like to watch it again, or if there are people in your organizations who you think might benefit from the content. So it will be available and uploaded very soon.
Now, if you want to learn more, please get ready to head into one of our chat rooms where our panel will be ready to answer follow up questions and to continue the discussion with you generally. If you registered in advance you will be guided to the room of your choice now. If you registered today then just wait and one of our team will direct you.
For now, thank you so much to all of you for your attendance, and attention today, and thank you to our panel who have put a lot of effort and time into thinking about how best to structure these answers for you. I look forward to seeing many of you again as our webinar series, The Life Cycle of a Smart Idea, continues. We're not switching this off at the moment. We'll leave it on while people gravitate towards the chat rooms. For the purposes of the main session that's an end to it. Thank you very much indeed for coming along today.