Beyond the hype: a practical guide to navigating AI for intellectual property

Recent advances in artificial intelligence (AI) have spurred a boom in press coverage and marketing but there remains much confusion surrounding the technology. For IP counsel this confusion is hampering efforts to evaluate the impact of AI on day-to-day work and the bigger impact on organisations in general. AI is both overhyped in the short term and underestimated in the long term. For example, two decades ago IBM’s Deep Blue shocked the world by besting Gary Kasparov in chess. However, this watershed moment did not spark an immediate explosion in algorithmic computing; instead it was gradually and quietly adopted into commercial products. Today we are surrounded every day by algorithms: recommendations for new products, songs and shows; personalised advertising; GPS navigation that can take you anywhere. Patent search and analytics software uses algorithms to score patents and create visualisations. This latest wave of AI is coming from deep learning and machine learning. According to Gartner’s 2017 hype cycle these are at the peak of inflated expectations, but that does not mean that they should be dismissed. Gartner also estimated mainstream adoption to be between two and five years away – and that was in July 2017. So are we on the verge of the robot revolution?

While you may not see robot colleagues walking around the office, the consequences of the AI revolution will still be profound and far-reaching. The rate of change from AI is much faster than previous disruptions – according to one study 10 times faster and at 300 times the scale (see www.mckinsey.com/business-func...). Most deep learning work to date has been research as opposed to commercialised products but the technology is maturing quickly and we will likely see many deep learning products come to market in 2018.

When trying to predict the future, it is helpful to understand the past. As a field, AI began in the 1950s and has since gone through alternating periods of intense excitement and disappointment. As a technical term, ‘AI’ is vague and can cover many things, but as a marketing term it conveys a sales message more compactly and excitingly than a complex technical explanation full of caveats. Unfortunately, this can lead to confusion and misconceptions. Generally AI is used today by the mainstream media to describe breakthroughs attributed to an approach called ‘deep learning’, which is actually a sub-set of machine learning – one means of trying to build AI. The field focuses on how to build machines that learn from data in order to get better at a task over time. Humans are not born with much knowledge but we do have the capacity to learn. Machine learning asks the question – can a machine be given the capacity to learn and what does it learn from?

How deep learning is revolutionising AI

Deep learning is all about data and that is revolutionising the impact that AI is having on many knowledge industries. Despite how powerful existing data science already was, we were only beginning to scratch the surface when deep learning began smashing records across fields – from computer vision to machine translation. While most commercially available AI solutions do not actually use deep learning, once these tools become available the impact will be both fast and widespread. For those interested there are plenty of resources to discover more about how deep learning works, but to understand the impact of AI on IP work we will take a more practical perspective. Why is it so important to recognise that deep learning is all about data?

More traditional AI – let us call them algorithms – are driven by experts not data. In other words, the development cost and value come from engineers analysing a problem space and finding the algorithm that works best. The hard-coded logic is effectively just a formula or series of rules built through a combination of expert knowledge and trial and error (costly and difficult to adapt). In patent valuation and scoring, for example, this has led to a black-box dilemma where practitioners want to understand the basis for software-based patent scores but the software companies cannot risk disclosing their hard-earned algorithms. Another example of legacy technology is semantic search and patent visualisation software. In Japan this technology was met with great excitement in the early 2010s and many professional patent searchers wondered: “Will semantic search eliminate our jobs?” Although the technology can perform well in some isolated cases, usually it still requires significant additional expert labour to deliver meaningful results, whether you are looking at a legal case such as invalidity or a more complicated case such as data collection for analysis. Since technology and patent language is abstract and can be interpreted in many ways, results tend to be noisy and unreliable. Why could deep learning be any better?

Figure 1. Hype cycle for emerging technologies, 2017

* Platform as a service
** Unmanned aerial vehicles
Source: Gartner (July 2017)

Even for an expert it is hard to explain how language works and even harder to force that into code that a computer understands. Deep learning gets around this by skipping the expert and going straight to data. Access to data, its quality and the cost of collecting and organising it are the new cost drivers instead of engineer time. Instead of an expert choosing the best formula, deep learning iterates through the data to find the best formula. Fundamentally the concept is similar to linear regression, which is taught in high-school math class. You can find a line of best fit from a set of data points. The broad term for the field deep learning comes from is neural networks. These are so named because the technology attempts to mimic the structure of the human brain, which is made of layers of neurons that send signals to each other. The term ‘deep learning’ comes from increasing the number of layers in the machine’s neural network to find more complex ways to extract information from data.

Data is not only the driver of cost but also the determinant of quality. The development challenges stem primarily from learning what types of models will accurately reflect the domain to which they are being applied. Of course the other big challenge is that the engineering process for building deep-learning systems is still poorly understood and remains more of an art than a science. The talent pool capable of doing a good job in designing such systems is still limited and the development process is unpredictable. However, the results so far are powerful. Researchers and data scientists are often surprised at just how effective deep learning models are. However, despite the excitement this is not a panacea.

What deep learning cannot do

The easiest way to understand deep learning is to understand what it cannot do. Despite the huge difference in performance, deep learning is just one tool in a data scientist’s toolbox and it has limitations. Understanding what deep learning cannot currently do will help you to develop an intuition for what it can do:

• Deep learning requires lots of data to train on. If data is limited, traditional approaches actually perform better.
• Learning only applies to the task learned or a very similar task. We can build machines that perform specific tasks better, faster and cheaper than humans but we still need to build and train a new AI for each task.

Provided that tasks are structured and routine, deep learning results will be far better than traditional approaches. Deep learning has already led to brand new opportunities in fields such as self-driving cars and natural language voice interfaces for consumer electronics. For IP counsel the impact will be felt in the automation of labour-intensive knowledge work.

Figure 2. Evolution of artificial intelligence

When do the robot lawyers show up?

Even before the current advances in deep learning, machine learning and data science in general have already been reshaping industries. In the legal industry, for example, e-discovery technology has transformed how courts and attorneys think about discovery. Knowledge work automation is estimated as being worth between $5 trillion and $7 trillion by the McKinsey Global Institute, which ranks it as the second largest disruptive opportunity behind mobile internet (see www.mckinsey.com/assets/dotcom...).

E-discovery is an example of software automation that has already found widespread application in the legal industry. In 2011, the New York Times reported that “thanks to advances in artificial intelligence” new software could analyse 1.5 million documents for less than $100,000 as compared to a 1978 case that required $2.2 million to examine 6 million documents. Perhaps more important than the cost savings was the finding that the software actually did a better job – “human colleagues had been only 60 percent accurate”. The title of the article was: “Armies of expensive lawyers, replaced by cheaper software”. It is little wonder that today e-discovery is commonplace and we can see how this automation has changed the market. However, will this replace the job of the attorney?

As the New York Times article suggests, one would expect software to put lawyers or at least legal clerks and paralegals out of work. A 2016 Deloitte Insight report (see www.legaltechnology.com/latest...) predicted that 39% of legal jobs could become automated. However, the effects so far have been the opposite. By reducing the cost of discovery, automation has increased demand, which has led to more judges willing to allow discovery – this resulted in a 1.1% annual increase in the number of legal clerks in the United States between 2000 and 2013. Table 1 shows similar examples throughout history, from the automation of weavers to bank tellers to legal clerks.

As these examples show, the effects of AI are nuanced and sometimes counter-intuitive. Although e-discovery came well before the game-changing advances of deep learning, the impact is similar in that it is tasks – not professions – that are suitable for automation. In many cases software will not only reduce the cost and time of these tasks but will provide better quality. The conversation should focus not on “will AI take my job?” but rather on “what parts of my job can I outsource to AI so that I can create more value?”

Despite general talk about automating legal work, lawyers actually rank low in a study on the future of employment:

For example, we find that paralegals and legal assistants – for which computers already substitute – are in the high risk category. At the same time, lawyers, which rely on labour input from legal assistants, are in the low risk category. Thus, for the work of lawyers to be fully automated, engineering bottlenecks to creative and social intelligence will need to be overcome, implying that the computerisation of legal research will complement the work of lawyers in the medium term. (see www.oxfordmartin.ox.ac.uk/down...)

In fact lawyers were ranked with only a 3.5% chance of becoming automated compared to 170 occupations with greater than 90% probability. Other low-risk occupations include purchasing managers, operations research analysts, political scientists and credit counsellors. High-risk occupations included abstractors and searchers, tax preparers, legal secretaries and bookkeepers. These professions appear to fit into two categories: one profession that routinely processes and prepares information and a second that interprets information in a variety of circumstances. What does this mean for the landscape of legal professions?

Paralegals and legal secretaries are at high risk because their work is routine and highly structured. The same can be said for some of the work that attorneys carry out. In fact, McKinsey estimates that 23% of a lawyer’s job can be automated. Often this type of work is disproportionately represented in the work of junior attorneys and associates. What happens when work becomes increasingly automated? IP counsel already cite lack of talent as a major problem. In the future where will new talent cut its teeth?

Jobs do not get automated, tasks do. While lawyers themselves are not likely to be automated as a profession, the implications of automation will greatly change the environment around them. In intellectual property, routine tasks such as patent administration work, prior art search, initial drafting, portfolio management and reporting will become increasingly automated. The good news is that this may help to alleviate the pain from lack of qualified talent. Patent information will become more readily available and easier to access. That begs the question: how do we evaluate and use intelligent tools and where do we allocate human resources?

Since the industrial revolution we have seen machines and later computers continue to become better and better at work that humans used to do. However, the impact of automation on jobs is not straightforward. As the new technology dramatically improves, new skills are required to meet changing market demand. The result is often an increase in overall jobs along with a shift in the nature of the work encompassed by these jobs. Like the invention of the steam engine and the internet, AI is a fundamental technology that has far-reaching implications. The organisations that are able to restructure themselves to adopt this new technology will reap huge benefits.

Why patents need deep learning

The potential impact of AI on patents is profound and fundamental. It is worthwhile remembering that the patent system was developed to promote innovation. To reward progress, inventors are granted a temporary legal monopoly; in exchange they have to make the information about their invention public to promote further innovation. This framework underpinned much of the technological progress of the industrial revolution. For the first time in history, the entirety of human technological progress was indexed and available to the public. That is a pretty modern view of the value of information. However, this system is no longer functioning as intended. Patents are the source of a lot of frustration – over-priced solutions that under-deliver uneven access to information, while bad actors take advantage of the system. Despite more patents filed and granted than any time before in history, the information is less accessible than ever.

While patent data has grown exponentially in the past few decades, workflows have hardly changed over the past century. Arguably the biggest change may be the digitisation of patents and the introduction of keyword searches in the 1990s. However, this has actually exacerbated the situation. The processes required to file, examine, litigate and enforce are all manual and can only scale in a linear way. Over the past two decades the mismatch in supply and demand has led to a tipping point.

Globalisation combined with digital patent information raised the bar for diligence. The additional risk led corporations to patent more aggressively. Patent offices under pressure to examine rapidly growing numbers of applications naturally used keyword search and databases to increase efficiency. Where examiners used to have time to develop a robust understanding of their field and the invention at hand, now technologies bleed across domains and there is more information than a single human can effectively remember. The increased use and reliance on keywords and phrase matching is an incentive to draft applications with increasingly confusing language, which only increases the overall burden on the system.

Deep learning works with data at truly massive scale to extract features and patterns. This approach is able to capture much more complexity than those used in e-discovery. The result will be tools that eliminate weeks of prior art searching and reduce office action response cycles. Humans can review only a limited number of documents but an AI system can review every patent in existence.

Figure 3. Growing US filing gap between applications and grants

Prior art searches are one example of a structured and routine task. We built an AI system to find the most relevant patents given an input patent and tested this against patents that have had art successfully cited against them accordingly to data collected from the US Patent and Trademark Office Patent Trial and Appeal Board (PTAB). We checked to see whether our AI was finding art successfully cited in an appeals process. Note that we did not give the system the PTAB data to train on – this experiment was designed to test how well a deep-learning based system could identify relevant documents in a complex situation such as patent validity. The AI seemed to find cited art at least as reliably as a professional searcher. More importantly, this only takes seconds for the AI while a typical prior art search takes days or weeks.

This example raises an important point when considering AI solutions. Deep learning is based on probabilities. Unlike legacy systems where you specify an input and get back an output, a deep-learning based system is giving you results with varying degrees of certainty. In the example above we could easily cherry-pick a perfect example where our AI found relevant art as the number one result. Conversely, we could show an example where our AI did not find the best art at all. So when considering deep-learning solutions it is important to look at the statistical performance over a meaningful number of examples – and think through what kind of human labour that will require.

This is just one example – there are countless other areas where deep learning will dramatically reduce the cost of routine patent work. With increasing trends in data-driven decision making and cost pressure on IP departments, organisations will be forced to seek new solutions. The IP function in organisations will begin to shift from being a cost centre to a value centre as admin work becomes automated and the focus shifts towards making data-driven strategic decisions based on patent information. Senior IP counsel and management will face challenges in how to incorporate AI into the organisation. Will AI necessitate a shift away from a billable-hours culture in organisations’ legal departments? As information becomes cheaper and more widely available, organisations will be under increased pressure to strategically manage ballooning patent portfolios.