The Chief Justice of the Federal Court of Australia observed:³

AI is currently receiving great attention, including in areas such as the law.⁴ It is part of the fascination with and implementation of the technology age, being an era of unparalleled interconnectedness and resulting efficiencies — within a global population of 7.78 billion people, there were, as at October 2019, 5.15 billion mobile phones, 4.48 billion internet users, 3.7 billion social media users, 2.5 billion email users, and in 2021 there are 2.8 billion Facebook users.⁵ These are likely to all join up as the “internet of things” (often referred to as “IoT”) in the not too distant future.

In Australia, the Commonwealth Government’s national science agency, CSIRO, has developed an AI Action Plan, which includes establishing the National AI Centre and four AI and Digital Capability Centres. CSIRO is home to one of the world’s largest AI applied capabilities, with more than one thousand researchers.

This article outlines the possible impact of AI⁶ on the future work of the key players in litigation — the courts, solicitors and barristers. It does not address the now well-accepted technologies facilitating filing of court documents electronically and general digital case management within courts.⁷ The article addresses the question of whether AI might be (to use jargon) a “disrupter” of legal services or an “enabler” or something between the two, and if so in what areas. It is intended to assist with an understanding of AI by way of overview,⁸ with particular emphasis on litigation. It is necessarily general in some respects as it attempts to predict the future in the very uncertain environment of software development. In this regard, the article is exploratory in nature and intended to provide food for thought about AI’s potential role in litigation.

Preliminary conclusions about the future are outlined below, which in turn may inform the possible future of litigation more generally.

It is impossible to address the future of AI in the legal industry or elsewhere without first discussing two concepts in detail — “intelligence” and “work”. These are then applied to the relevant tasks under consideration, each discussed in turn.

For the purposes of this article, and to assist the discussion, intelligence has been defined in these terms: the capability to perform work to achieve the relevant task.¹⁰ Determining the pathway may involve abstract reasoning, the application of logic, testing, problem solving, induction, deduction, the making of inferences, and the like. To understand and apply this definition, it is necessary to be clear on what “work” is within the above definition.¹¹ “Work” for these purposes is the determination of a pathway to achieve the relevant task. Determining the pathway may involve abstract reasoning, the application of logic, testing, problem solving, induction, deduction, the making of inferences, and the like.

Apart from “work” being a place, it can be said for present purposes that there are essentially two types of work (which are often conflated):

The first type of work is in broad terms “human” work, and the second is in broad terms “computer (AI)” work. These two types of work are explored in more detail below.

It is widely predicted that at some point in the future, human and artificial intelligence might eventually converge. This has been contemplated for some time on the basis that the human brain is (merely) an organic algorithm.²⁵ If this occurred, perhaps all work (and therefore tasks) humans can do could also be done by computers — including that done by courts, solicitors and counsel. This section explores the possibility of convergence.

The answer to the potential for convergence lies with, at the very least, four main factors.²⁶

The first factor is in the incredible increase in computational power. It is said that the human brain can be reduced to a series of processes that are chemical and electrical. If this is correct, there exists the possibility that we become able to expose the processes within the brain and therefore the reasoning currently sitting under the “intuitive” or “creative” judgments made by humans.²⁷ As predicated by Moore’s Law,²⁸ the fourth generation iPad now has more computing power than the most powerful super-computer in the world some 30 years ago, the Cray-2.²⁹ We are, apparently, nowhere near the limit of computing power. It has been estimated that the limit is 33 orders of magnitude (10³³ times) beyond today’s state of the art for how much computing a clump of matter can do; even if computer power doubled every couple of years, it would take over two centuries until the frontier was reached.³⁰ It is possible that in the foreseeable future, computing power may extend significantly into the mental processing power of the human brain. This coupled with “machine learning” (see below) may see computing power reach capabilities once thought impossible (consider the changes arising from the introduction of the iPhone alone). It has been estimated that the computational capacity of the human brain is roughly the power of, what is today, an optimised $1,000 computer.³¹ More recently, it has been estimated that by 2020, the average desktop computer will have the same processing power as the human brain, with approximately 1016 calculations possible per second.³²

Associated with this area of exponential growth in computing power is the second main factor — the development of so-called “machine learning”, using new neural network techniques applied to vast amounts of data. Machine learning is computation using algorithms that improve through experience. The “learning” occurs because each relevant neuron updates its state at regular time steps by averaging together the inputs from all connected neurons, weighting them, optionally adding a constant, and then adding what is known as an “activation function” to the result to compute the next state.³³ In this way, the machine “forward-predicts” the next better step. Machine learning (or deep learning) is responsible for some incredible results: from a computer giving an accurate caption of a photo even though it does not know what the photo represents;³⁴ to mastering dozens of computer games without instructions;³⁵ to passing Tokyo university entrance exams (which included maths, language and general knowledge questions); to beating the best Go player (Lee Sedol) in the world in 2016.³⁶ However, deep learning cannot (yet) do something that is fundamentally important — explain its reasoning and findings.³⁷

The third main factor is the incredible reduction in cost and size of memory capacity for computers. Memory capacity has been improved dramatically over time, without requiring any software changes. Hard drives have reduced in cost by about 100 million times, and the faster memories useful for computation rather than mere storage have become 10 trillion times cheaper.³⁸ The human brain stores around 10 gigabytes electrically and around 100 terabytes chemically/biologically. The world’s best computers can now out-remember any biological system — for a few thousand dollars.³⁹ Overall, the cost of computation halves every couple of years, with the effect that over the last century or so, the cost of computing has dropped about one million million million times (10¹⁸).⁴⁰ This fact alone puts into challenge the use of labour when performing certain tasks. Computers are no longer building-sized; they are in our pockets or on our wrists.

The fourth main factor is the amount of available data, which is growing exponentially, and its collection is becoming more systemic and collaborative across certain professions, industry and government. The insurance industry, for example, has huge collaborative networks and systems designed to capture large volumes of data that can be mined by AI, for the benefit of the participants of the networks/systems. It is important to note, however, in the context of the legal industry specifically, that sophisticated AI algorithms and learning alone cannot provide valuable solutions; there has to be sufficient data upon which the algorithms can be put to analytical work.⁴¹ It is true that sophisticated algorithms can sometimes overcome limited data; but even if there is a significant body of data available for analysis, this must be viewed cautiously because it may have inherent biases. For example, negative data is almost never published.⁴² Further, commercial repositories of legal texts may be big enough to be useful for statistically learning contextual information about legal terms and phrases that can then be useful semantic information for other machine-learning applications. However, generally, law does not seem to present opportunities for big data.

In some industries, significant data-gathering or consolidation occurs, permitting better AI analysis. An example is in the insurance, banking and finance, and medical industries. This gathering and co-operation between entities raises significant legal issues regarding privacy and data integrity. Arguably, it is the capacity for a particular industry or profession to gather data in large amounts that will determine its ability to exploit AI in the foreseeable future. The computing power will come — but aggregation of data in certain areas may not. We know already that aggregation of data in, say, the medical fields is progressing well, and aggregation of data in litigation is even more so given its specific fact-dependent nature.

There are several well-recognised categories of tasks where AI is already well established.

The first category is image recognition or so-called “computer vision” — the ability of the computer to know what it is looking at. AI is being used, for example, to sort vegetables by recognising their sizes and varieties. It is used in face recognition, for example, to access the new iPhone X. And in China, AI is used extensively for face recognition. For example, an insurance company there,⁴³ with a market cap of $120 billion, is using AI to provide online loans quickly, verifying more than 300 million faces in various applications, complementing other cognitive AI capabilities including voice recognition.⁴⁴ That company alone has some 110 data scientists. Further, also in China, where there are tens of millions of cameras, it is used to penalise people for traffic infringements.

Another example of AI is in the use of “natural language” algorithms. Applications here include analysing whether customers are happy or discontented when making inquiries of companies, filtering inquiries prior to having to speak to a real person (such as when you call Telstra), and even attempting to determine (in the United States) whether 911 callers are witnesses to crimes or the criminals themselves (calling as “witnesses”). Natural language AI is also used in Siri and Google Home applications when you want quick answers from your device. The AI can get to know a person and does not rely on independent, full-premised questions; it can take premises from the preceding question, just as a human does.

Further examples of the use of AI include:

With all these examples, it would appear that AI already has broad application across industry. It should be noted, however, that one of the biggest AI providers in the world, IBM, with its super-computer AI product, Watson, lists the following industry solutions:⁴⁹

It is notable that “legal services” is not listed. However, IBM has built ROSS (AI) based on Watson, the cognitive computing platform. ROSS is a legal research tool that enables law firms to slash the time spent on research, while improving results (discussed further below).

So far this article has focused on the work a human or computer might perform in order to achieve a task. However, even in circumstances where yet more work and tasks can be performed by AI (as it comes closer to “human intelligence”), the question arises whether there are some types of work that AI cannot or for that matter should not do.

The current computer programs developed to assist the legal industry can answer legal questions in a superficial sense, but the underlying AI cannot explain the answers. However, computational models of legal reasoning are currently being developed, with such models in some cases being able to generate arguments for and against particular outcomes in problems. In particular, the development of AI to assist with legal arguments, reasoning and to predict litigation outcomes is now a well-developed field, and is continuing to develop rapidly.⁵³ These computational models may also attempt to breakdown complex human intellectual tasks, such as estimating settlement values in disputes, or analysing offer and acceptance problems, to provide insight to legal practitioners.⁵⁴ The answers given are not philosophical but scientific in nature, in that the computer programs evaluate tasks and outcomes according to data.

While this area of AI has made progress, there is currently a bottleneck in further progress toward contributing to legal practice. To date, the substantive knowledge on legal matters deployed by currently developed computational models have been extracted manually from legal sources (cases, statutes, contracts and other texts). That is, currently, legal practitioners have to read the legal texts and provide their content (in a form the models can apply). The inability to automatically connect the computational models (for legal reasoning) directly to legal texts has limited the ability of programmers to apply programs in real-world legal information retrieval, prediction and decision-making.⁵⁵

In broad overview, these developments in legal AI, involving question and answering techniques, knowledge extraction from text and “argument mining” are termed “text analytics” or “legal analytics”. Such analytic techniques in the future might overcome the bottleneck if the relevant computational program can be developed so as not to rely solely on manual techniques (by humans) to input what legal texts mean in ways programs can use, but rather so that knowledge can be input automatically. If this occurred, AI could, potentially, link these text-analysis tools to computational legal reasoning and legal analysis algorithms, to produce a wholly AI-derived legal solution. This has not happened yet; however, the amount of work being done in this area should not be underestimated — legal analytics is a well progressed field.⁵⁶

According to some, the legal industry is at the “cusp” of what is shaping up to be an industry-changing revolution.⁵⁷ The new analytics is said to enable a legal practitioner to “mine” massive quantities of data that in earlier times would simply have been impossible,⁵⁸ aiding, among other things, predictive analytics. According to LexisNexis:⁵⁹

However, this kind of analysis might be better characterised as risk-assessment analysis, not strict legal analysis of relevant issues in a proceeding. In this sense, so-called “legal analytics”⁶⁰ is more of a business tool (eg for strategising a case and making predictions).

With these constructs of intelligence, work and tasks in mind, and an overview of the development and application of AI to date, one now has a framework in which to assess the possible impact of AI on the particular work of the courts and participants in litigation. Court systems are often characterised, fairly or otherwise, by lengthy delays, high costs and occasional injustice. The question is whether and/or how AI can make courts more efficient and just (including by making more consistent decisions). To do this, it is necessary to outline the work of those participants in the justice system and to understand the analytical tools/processes deployed in order to perform that work.

As can be seen from the above, AI appears for the foreseeable future to be unable to do the work of judges, particularly at trial level in superior courts, unless the processes and methodologies of the law in relation to the quelling of disputes fundamentally change. As former Chief Justice of the High Court of Australia, Murray Gleeson AC, has observed, it is for the parliaments to decide what controversies are justiciable, and to create, and where appropriate to limit, the facilities for the resolution of justiciable controversies.¹⁰⁴ As his Honour has pointed out, parliaments regularly expand and contract the subjects of justiciable controversy.¹⁰⁵

There would seem to be many circumstances where AI might step in and take the place of courts where courts make explicit yet very discrete discretionary judgments, for example: whether to order an injunction; whether a court case should be moved to another jurisdiction; whether to grant security for costs; whether a party has waived or abandoned its right to arbitration; and whether to award costs to a particular party. In virtually all of these cases, there is a specific body of principles, guidelines or factors that have been developed by earlier decisions.¹⁰⁶ These “markers” or (to use the phrase adopted above) “recognition points” of what direction the court in question should head in and conclude on assist the court to determine what is “fair” having regarding to earlier thinking on the issue in question. This process might be said to be approaching the “recognition points” in the algorithms that were mentioned above and therefore, for these particular types of cases, these processes may be capable of being reduced to an algorithm(s). For example, whether a proceeding should be transferred to another jurisdiction might be determined by the following factors: the cost to each party of staying or going; the place where the principal activities involving the issue took place; and the law governing the relevant issue. These could be put into a relatively simple algorithm, with weighting given to the factors, and with AI producing a determination. There would be no need for a hearing. The algorithm could be programmed to look for like cases, and to distinguish other non-like cases.

Another example might be in the area of waiver in relation to arbitral rights under contract. The algorithm could determine whether the minimum number of judicial steps have been taken by the plaintiff to conclude that there has been a waiver of the right.

While this may sound attractive, experience shows that such limited discretion cases are often more complex than the above might suggest, and may involve factors that are not within the dedicated algorithm. There may also be a larger context to the question under consideration. Human judgment therefore might be necessary in order to derive a fair outcome. Further, in many jurisdictions, interlocutory decisions are able, with special leave, to be appealed. This “check” on such decisions ensures fairness for outlier cases. Perhaps the check could remain; an appeal could be available from an AI decision to a human judicial officer if certain threshold criteria were satisfied or not (as the case may be). Overall, it is possible that one day there may be a whole range of areas of litigation where interlocutory applications are dealt with solely by algorithms (AI) rather than at hearings.

Similarly, it would no doubt be possible to set penalties as derived by AI from an algorithm with known inputs and assessments. Again, however, setting penalties is not usually done solely by reference to the closed universe of evidence before the court in the particular case and judges no doubt draw on their observations and subjective judgment as to what is appropriate. If AI were to undertake such a task, using both “internal” and “external” evidence, in the case of the external evidence AI would need to mine data externally available to it as to the appropriate penalty to set. This appears at present to be unworkable.

Let’s take another example. In many if not all proceedings, the court decides whether and when the parties should go to mediation. The courts have, with the increases in costs of the courts, and delays, increasingly engaged mediation as a “pre-litigation” step in order to achieve fair and speedy outcomes at acceptable costs. It might be possible to have AI determine, based on a set of criteria, which cases should be the subject of compulsory mediation, even if one party does not agree to one. This might be based on a mining of a database that shows that particular kinds of cases have in the past settled at mediation in say 95% of cases. It might also be possible for AI to act as the mediator, based on the suggestions put to it by the parties, and suggest a possible settlement range. However, even if AI could be put to this task, mediations themselves are necessarily humanistic matters involving highly subjective judgments about what the outcome of a proceeding might or should be, the ancillary risks that might pose, and what can be afforded. There is also a “human dimension” of work in mediation, relating to creative solutions and intuitive assessment.

Finally, a good case for AI in the court’s process relates to the question of costs. It might be possible for AI to track the progress of the proceeding by reference to the court’s file, and to determine the costs that should be payable at the end of the proceeding. This might eliminate the work of the costs court and save tens of millions of dollars over time.

Assuming our current notions of justice remain, AI will have significant difficulties, at least in the foreseeable future, attempting to do “judge” work. At present, it is simply impossible. As the former Chief Justice of the High Court of Australia, Robert French AC, has pointed out, it is the courts and only the courts that can carry out the adjudication function involving the exercise of judicial power.¹⁰⁷ The courts exercise the power of the state, interpreting laws, ensuring procedural fairness and rendering binding decisions across society at large. However, the question arises whether less complex legal analysis and decision-making can be performed by AI.

It is generally accepted that some of the work of tribunals may be less complex than superior court litigation, because the knowledge field and expertise required is narrower, and there is less need to develop overarching principles. Further, for some tribunals, case-by-case outcomes are generally acceptable without the need for formal rules of evidence and broad principles. AI may have some significant role for these tribunals over time. For example, in relation to refugees, the decision-maker may be an AI algorithm based on input data of the applicant’s background and an assigned percentage likelihood that refugee status is justifiable. Another example, current at present, is whether someone is a citizen of a particular country,¹⁰⁸ or even a dual citizen. Apparently INDIGO, a large legal expert system deployed by the Dutch immigration administration, deals with issues like this.

This kind of work of AI would require data to be inputted, which presumably would need to be certified by a legal practitioner or other qualified person. Such outcomes might also need to be audited by a “human” decision-maker for reasonableness against some normative standard. The audit might be like a Tax Office audit, targeted at certain decisions. And if AI was rejecting virtually all applications, the algorithm would presumably need to be revisited.

Finally, in some parts of Europe, family court settlements are determined without hearings, using allocations of assets between husband and wife based on an algorithm of what is fair. In the Family Court of Australia, software is being developed that generates advice on how property from a marriage would be split under a court determination.¹⁰⁹

Down yet a further level of complexity and dollar value of amounts in disputes are “small claims”. The question is whether these can be “AI’d” out of existence. This has in part already been suggested. There is a product (a web-based interface) called DoNotPay,¹¹⁰ which is dedicated to assisting people to make small claims or defeat legal infringements such as parking or speeding fines.¹¹¹ The user fills in a form on screen, which is dedicated to the particular issue at hand (eg defending a fine), and the form is submitted to the relevant adjudicator (such as the small claims court). By using the product, a person does not need to use a lawyer. But that is only to start the claim and, in any event, the claim is of a size that would not justify a lawyer. This product will not, therefore, revolutionise the litigation system; however, it might make one small segment of it more efficient at the front end. It is a laudable product as an “access to justice” device; it enables people to take a formal litigation step.¹¹² But it does not do much more. An associated AI product has been developed in relation to class actions, where individuals can log their claims individually.¹¹³ The basic idea is that the product lets a complainant approach a defendant electronically so as to be able to settle a claim online.¹¹⁴ For more commercial work, there are legal online services such as those offered by Allen and Overy in London, generating more than £12 million for the firm each year from subscriptions.

If AI developers are able to take this initial “dip in the water” at doing litigation without a lawyer to further stages of development, the litigation landscape might change and possibly materially. However, at present, that looks a long way off. The question is not whether AI can improve in order to do this, but whether the structures of our legal system can be redesigned, giving AI more work to do because much of the discretion is taken out of the system. That is a large philosophical question, not merely an AI question.

Yet further down the tree of dispute resolution are the systems established by private individuals to quell controversy. Examples of this include dispute resolution provisions for sporting bodies or for consumer transactions, such as purchases over eBay.¹¹⁵ In effect, consumers contract to resolve disputes electronically and agree to be bound by the dispute resolution process. It has been estimated that three times as many disagreements each year among eBay traders are resolved using “online dispute resolution” than there are lawsuits filed in the entire United States (US) court system.¹¹⁶ This may drive efficiencies, but not all are satisfied that it is appropriate. The Paris Bar has expressed disapproval with non-lawyers providing legal services by trying to solve legal claims without lawyers. The Paris Bar has worked with regulators to create the Paris Bar Incubator, which works to focus new legal technology.¹¹⁷

Finally, even with AI dispute resolution, one would expect that a human’s judgment would at least on occasions be necessary by way of a control or an audit over that work. In any event, if there are to be such AI outcomes, this fundamental question arises: who will be the “owner” of the AI and/or the system governing it?¹¹⁸ The natural answer is the court, the tribunal or the organisation (eg eBay) in question. But, as mentioned above, will a court or tribunal have the capability and resources to maintain and improve the AI over time? Will it have the resources to maintain the controls within the AI to ensure that it is meeting its purpose over time? Will it have the resources to audit those controls? Or will the court/ tribunal ultimately “outsource” the development of the AI to contractors, as it might do its IT systems? These governance issues do not just pertain to the court system.

All of the above questions seem to generate a more general question: whether a court has to be a physical place, or whether it is in fact the service of justice? Certainly, the presence of a court is a symbol of government authority and facilitates highly complex problems (both logical and philosophical) being dealt with iteratively by competent legal experts and the court itself. However, alternative dispute resolution online, and courts themselves online, are a developing and significant trend.¹¹⁹ They facilitate affordable access to justice, timely dispute resolution and binding results (ending lengthy appeal processes).

In this area, it may be possible that AI could be used to perform decision-making through pre-agreed contractual arrangements. As stated above, eBay is the classic example of this. Further, there are examples of low-complexity online courts already in existence around the world — in the Netherlands, British Colombia and the United Kingdom.¹²⁰ These online processes are intended to enable an exchange of facts, identify issues for resolution, and suggest consensual resolution of disputes if possible, failing which the dispute may go to a hearing or the dispute may be resolved “on the papers”.

So far this article has been essentially exploring the notion of whether there can be “behavioral equivalence” between traditional human-based, judicial decision-making and AI-based decision-making; that is, whether outcomes can be the same under each pathway. So far, at least to this author, behavioral equivalence is impossible except at the most basic procedural level such as rudimentary procedural tribunal work and the like.

But let us assume nevertheless for the moment that behavioral equivalence is possible. The question arises whether, although the outcomes are the same, there are nevertheless significant other differences which remain.

To consider this further, let us suppose that a particular litigant attends court, sees the court processes in motion, observes the judge in question, and assesses the reasoning for the result by reading the judgment. The litigant will necessarily make certain value judgments about the litigation and the result. The litigant will no doubt ask themselves various questions, including: was the process honest? Was the court courageous in its determination to upheld rights? Was the overall process fair? Was the result fair?

Now assume the same result has been achieved using AI. These questions do not go away when AI is introduced to determine the result. These same value judgments about honesty, trustworthiness, courage, and fairness, will all still inevitably be made by the litigants.

The literature on AI and the courts has a dearth of learning on how AI may affect (perhaps adversely) on those values as attributed to the court system by litigants or even observers.¹²¹ Despite this, it is essential for governments to assess how the tradeoffs between human judgments and AI judgments in quelling controversies via adversarial (or even inquisitorial) systems are valued by litigants and society in general. Much work is needed in this area to determine, before introducing AI into a particular process, or so as to replace a process, how the involvement of AI will affect the values which litigants and society place on the relevant court system and its status in society.

Law firms advise on and implement complex transactions, including mergers and acquisitions. AI now exists for this kind of work and is being embraced by large law firms across many continents.

In late 2016, American academics published a paper entitled “Can robots be lawyers? Computers, lawyers, and the practice of law”.¹²² In summary, the research accessed the time records of lawyers in major American law firms, allocated work across certain work streams (eg document management or legal writing), and then attempted to determine what work could or could not be done by AI. Surprisingly, the analysis revealed the absence of a strong association between the ease of automating a task and whether the task was performed by a junior associate, senior associate or a partner. Nevertheless, the overall conclusion was that automation could have a material employment effect on lawyers.

A major AI product now available in the legal services market is called Luminance, based in Cambridge in the United Kingdom.¹²³ It has developed data systems for NATO (the Northern Atlantic Treaty Organisation) and the British National Health Service. Its competitive advantage is that it applies “new maths” to solve complex real-world challenges where there are very large, heterogeneous, constantly evolving datasets. This is what Luminance says about the product:¹²⁴

According to its website, Luminance provides lawyers with the most rigorous analysis of their documents, instantly highlighting anomalous areas or risks that require urgent attention, working across the entire dataset, negating the need to rely on sampling, giving lawyers confidence that no critical document or clause has been missed. The AI does the burden of low-level cognitive tasks common in due diligence, compliance, insurance or in-house contract management, enabling lawyers to spend more time advising clients on business-critical issues. According to Luminence’s website, one of its clients has been able to upscale the review of documents from a rate of 80 per hour, to some 3,600 per hour. Luminance technology uses inference, deep learning, natural language processing and pattern recognition, supervised and “unsupervised” machine learning. Luminance has a broad and increasing client base across the globe.

Another type of AI product is called “COIN” or Contract Intelligence, developed by J P Morgan Chase and Co. The product engages machine-learning algorithms to review and interpret commercial loan agreements. The company has estimated that this AI saves 360,000 hours each year of lawyer and loan officer work.¹²⁵ It has been said that by the use of these kinds of products, AI is “closing in on the work of junior lawyers”. That is probably right.

This begs the question whether there is such a product for litigation. Law firms assist in the litigation process in a large number of ways, including:

AI can and will increasingly be deployed in aide of these tasks.¹²⁶ For example, as to the third point (reviewing documentation), electronic discovery processes already take the place of lawyers or paralegals looking through large volumes of material, thereby significantly reducing time and cost.¹²⁷ Courts have more recently embraced discovery using algorithms, including so-called “predictive coding”, as a means of keeping the discovery obligation manageable.¹²⁸

As to the fourth point (advising on the law), predictive coding has been applied, selecting relevant statutory provisions from those retrieved with keyword searches.¹²⁹ Further, significant research tools are now available. In the United States, for example, you can submit a legal question to a firm, which will use AI to analyse the issues and produce a suggested memo to be provided to the relevant client. There is a turnaround time because humans have to turn the memo into a more credible product.¹³⁰ As mentioned above, ROSS (an IBM product) is a legal research tool. It enables law firms to slash the time spent on research, while improving results. According to IBM, existing technologies such as keyword search poorly makes sense of the volume, variety, velocity and veracity of legal data. Watson’s cognitive computing capability enables ROSS’ intelligence. The ROSS application works by allowing lawyers to research by asking questions in natural language, just as they would with each other. According to IBM, because it is built upon a cognitive computing system, ROSS is able to sift through over a billion text documents a second and return the exact passage the user needs. According to IBM: “Do more than humanly possible. Supercharge lawyers with artificial intelligence.”¹³¹ However, it should be noted that a recent review of ROSS¹³² has found that it is still a “supplement” to traditional Boolean searches. In the review, young lawyers were asked to answer American bankruptcy law questions using ROSS. ROSS outperformed other search technologies but still only found about half the relevant authorities within the first 20 results.¹³³ There were, however, significant reductions in research time, which transpose to lower costs to clients and therefore a more competitive law service. The above review concluded that the gains from using ROSS did not constitute a dramatic transformation in the use of technology in legal services. Rather, the tool was characterised as a “significant iteration in the continuing evolution of legal research tools that began with the launch of digital data bases of authorities and have continued through developments in search technologies”.¹³⁴

As to point 5 (assessing prospects), there currently exist AI products that will seek to predict litigation outcomes. This is based, inter alia, on not only the facts and the law but AI analysis of the particular judge allocated to the case and the particular counsel retained. It is not inconceivable that an AI product might be developed that will assess the risk on a “real time” basis — that is, during the trial — as the transcript and submissions are made to the court over the course of a trial. It has been said in relation to the US Supreme Court that computational statistics can often yield more accurate predictions of the likely behaviour of courts than retained lawyers in the cases.¹³⁵ The predictive litigation outcomes are particularly suited to the US legal system where the level of activity is truly immense. For example, there are 3,124 state courts in America, and over 15 million civil lawsuits filed each year (roughly 41,000 claims a day). This gives rise to a significant opportunity to “mine” the data from such cases and to develop predictive algorithms.¹³⁶ Such data is unlikely to be available other than in the huge legal litigation markets such as the United States.¹³⁷ Predictive algorithms, in respect of particular judges, is prohibited in France.¹³⁸

Anterior to all this is the question of whether litigation should be commenced in the first place. Here we might see the introduction of “self-executing contracts”,¹³⁹ able to initiate legal proceedings if AI detects breaches. Of course, that is a rather naive approach to litigation as launching it is substantially about reputation and risk as much as legal rights.

As to legal research, AI will develop apace, but it has to be said it is currently well advanced already. Legal research now takes a fraction of the time it used to take, say, 10 years ago.¹⁴⁴ Counsel will no doubt be advantaged by such developments.

AI might also be able to do the pleadings for a case. Counsel (or their instructor) would simply input the facts in question. The AI algorithm would then analyse the facts, draw on its database, and then produce a statement of claim, with particulars, containing the relevant causes of action and the appropriate relief. There might be yet other applications for AI. It was touched on above whether AI products could be deployed to predict litigation outcomes. This is likely to increasingly emerge, displacing the traditional quantitative guesses of counsel that the case is “50/50” or the qualitative assessments that the case is “well arguable” or “unlikely to succeed”.

However, where AI must necessarily fall short is with the art of the advocate (sometimes called court craft) — for this requires an understanding of the “social processes” of the courtroom, including as between the judge, opposing counsel and witnesses, taking into account the observance of duties to the court. Further, problem-solving towards a fair resolution of a dispute requires a didactic interaction between counsel for each party and the court. In effect, at least three super-computers (brains) are engaged on the one series of problems within a proceeding. Each hypothesis, fact, analysis, reasoning and conclusion is subject to interrogation. This permits deep reasoning across the so-called “open-textured rules” landscape referred to above. Further, the “art of the advocate” involves fine and frequent judgments in a complex environment. It is hard to see AI affecting this aspect, particularly where qualitative judgments are called on to be made regularly.

There may be areas where the work of counsel will increase with AI. One such area might be in relation to the testing of the relevant algorithm or algorithms producing the “computational law”, and the challenges (or support) for the relevant databases relied upon by those algorithms. Whether this will require special new expertise by trial counsel — for example, in the area of computing and/or engineering — is yet to be seen. Another area that might increase the work of counsel is in the possible certification of factual data for input into AI algorithms, particularly for tribunals whose work has been subsumed by AI. This would limit (but not eliminate) “gaming” of the relevant system. The current Civil Procedure Act 2010 (Vic) requires certain certifications by practitioners before proceedings can be commenced. This might be extended through the use of AI.

As foreshadowed above, there remains the broader question of what the Bars around Australia should do regarding AI. Since 2015, the Canadian Bar Association has had an Information Technology Committee, which, inter alia, monitors developments in computer technology, case law and regulations relating to and regards AI as a specific area of interest. The Committee is also tasked to keep members informed, to comment on legislation, and to make recommendations in relation to changes in the law in this area. The Victorian Bar has a similar committee. It is clear enough that the Bars across Australia should make every endeavour to monitor and keep on top of the AI revolution. And perhaps computing and AI should ultimately be compulsory continuing legal education for counsel (a need law schools are increasingly aware of).

The regulation of the development of AI is a new frontier. AI development in Silicon Valley is regularly described as “the wild west”¹⁴⁵ Regulators are grappling with the kinds of policies that should be put in place to deal with AI development, and governments across the globe are undertaking governance reviews of AI and monitoring what other governments are doing.¹⁴⁶ However, already, there are well developed principles across a range of disciplines, including in engineering activities.

An obvious regulatory field for consideration is driverless cars. Should they be required to pass certain regulatory standards before they can take to the road? The answer seems obvious, but what would be the content of those regulations? Would the car need to pass a test (like a person does to obtain a licence)? If a driverless car crashes, who is at fault?¹⁴⁷ Simply identifying the relevant defendant may be a challenge: the company writing the algorithms; the data upon which the driverless car relied; the satellite companies that supplied the data; the electrical engineer who serviced the car? In Germany, the legal authorities are currently thinking about how the next level of laws should cope with this challenge.¹⁴⁸ Such dilemmas may well call for a fundamental policy reset, involving “no fault” litigation, or insurance coverage for driverless cars.¹⁴⁹ If there is damage suffered from the use of AI, will the court have to look at the dataset the AI relied upon and determine who imposed a bias (if any) in the data, who set the size of the dataset, and who did or did not review the AI algorithm? If there is damage suffered from the use of AI, and there was no “protocol” for the use of the AI, is this itself prima facie a negligent act? There are also heavy societal governance issues associated with the use of AI for objects such as cars. Who should be in control of or audit AI’s use? Take, for example, a company that sells a car with advanced AI that is defective. If the directors do not insist on an audit (or independent certification) of the AI and/or its use, are the directors negligent?

Principles have also now been developed relating to the use of AI in the administration of justice.¹⁵⁰

Courts themselves will have to grapple with other new controversies arising from the use of AI. It may be that particular laws do not adequately cover new scenarios, for example: who owns an AI algorithm derived by another AI algorithm?¹⁵¹ With the inexorable aggregation of data for the purposes of enabling AI to perform useful analysis, how are privacy and ownership rights over the data best protected. Much of the mining of data by AI is on “the Cloud”. Despite privacy and confidentiality protocols, is the Cloud in fact appropriate for data mining? Particularly in legal proceedings?

This aggregation and mining of data gives rise to a further issue: the evidentiary regime in litigation relating to such data. The Evidence Act 1995 (Cth) operates essentially as a filter on what is or is not probative evidence. If AI draws on (mines) data from a huge dataset, how is the court to determine the probative value of that extraction, and therefore the conclusions derived from that extraction? There may not be any peer review of the extraction or other check or control on it. It might just be “pulled” from the internet. It might be highly valuable but be given very low probative weight because its source is not verifiable. These are all imponderables of future litigation.

One obvious but discrete area for consideration regarding AI is the possible development of non-discretionary elements within legislation or regulations — that is, elements designed so that there is no call for an evaluative process. This might apply, for example, in relation to citizenship status, social security payments, or certain compensation amounts; these could all be determined by algorithms, updating themselves (learning) on a predetermined basis. Essentially, this would be the work of automated design systems (ADSs) as mentioned earlier.

Further, as mentioned above, the use of digital cameras to recognise people and fine them for offences gives rise to a whole area of regulatory oversight issues.¹⁵² Further, in the area of AI weapons, we may well see international treaties or protocols for their use. Finally, there is the whole question of the control of a country’s population by AI systems.¹⁵³ This is a fertile area for the courts, lawyers and counsel.

It appears AI will not be constrained by computing power, memory capacity, cost or data collection (as to the latter, at least in many fields such as insurance, finance and medicine). The interesting question therefore is this: what will constrain it? There are two important factors. First, AI cannot as yet read or comprehend, or itself reason, so it has significant limitations in understanding. In effect, it has no understanding of why it is doing something,¹⁵⁴ and AI cannot as yet innately reason to a conclusion. Secondly, there may be significant forces at play that impose “protocols” or constraints on the use of AI — for example, there may be an international treaty that only “amoral” algorithms can be developed.

AI is at the “hype” stage at present, but significant studies across industry suggest the overall impact will be high. One study found that, across the technology, media and telecommunications industry, more than 70% of respondents expected large effects from AI in five years.¹⁵⁵ Even in the public sector, 40% of respondents had this prediction. Having said that, according to one source, the number of convincing examples of the implementation of AI as commercially viable products do not make it inevitable that there will be generalised negative results for professions. An October 2016 survey of the adoption of machine-learning-based AI in legal services by large British law firms over the previous year, showed that there were only three reported applications that had been actually developed, plus a further nine collaborations/agreements partnerships in respect of that.¹⁵⁶

It seems that the popular view is that the more dramatic effects of AI may occur within 10–20 years, when data consolidation has been better developed.¹⁵⁷ While there may be significant job losses due to automation (eg in call centres and banks), the 2016 World Economic Forum report suggests that upcoming disruption to employment will be multifaceted rather than narrowly influenced. For example, there are areas of expansion in employment; Infosys has trained more than 120,000 employees in design thinking.¹⁵⁸

It should be noted that there is a significant investment around the world in AI development — in 2016, there was an estimated $26–40 billion spent on it.¹⁵⁹

In the United States an AI Index has recently been developed, put together by Stanford University and others. In November 2017, the first (annual) AI Index report was published.¹⁶⁰ The report deals with the volume of activity in the AI field and its technical performance, and makes comments on “derivative measures” and observations about “towards human-level performance”. The report has significant limitations in that it does not contain data about AI research and development by governments or large corporations, and is US-centric. Nevertheless it shows a heavy uplift in activity, and that AI can beat human capability in certain areas (such as object detection), can match speech recognition, and is nearly as quick as a human to find the answer to a question within a document. Yet the report notes that AI’s capability, in contrast to a human, significantly degrades when tasks become “generalised”.

It is true that there are calls, in some quarters, to slow the development of AI for ethical and safety reasons. For example, in mid-November 2017, hundreds of AI researchers from Canada and Australia wrote to Prime Ministers Justin Trudeau and Malcolm Turnbull, calling for an international ban on “weaponised AI”.¹⁶¹ These kinds of concerns have been identified and raised as far back as 30 years ago.¹⁶² Assuming this were to occur, the question might be: how would this be enforced? By what court? And how would the court satisfy itself that a particular piece of weaponised AI is or is not within acceptable protocols unless the court is itself able to interrogate the AI?

Nevertheless, it seems the pace of AI development will only increase. The world’s leading experts cannot agree on when, if at all, AI will develop to the point of human-level intelligence, but some have said by 2055 or thereabouts.¹⁶³ It is not inconceivable that one day human intelligence might not be the apogee of intelligence. But the literature suggests that this is out past at least 50 years.

There are now literally thousands of Silicon Valley firms working on AI products in the legal services space. This may well increase in number. And as we have seen, under Moore’s Law, computing power will continue to develop with immense capacity in the future. It seems unlikely that the development of AI in legal services will slow — rather it is likely to increase dramatically as firms try to develop products that can produce steps in the litigation process that are cheaper and more effective, giving service providers competitive advantages in the delivery of legal services.

Finally, it is also not inconceivable that one day AI will prevent or eliminate controversy in society, because there will be higher levels of certainty. If this occurs, there may well be no controversies and therefore courts. If AI could eliminate controversy, because everything is far more certain, and where the “risk” of a counterparty defaulting is perfectly priced in a contract, there may be no need for litigation. But that outcome seems a very, very long way off from today.

It seems that in the short to medium term, AI will continue to be developed, and applied, in at least four areas in litigation:

If AI is to have an impact on the way superior court litigation is conducted, it is likely that a necessary condition for its influence will be less flexibility built into the litigation system — that is, less opportunity to provide corrective mechanisms for procedure or substantive irregularities or decision-making. This possible reset in the priorities of the courts is essentially a legislative task.

At lower tribunal levels with less complex controversies, AI may play a much more significant role. However, that role may develop more slowly than first thought; the development would need to come from government, which is not the natural engine room of entrepreneurial improvements in systems. Developments in the private sector to improve client services and lower costs may not always align with the societal framework changes needed to deal with AI. In any event, the natural engine rooms for developing AI further seem to be academia, businesses with an appropriate risk capital profile (venture capital), and businesses with huge balance sheets (such as IBM).¹⁶⁴ However, it is very possible that, even if there is material change in this regard, this may ultimately be seen to be sacrificing justice for efficiency. The right balance between these two is always contextual: an advanced wealthy democratic society may well tend to skew the balance in one direction; a different kind of society may skew it the other way.

The legal profession, including the Bars, must be on the look out for ways to improve access to justice, at lowering costs, and at improving time scales for litigation. As Nettle J has stated:¹⁶⁵