Some reflections on the use of artificial intelligence and blockchain technology in insurance litigation

Australian Professional Indemnity Group National Conference

Justice Jackman 4 September 2025

A great deal has been written about the risks and opportunities of artificial intelligence and blockchain in the legal system. There is a broad spectrum of views, from something akin to revolutionary zeal at one extreme to Bourbon obscurantism at the other. My views fall between those extremes. There is the good, the bad and the ugly. Most of us are aware of AI’s remarkable achievements and its increasing reliability and scope of application, but can resist the temptation to think that just because it does some things very well, it can do similar things equally well. Blockchain technology is also becoming more widely understood, mainly through the use of cryptocurrency, or crypto assets. In this paper, I will seek to outline the nature of AI, especially Generative AI, to try and identify its main strengths and limitations. I will then consider ways in which I see it impacting on insurance claims and in particular the litigation of those claims. I will not venture into other areas where Generative AI has useful applications in the insurance industry, such as in the pricing of premiums based on risk cost modelling, and insurance claims management where the extraction and analysis of key data by AI can improve the claims process. I will also indicate some ways in which blockchain technology has created new means of alternative dispute resolution.

I do not have the self-assurance, or the self-harming need, to attempt a definition of “artificial intelligence”. A definition would be necessary only if one were drafting a statute or regulation to deal with the subject-matter. For present purposes, it is sufficient to identify the central conception of AI as a range of techniques using machines to replicate the problem-solving capabilities of the human mind. That is, AI seeks to perform activities which would be considered intelligent if they were done by people. One could readily pick holes in that description if it were put forward as a formal definition. For example, it would include a pocket calculator, which is too mundane a machine to capture the essence of what we are discussing. More importantly, the idea of imitating the capacities of the human mind does AI a disservice, in that many of its applications are superhuman in their speed and functionality, and the models are so vast and complicated that we cannot understand their processes using concepts that can be held in human minds. The essence of AI is that it involves mimicry and imitation in its outputs, although not necessarily in its internal process. There are those who insist that AI involves no more than mimicry machines, and that the letters “AI” should really be treated as if they are stood for Advanced Imitation. That wrongly suggests that the machine is actually reproducing the function or process of human reasoning, whereas it is merely giving the appearance of doing so in its outputs, which are the result of superhuman capacity to crunch data and find statistical patterns. When AI plays chess, drives a car, or writes an essay, it is imitating activities which we normally associate with human reasoning, but it is doing so using statistical methods which do not (or which do not necessarily) correspond to human ways of thinking. AI thus outwardly mimics humans, but it does not operate internally like humans.

The early manifestations of AI were rules-based, also known as symbolic AI. That sought to replicate human problem-solving by encoding expected behaviour through a series of rules which could be applied to a body of knowledge. The computer was programmed with predetermined responses to a set of conditions. The AI system would follow paths which were laid out in decision trees. That had the benefit that the system could explain its lines of reasoning, but it had the drawback that the system could only produce outputs that had been explicitly inputted into its knowledge base. However, we now have so-called Machine Learning, whereby the machine is able to learn from statistical patterns within the data which it identifies, and the machine can be trained (ie taught by experience) to perform a specific task. That involves the creation of algorithms and statistical models that enable computers to accomplish tasks (such as making predictions or choices) without the need for explicit programming. In that way, the computer system can be programmed to improve its performance over time. And within the field of Machine Learning, there is Natural Language Processing (NLP), which uses language, whether written or oral, as the data. An everyday example is the email spam filter: the algorithm is trained on a set of emails, determining which are spam and which are not, and it must learn how to classify incoming emails which it has not seen before.

Taking that a step further, Deep Learning is a branch of Machine Learning which mimics the neural networks of the brain, being the basic structure of how humans think and process information, to allow patterns to be learnt from data. And within Deep Learning, we find Generative AI systems. The fundamental idea in Generative AI systems is that they build up an underlying representation of the data they are provided with, and try to fill in the blanks, thereby creating new data that mimics an existing data set. The system thus produces new instances of similar data from an apparent understanding of the underlying structures and rules that make the data what it is. I say “apparent understanding” because the system is in reality determining statistically which “token” (such as a word) comes next in a sequence generated in response to a user prompt.

Large Language Models (LLMs) are in the heartland of Generative AI systems. I vividly remember French and German classes at school in the 1970s in which we were given what were called Cloze tests. They involved a passage in the foreign language with blanks which we had to fill in by choosing the right word. A language model is effectively a machine which guesses the right word in that kind of way. It will start badly by guessing words randomly, but with exposure to many millions or billions of examples it will learn to guess correctly, thus making its output look like an informed prediction by someone truly familiar with the subject-matter. Importantly, a language model learns the mechanics of language through statistics, and in that way it can predict outputs from new unseen inputs. But the language model has absolutely no concept of truth or factual accuracy, nor does it have any familiarity with the real world. Large language models can generate text which reads admirably, but they achieve that through the artificial objective of statistical consistency with the prompts which they are given. They are better at words than they are at meaning. That is why, in the now infamous New York case of Mata v Avianca Inc 22-cv-1461 (PKC) (S.D.N.Y. 4 May 2023), the claimant’s lawyer, who had relied on an artificial intelligence tool to produce submissions to the court, was found to his intense embarrassment (and no doubt to the anxiety of his professional indemnity underwriter) to have cited non-existent cases, even to the point of advancing fake quotations. These kinds of errors are quaintly referred to as “hallucinations”, another anthropomorphism. My own view is that hallucinations are an inherent feature of these systems, at least within the current technological constraints. That is because they are not actually “errors” but a byproduct of how these systems generate language, namely by prediction from statistics. As an aside, over-reliance on the technology seems very likely to be a growth area in professional negligence cases.

Before the invention of Generative AI, there was only a small set of tasks performed by lawyers which could be mimicked through AI, such as the extraction and labelling of data. Generative AI, however, has substantially increased the tasks which machines are able to mimic, as anyone familiar with chatbots such as OpenAI’s GPT–4, and in the last month GPT–5, will be aware. Such chatbots, even if they are not optimised for providing legal information, can give basic legal information and advice, respond to basic legal inquiries, and guide users to relevant resources, and can do so very cheaply and often for no charge at all.

Put in terms of statistical theory, the goal of Generative AI is to mimic the probabilistic distribution that governs why certain data instances exist in the real world and why others do not. Generative AI thus creates new and unique instances of data that look as though they could belong to the original data set, and it does so with great sophistication and nuance. To my mind, ChatGPT’s composition of a poem on cricket in the style of John Keats is better than its attempt at a poem on rugby in the style of Gerard Manley Hopkins, but they are both convincing imitations. But its basis in statistical method means that it has very significant limitations when applied to litigation. Let me give several examples which are relevant to insurance litigation in particular.

First, as every insurance lawyer knows, there is a fundamental difference between causation and correlation. That distinction, for example, lies at the heart of disputed medical evidence in product liability cases. Generative AI is very good at picking up correlations and regularities in data. But its lack of knowledge of the actual world means that it has real difficulty in analysing causes which may be responsible for the correlations in the way that human reasoning does. So I asked ChatGPT if GPT–4 can distinguish between causation and correlation. It replied as follows:

Yes, GPT–4 can distinguish between causation and correlation in theory and provide explanations of the difference. However, like any AI model, it does not infer causation on its own from raw data – it can only analyse arguments, identify logical fallacies, and reference statistical methods used to establish causality. (emphasis in original)

That showed the machine imitating an admirable degree of self-awareness of the kind captured in Dirty Harry’s famous exhortation: “A man’s got to know his limitations.”

Having said that, the areas where AI has clear limitations of AI are rapidly shrinking. Consider, for example, the problem of onus of proof on the civil standard of the balance of the probabilities. As Sir Owen Dixon said in Briginshaw v Briginshaw (1938) 60 CLR 336 at 361, when the law requires the proof of any fact, the tribunal must feel an actual persuasion of its occurrence or existence before it can be found, and it cannot be found as a result of a mere mechanical comparison of probabilities independent of any belief in its reality. The facts proved must form a reasonable basis for a definite conclusion affirmatively drawn of the truth of which the tribunal of fact may reasonably be satisfied, not merely a choice between guesses on the ground that one guess seems more likely than another or others: West v Government Insurance Office of New South Wales (1981) 148 CLR 62 at 66 (Stephen, Mason, Aickin and Wilson JJ). That is an issue on which one might think that Generative AI can never be a substitute for human reasoning. The point may be illustrated by the hypothetical example given by Jonathan Cohen in his book The Probable and the Provable (OUP, 1977) of 499 people paying for admission to a rodeo, but 1,000 are counted on the seats, including the defendant. Is the defendant liable on the basis of that information for non-payment? As Cohen contended, a court would not find in favour of the promoter of the rodeo in an action against the defendant on the basis of a bare 0.501 probability that the defendant did not pay. As Justice DH Hodgson said extra-judicially, the bare 0.501 probability would not be based on adequate material concerning the circumstances: “The Scales of Justice: Probability and Proof in Legal Fact Finding” (1995) 69 ALJ 731 at 735–6. I posed that very problem to GPT–4, which told me with great (but misplaced) confidence that it is more likely than not that the defendant did not pay for admission on the legal standard of the balance of probabilities. However, this particular answer could be improved by a better model that is trained to deal better with the concept of proof on the balance of probabilities. The new paid model of ChatGPT (GPT–o1) in fact points out that the concept does not depend only on statistical evidence as that would lead to liability on the part of every spectator when 499 of them did pay, and says correctly that the plaintiff must show that this particular defendant is more likely than not to be a non-payer.

As a separate matter, insurance law is permeated in Australia with the elusive and vague concept of utmost good faith, which sections 12–14 of the Insurance Contracts Act 1984 (Cth) make a paramount term of contracts of insurance. That is necessarily a value-laden issue, which requires sufficient experience of the real world to know what commercial decency and fairness (and not just honesty) require: see CGU Insurance Ltd v AMP Financial Planning Pty Ltd (2007) 235 CLR 1 at [15] (Gleeson CJ and Crennan J); and Allianz Australia Insurance Ltd v Delor Vue Apartments CTS 39788 (2022) 97 ALJR 1 at [95]–[96]. As the unsuccessful counsel in Allianz v Delor Vue in the High Court, I was delighted to be told by GPT–4 that the facts in that case gave rise to real concerns over whether the insurer had acted with the utmost good faith. As many of you will be aware, there is nothing a barrister (or in this case a former barrister) likes more in the absence of a win than seeking to claim the moral high ground. But the voices of infallibility on the High Court, by majority, held that my argument was wrong.

Just as Generative AI has no concept of truth, it also has no concept of fairness or decency, although it may well pretend otherwise. Perhaps a machine could become so good at mimicking decisions based on fairness and decency that it would almost always predict what a judge would conclude. That is a useful hypothesis by which to test the moral debate. It is my firm view that the debate is about more than accuracy. It is also about respect for, and the dignity of, people who are subjected to legal power. We want to know why a decision is made against us and we want that explanation to come from a fellow human being who takes responsibility for moral and legal judgments, and who has a grasp of what fairness and decency require in the real world.

A further problem is that the more accomplished Generative AI is in filling gaps, the less sensitive it will be to the existence of a hard case, where the available legal materials do not lead logically to a single answer. The law has an open texture, the metaphor being of gaps between the warp and the weft in woven cloth. The law can never be entirely deductive. By their nature, rules have a core of certainty and a penumbrum of doubt. When a judge encounters a problem in that fringe of open texture, the decision will be guided by broadly evaluative reasoning, such as an assumption that the purpose of the rules which they are interpreting is a reasonable one, or that the rules are not intended to work an injustice or offend settled moral principles (H.L.A. Hart, The Concept of Law, OUP, 1961, p 200). At times, a court is faced with a highly technical argument, which appears contrary to the real merits of the case. Judges, being human, then tend to ask themselves “Must I accept that argument?” rather than “Can I accept that argument?” Judicial decisions can thus involve moral values or a choice between moral values, and morality does not always have a clear and discernible answer to offer (ibid., p 200). The ultimate decision is neither arbitrary nor mechanical, but involves an impartial and reasoned survey of the alternatives and a well-developed sense of judgment. Judges speak of “weighing” or “balancing” the competing interests. I do not think that artificial intelligence can replicate that kind of reasoning, nor in my view can it reliably predict the outcome in such a case. It certainly cannot dignify the person whose interests are at stake with a proper explanation produced by genuine intellectual engagement with the problem, rather than merely predicting the “correct” sequence of words as a large language model does.

There are obvious benefits of using AI in litigation, including insurance litigation. This can be seen in the competition in the United States between Lexis and Westlaw, which have augmented their legal research databases with search tools that can handle free-form questions rather than key word searches. There are undeniable time and cost efficiencies in using AI, particularly in relation to document review and analysis, legal research, and sometimes in preparing initial drafts of documents. In that last reference concerning draft documents, I stress the use of the word “initial”. In ACCC v Master Wealth Control Pty Ltd (Penalty) [2024] FCA 795 at [78], I dealt with a criticism by the ACCC of the respondent for having proffered to the court a draft document which appeared to have been produced by Generative AI. However, I did not see any problem with that in circumstances where there was nothing to suggest that the respondents’ lawyers had not scrutinised and settled the draft thereby produced and satisfied themselves that it was appropriate for their purpose. I disagreed with the respondent’s position, but not because of the use of Generative AI.

There is also a case to be made for the use of Generative AI in improving the quality of legal services. I wonder whether the problem which was dealt with by the New South Wales Court of Appeal in HDI Global Specialty SE v Wonkana No 3 Pty Ltd (2020) 104 NSWLR 634 would have arisen if the drafting of the business interruption insurance policies in question had been cross-checked by using Generative AI. As most of you will be aware, those policies excluded from cover “diseases declared to be quarantinable diseases under the Quarantine Act 1908 (Cth) and subsequent amendments”. However, the drafting had overlooked the fact that the Quarantine Act 1908 had been repealed about five years earlier by the Biosecurity Act 2015 (Cth), and it was under the new Act that COVID-19 was determined to be a listed human disease. The exclusion therefore had no practical application, with a potential consequence for insurers and insureds worth many millions of dollars. Apart from the drafting and review of documents, lawyers must digest large amounts of information, both in terms of evidence and in terms of legal precedent and commentary. AI systems are capable of digesting very large amounts of data in a timely and cost-effective way, and in a way which is increasingly reliable. In the context of litigation, Generative AI can extract legal principles from published judgments, search vast quantities of evidence for key words or correlations, and even predict the outcome of some cases by synthesising past decisions. None of that will eliminate human processes, but it can provide a very beneficial adjunct to augment the skills of lawyers.

We must ourselves, of course, be careful not to succumb to the temptation to think like machines in solving legal problems. This is not a recent phenomenon. For example, it is often submitted in insurance litigation that, in construing a contractual clause which uses a defined term, one can simply rewrite the operative words of the clause by substituting the definition for the defined term. But as the High Court said in the insurance case of Halford v Price (1960) 105 CLR 23 at 28 (Dixon CJ, with whom Menzies and Windeyer JJ agreed), and 32–33 (Fullagar J), that is only the first step. One must also look to the contract as a whole to ascertain if that literal approach is appropriate. As Lord Hoffmann said in Chartbrook Ltd v Persimmon Homes Ltd [2009] 1 AC 1101 at [17], defined terms in a commercial contract should not be treated as akin to algebraic symbols, but rather are labels which are seldom arbitrary and are usually chosen as a distillation of the meaning or purpose of a concept intended to be more precisely stated in the definition. In my view, the well-trained and thoughtful legal mind has a clear edge over a machine in that regard.

There is clearly scope for technology to provide insureds with a low-cost way of ascertaining, at least to some degree of satisfaction, what their rights are and how they might deal with them. That point was well made by James Allsop shortly after his retirement as Chief Justice of the Federal Court, in a paper which expressed very considerable caution generally about the potential uses of Generative AI in litigation in light of the importance of the social bond between the state and the citizen which is reflected in our courts: The Humanity of Law: Selected Essays of James Allsop (edited by Ruth Higgins and Kevin Connor, The Federation Press, 2024, p. 165). The paper acknowledged that GPT-4 was said to be able to pass the US Uniform Bar Examination multiple choice paper at a higher level than the average candidate, and it could even pass the legal essay and problem-solving task (p. 164). The insureds who resort to Generative AI for the purpose of assessing their prospects against their insurer are likely to be those who may well have baulked at the cost of consulting a qualified lawyer before pursuing their claim. That does not, of course, mean that, a higher number of litigated claims will necessarily result. The use of Generative AI may well provide many insureds at an early stage with sufficient information to dissuade them from pursuing their claims further or with vigour, and the risk of an adverse costs order by a court will remain as a significant deterrent for speculative claims. However, access to reliable legal advice at an affordable cost is one of the desiderata of the rule of law, and Generative AI has the capacity to play a significant role in meeting that principle. The potential for Generative AI to assist community legal centres in reducing unmet demand, for example, is the subject of an illuminating and measured article by Will Cesta, “Large language models and community legal centres: Could chatbots help reduce Australia’s justice gap?” (2024) 49 Alternative Law Journal 181.

Generative AI also has a potentially significant role to play in what is currently a highly inefficient process favoured by modern litigators, namely the formal mediation of disputes. For the last 35 years or so in Australia, formal mediation has been a popular means of seeking to compromise litigation, even to the point where many regard it as a moral imperative. Perhaps surprisingly, it is often employed as a first, rather than a last, resort in seeking to compromise litigation. That results in the parties having to pay their legal representatives to sit around unproductively for very lengthy periods of time waiting for one’s opponents to respond to offers or proposals, and to share the costs of a professional mediator to supervise proceedings. There will be cases where that is desirable, but in my experience there are many cases where it results in a substantial waste of time and money, except perhaps for the (no doubt) incidental and wholly unintended benefit of gathering information on one’s opponent’s case. The insurance industry has a strong vested interest in remedying these inefficiencies, given that insurers are probably the most regular participants in mediations, whereas for insureds it is typically a one-off experience.

An alternative has been developed in the United States by a number of online platforms, which include at least one specifically targeted at insurance claims, with catchy brand names such as Cybersettle.com, Smartsettle.com and ClickNSettle. The essential concept is a blind bidding negotiation mechanism whereby the claimant and the respondent each submit the lowest and highest settlement figures respectively that they are prepared to accept. The platform’s algorithms then compare the offers confidentially and see whether the parties are within the range of agreement. Neither party sees the other party’s offer or demands. An offer of settlement is sent by the software to each party, without disclosing the amount to the other, allowing each party to submit a counter-offer, and then again submitting an intermediate point to the parties for acceptance or rejection. The proposed resolution is arrived at by Generative AI using data from previous attempted negotiations. The double blind method appears to avoid the irritating problem in mediations with so-called “splitting the difference”, which of course rewards the party that holds back. The algorithm does not seek to replace humans but to help them arrive at an outcome by obtaining information from both parties to the dispute, and combining those data points with its own data set to suggest solutions. That may well provide a faster and cheaper way to resolve disputes than conventional mediation, and at the very least offers a useful starting-point. As Professor Richard Susskind argues in Online Courts and the Future of Justice (Oxford University Press, 2019, pp 138–41), there is a powerful case for the courts themselves to be providing this service. That is especially so in Australia where courts routinely provide mediation services through their registrars.

Turning then to blockchain technology, a blockchain is a ledger method for recording transactions. The data is organised in blocks or groups across many computers that are linked and secured, but each block can only hold a certain amount of information, so new blocks are added to the ledger and this forms a chain. Blockchain refers to each block being chained to its predecessor.

Importantly, there is no single point of ownership or control of the blocks or the data in them. Blockchain is a distributed ledger technology that involves a list of transactions that is shared across many computers rather than being stored on a single centralised network or server. The network of computer devices is referred to as “nodes”. Equally important is that if the ledger says that a particular address has crypto assets then that makes it so. Unlike conventional ledgers in business accounting, these ledgers cannot misrepresent the real position, because they are themselves the reality. The ledger both records and determines who has what crypto assets. Any data sought to be added to the blockchain undergoes a validating procedure by way of a consensus mechanism, based on algorithms. Any unauthorised modification to an existing block triggers a warning to all the other nodes on the network, and, at least with public blockchains, any new data is stored only with the approval of the nodes authenticating the modification. Public confidence in the blockchain, such as the Bitcoin or Ethereum blockchains, thus depends on whether the requisite consensus renders the blockchain tamper-proof.

In terms of alternative dispute resolution more generally, there is a number of online dispute resolution platforms (known as ODRs or smart dispute resolution, SDR) which are based on blockchain and its capacity for “crowdsourcing”. These combine the technology with a reliance on what are claimed to be the psychological insights of game theory. At the risk of oversimplification, the basic concept is that a number of potential jurors deposit cryptocurrency, and the more they deposit, the higher the chance of them being selected as a juror for a case. Those selected are anonymous to each other so as to prevent collusion. Once selected, the jurors receive the evidence and decide the case. Those jurors who vote with the majority are rewarded by a proportional fraction of the sum of all deposited tokens, whereas the jurors who vote with the minority have their tokens taken away. This is said to create the incentive to vote honestly and reasonably.

I am deeply sceptical of the assumption that an outcome which is objectively fair will be arrived at by guessing what participants generally will think is fair (or more accurately guessing what others will guess what other participants will guess to be fair and so on adinfinitum). It is worth recalling the Keynesian beauty contest which John Maynard Keynes deployed in Chapter 12 of The General Theory of Employment, Interest and Money in 1936 (p 156). Imagine a beauty contest in which the judges do not choose the contestant whom they think is the most attractive. Rather, they choose the contestant whom they think the other judges will regard as the most attractive. Keynes saw that as illustrating how short-term stock market fluctuations are not caused by changes in underlying value, but instead by investors attempting to figure out what others think the typical investor finds valuable. In other words, there will be disparities between what a reasonable market participant regards as reasonable and what the same participant thinks others will regard as reasonable, and there is money to be made in the stock exchange and other capital markets by correctly predicting the majority’s choice rather than standing by one’s own personal judgments. Indeed, the current market price of many crypto assets provides a compelling illustration of the phenomenon. Nevertheless, those who use ODRs and SDRs do not appear to be concerned with any those disparities.

The original application of that technology was in the field of disputes concerning blockchain, smart contracts and distributed ledger technology, as illustrated by platforms such as Kleros, Juris, JUR and Aragon. One can perhaps understand the technology gaining market acceptance in that field, given the difficulties of applying conventional dispute resolution in a field where participants act anonymously (or more accurately by way of alphanumerical pseudonyms) and in different jurisdictions. The technology perhaps also suits a system whereby distributed ledger systems depend ultimately on the informal consensus of participants. The rules which govern those dealings tend to be self-enforcing in practice because only transactions made in compliance with the consensus that underpins them, and thus which are duly entered in the ledger, will be accepted by participants as valid. The proponents of the technology sometimes draw a rather strained analogy with the lex mercatoria, or “law merchant”. During the Middle Ages in Europe, a body of legal conventions was created through customary commercial practice and enforced by private courts in the major merchant trading centres. The law which developed was largely made and administered by merchants for themselves, based on notions of good faith and fair dealing, free of technicalities.

But can blockchain technology be deployed in resolving disputes outside the field of blockchain itself? The answer is “yes”, and some of the platforms already mentioned claim a wide range of use applications, including insurance disputes. The point is illustrated most radically by the oddly named “RHUbarb” platform and its business companion, the PeopleClaim platform, which according to the PeopleClaim website has negotiated over 30,000 cases. The promoters call their process “online community dispute resolution”, which they liken to the ancient Athenian trial of civil disputes by way of a public citizens’ forum typically comprising several hundred jurors. However, they do not tell you that the citizens of ancient Athens were paid only a small fixed fee to participate in that process and were certainly not paid extra if they conformed to the popular opinion. Aristophanes even wrote a play on the subject, namely Wasps in 422BC, which satirised the enthusiasm of old men for jury service. The background was that the jury fee introduced by Pericles was less than an able-bodied man would earn by an ordinary day’s work, with the result that many of the men who volunteered were too old for work. The analogy is therefore a good deal weaker than the promoters claim.

Participants in the online community dispute resolution post a deposit of the cryptocurrency known as RHU (being the RHU in RHUbarb) which they will forfeit if their vote falls outside the final consensus, in favour of participants who make up that consensus. With the consent of the respondent to the claim, the poll verdict binds both parties. The RHUCoin website gives a worked example of how its poll verdicts work, in which an insured, Ella, has her insurance claim denied. Ella posts three possible solutions online: (1) the insurer pays her $2,400; (2) the insurer pays her $4,800; and (3) the insurer pays her nothing. There is also scope to allow voters to suggest other solutions. Ella then chooses to submit the claim to 1,000 network participants comprising 400 consumer advocates, 250 insurance professionals, and 350 Californian consumers. I have nothing against consumer advocates or Californians but you might sense at that point that Ella is gaming the system. Ella and the insurer (who can’t have been very alert) then decide that the poll verdict will be binding. If they had treated the poll verdict as non-binding, the poll verdict would still be taken and the parties could negotiate against the background of that poll verdict. The RHU network members are then invited to participate, with each participant depositing one RHU. When the poll is taken with the 1,000 network participants, 380 votes are cast in favour of a $4,800 payment, 280 votes are cast in favour of a new option (suggested by one of the participants) of a $5,200 payment, 220 votes are cast for no payment, and 120 votes are cast for a $2,400 payment. The consensus is the 380 votes for $4,800. The 620 voters who chose non-consensus solutions forfeit their RHU deposits, and the 380 consensus voters each receive 1.63 RHU from the forfeited deposits, being a 63% return on investment. Ella, the insurer and the poll voters are automatically notified. Ella’s RHU wallet is instantly credited and the consensus voters are depicted with big emoji smiles. According to the website, “life is good for everyone”. The website rather breezily dismisses the concern that voters will act tactically instead of fairly by voting for the result they feel will be chosen by the largest group instead of the most equitable outcome, by saying that in most cases the fairest resolution will be the one that earns consensus. No mention is made of the Keynesian beauty contest in which anticipating what the popular opinion expects the popular opinion to be takes the decision away from what is objectively fair and reasonable, not towards it.

On the PeopleClaim website, the same promoters are candid about the pressure they seek to exert on respondents (such as insurers) to bring about a resolution, saying that although they do not guarantee a successful resolution:

if you choose the public posting we do guarantee that if [your claim is] not resolved to your satisfaction it will remain online and available to search engine queries and public comment. This provides a powerful and ongoing incentive to fair resolution.

If that minatory prospect is not sufficiently blunt, the promoters then say:

Businesses today want to do the right thing, or at least appear to do the right thing. Exposure of bad behaviour has become a stronger motivation in many cases than going to court. Public posting on PeopleClaim allows search engines to index your claim and make it easy for people to discover your problem. Bad behaviour or unfair treatment is a turnoff to other consumers using the web to make purchasing or hiring decisions. The more people file against the same party the stronger the incentive to settlement, which means that most businesses would rather fix a problem that’s disclosed online than let it remain posted.

Welcome to the brutal world of online and not-so-smart dispute resolution. One of the great virtues of our legal system is that cases are not decided as popularity contests. The unpopular litigant, perhaps one who is despised by public opinion, is treated with exactly the same human dignity as the most fashionable celebrity. All are entitled to the equal protection of the law. It would be invidious to say where insurance companies fall in that spectrum but it is sufficient to say that I find it difficult to imagine that any insurer potentially confronted by what it regards as unjustified consumer complaints would willingly submit to this process. But this commercial experiment is a useful vehicle for identifying and testing the criteria that should be applied to measure the success of a system of dispute resolution, including alternative dispute resolution.

First and foremost, any system of adjudication must be fair, both in a procedural and a substantive sense. Procedural fairness demands impartiality, which requires that decision-makers not have a material financial interest in the outcome of the dispute. ODR platforms operate on the directly opposing premise that decision-makers can and should profit from their decisions, in the belief that game theory will apparently ensure that the profit-motive will incidentally produce a fair outcome. Whatever view one takes as to the application of game theory in this regard, intentions rather than merely outcomes are fundamentally important to the legitimacy of any acceptable system of adjudication. Economic liberalism, in my view, is best pursued within a framework of stable and enduring institutions and processes under the rule of law which serve the common good, and provide a patterned structure within which individuals can pursue their goals in an ordered setting. It is theoretically conceivable, for example, that judges or other adjudicators may decide cases in a self-interested way, with an eye perhaps on private economic advantage or career advancement, but the person who is moved by calculations of self-interest waters down any concern for the functions of law (or accepted commercial morality for that matter) as an answer to real social problems and renders less stable the process which makes such adjudication viable. As Professor John Finnis argues, that person dilutes his or her allegiance to law with doses of the very self-interest which it is an elementary function of law to subordinate to social needs (Natural Law and Natural Rights, OUP, 1980, p 14). And a system of dispute resolution which is built on such a premise and treats legal order as a game undermines or weakens the practical viewpoint that brings law into being as a distinct type of necessary social order. I recognise, of course, that the participants in so-called community online dispute resolution are private citizens rather than officials who are appointed by the state to administer or apply laws. But in a healthy society an allegiance to the legal system whereby people genuinely accept and use the standards of the law in assessing their own and others’ conduct must be shared beyond the world of officialdom, especially if the non-officials are engaged in adjudicating disputes.

Procedural fairness also requires that the parties to the dispute be given a regime for the discovery of documents which may not favour each party’s case, and an opportunity to test their opponent’s evidence by cross-examination. By contrast, on-line platforms enable each party to upload whatever evidence they wish to present and conceal damaging material. That is not conducive to fair adjudication.

Second, any system of adjudication requires accountability; that is, that those people who have authority to make decisions are accountable for their compliance with rules applicable to their performance, and do actually administer the relevant norms (whether legal or moral) consistently and in accordance with their tenor. ODR makes no attempt to satisfy that criterion. Rather, it places its trust firmly in a questionable view of game theory on the postulate that accountability can then be dispensed with.

Third, decisions must be explainable. The ODR platforms which are directed to disputes concerning blockchain transactions do make some attempt at this, with provisions imposed to varying degrees for the jurors who comprise the majority to give short reasons for their decision. Community-based ODR makes no such attempt, consistently with the basic premise that participants should decide on the basis of a calculation as to which of the range of possible outcomes will win the most votes. An honest participant in such a system might simply say as the real reason for his or her decision: “Well, nothing succeeds like success”. To return to an earlier point, human dignity requires that decisions are explained by adjudicators who demonstrate by their reasons a genuine intellectual engagement with the problem at hand.

Fourth, decision-making should be more or less predictable and reliable. Whether ODR platforms satisfy that criterion essentially depends on how much faith one is prepared to place in its view of game theory.

Fifth, decision-making should be efficient in terms of cost and time. That is ODR’s main selling-point, as a low-cost, accessible and fast mechanism. However, that benefit is achieved only at the expense of the other criteria to which I have referred.

All of this serves to remind us that law is both an art and a science. Its application is not merely a process of logical deduction from a closed set of legal rules. Nor is it determined by statistical probabilities. The very question of what parts of legal work can best be performed by legal technology rather than by the human mind is itself a question which only the human mind can answer. And when it comes to designing systems of dispute resolution, even those preceded by the word “alternative”, the systems must have a legitimacy in sharing the fundamental objectives of legal adjudication. There is much to be gained from legal technology, and potentially much to be lost by too fervent an application of it. Even if one were to assume away all the technical constraints which currently exist, the robo-lawyer, like the robo-judge, still belongs in the realms of science fiction, not in the world of real human beings looking to solve real problems in a real society.