Dispute Resolution: Master, What Does the Future Hold?
While litigation scaled fees may have been postponed, a larger challenge to the dispute resolution profession draws closer: the disruption that artificial intelligence (AI) promises. This article summarises the challenges that AI brings to dispute resolution lawyers.
But scrambled laws don’t hide my ignorance
To write some lines for you, my discs do spin
I sort them quickly, making lawyers wince
I’m trapped inside a box of wires and tin1 The title and verse were randomly derived from www.totopoetry.com, which uses AI to generate “poetry”
Recent developments in machine learning have been dramatic. An earlier generation of AI, which focused on brute force computation, is being overtaken by machine learning, with its human-like, eerily uncanny results.
The rate of change can be startling. It took only nine hours for Google’s AlphaZero to learn how to play chess, and then crush the reigning AI chess champion, Stockfish 8.2 For a summary see https://deepmind.com/blog/alphazero-shedding-new-light-grand-games-chess-shogi-and-go/ AlphaZero represents a new paradigm. It did not win because it could make more calculations: AlphaZero could calculate about 60,000 moves per second, compared to Stockfish’s 60,000,000. Instead, AlphaZero’s chess play was described (by humans) as “creative”, a “relentless positional boa constrictor approach that is simply unheard of”,3https://www.theguardian.com/sport/2018/dec/11/creative-alphazero-leads-way-chess-computers-science; https://en.chessbase.com/post/the-future-is-here-alphazero-learns-chess. “like discovering the secret notebooks for some great player from the past.”4 https://deepmind.com/blog/alphazero-shedding-new-light-grand-games-chess-shogi-and-go/
And that was in December 2017. Last month, AlphaZero’s developers announced a demonstration using AI to drive and accelerate bioscience research.5 https://deepmind.com/blog/alphafold/
Dispute resolution lawyers like to tell ourselves that our craft, consisting of persuasion and creativity, will largely be immune to AI disruption. We think the corporate lawyers will be the first to go. Yet AI already has been gnawing away at key parts of the dispute resolution process, like discovery and document review. If AlphaZero takes only nine hours to be the Master of chess, how much time would a specifically tailored AI need to “master” the common law? Probably much quicker than it would take you to read Treitel on Contract.
Imagine a machine that can, in seconds, trawl through the case law and provide you with the two or three authorities that best match the facts of your case. Now, would that not be a game changer? No more fear of being ambushed with a “supplementary” bundle of authorities moments before a hearing starts. No more having to put up with diva first-year associates with first class law degrees or having to attend CPD talks.
Actually, that machine is already here.6 Say hello to Ross: https://rossintelligence.com The larger question is what the future holds.
One way to make sense of AI disruption is to see AI as “prediction machines”.7 This term and the following analysis are derived from Agrawal et al, Prediction Machines: the simple economics of Artificial Intelligence (HBR Press 2018). What this means is that AI takes existing information (ie, data) and uses it to generate information that it does not have (ie, prediction).
For example, when AI translates a sentence from English to Mandarin, it takes existing information (its language database, including patterns of use) and uses that to predict the correct order and appropriate characters in Mandarin. Similarly, when Amazon.com suggests products for you to buy, it is taking data (eg, your shopping and browsing history) and using that to generate a prediction of your desires.
In the same way, when you make a transaction with your credit card overseas, your bank’s AI predicts the risk of whether that transaction is fraudulent. If the risk is high enough, you get a call from customer service. The AI in self-driving cars predicts what a human would do and controls the car accordingly. When AI identifies unlabelled objects, it essentially asks itself “does this image have the same features as the [object] that I have labelled before?”, thereby predicting what the correct label of that object is.
Machine predictions are not subject to the heuristics and bias that influence human predictions.8 See Kahneman, Thinking, Fast and Slow But the accuracy of machine predictions depends on the amount of data that it has. When rich data is available, machine predictions can surpass humans in speed and accuracy.9 Eg, AI is outperforming the average radiologist in identifying tumors etc: Aerts et al, Artificial Intelligence in Radiology, Nature Reviews Cancer 18, 500-510 (2018); https://www.economist.com/leaders/2018/06/07/ai-radiology-and-the-future-of-work More and better data leads to better predictions.
Accordingly, AI has been used accurately:
- To “predict” the outcome of US Supreme Court10 Daniel Martin Katz et al, A General Approach for Predicting the Behavior of the Supreme Court of the United States, https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0174698 and European Court of Human Rights decisions;11 https://www.ucl.ac.uk/news/2016/oct/ai-predicts-outcomes-human-rights-trials
- To analyse the win rate of lawyers against different judges;12https://premonition.ai/; https://home.ravellaw.com; https://lexmachina.com/what-we-do/
- To interpret commercial loan agreements and contracts;13https://www.bloomberg.com/news/articles/2017-02-28/jpmorgan-marshals-an-army-of-developers-to-automate-high-finance; https://www.lawgeex.com/
- To advise on sentencing and generate arrest warrants;14 https://www.telegraph.co.uk/news/2017/08/04/legal-robots-deployed-china-help-decide-thousands-cases/ and
- To evaluate possible litigation outcomes for a party before the case is filed.15 https://www.bakermckenzie.com/en/insight/publications/2018/02/adoption-ai-chinese-courts.
Conversely, machine predictions are poor when there is little data. AI would probably have struggled to predict Pakatan Harapan’s victory in the 2018 Malaysian general election.16 Malaysia’s 14th general election. AI is also not very good at predicting events that occur rarely — ie, “black swans”. Examples of black swans would be Napster’s development in 1999, which decimated the music record industry17Prediction Machines at p 61. and the arrest last month of Huawei’s CFO in Canada.18 https://beta.scmp.com/tech/tech-leaders-and-founders/article/2176654/huaweis-cfo-sabrina-meng-wanzhou-has-been-arrested In contrast, humans can be very good at making prediction with little data: we can reason by analogy, have a sense of the “fair” or “just” outcome even in novel situations, articulate paradigm shifts, devise strategies to deal with unexpected events.
Prima facie, it would follow that machine prediction is likely to be poor in dispute cases that (a) require a development of the existing law;19 Eg, extension of the duty of care in tort (b) require an application of the law to novel facts; (c) lack substantial volume of case law; or (d) require wider considerations and judgment of public policy and local context.20 Eg, most public law cases.
As the use of machine prediction becomes cheaper and more accessible, it will increasingly replace human predictions, and the value of human prediction will fall.21Prediction Machines at p 76. The flipside is that skills that only humans are good at may increase in value. In particular, demand will grow for the skills that only some humans are good at, or which some humans are much better at than other humans.
The most important factors that distinguish humans from machines are arguably creativity and empathy.22 See Kai-Fu Lee, AI Super-Powers: China, Silicon Valley, and the New World Order (Intl Ed) at chapter 8 Machines can optimise, but they cannot create, especially with a lack of data. Neither can machines empathise with humans or feel compassion.23 The importance of empathy should not be underestimated. See e.g. doctors and the importance of “bedside manner”: https://medschool.ucla.edu/body.cfm?id=1158&action=detail&ref=699
The next question is where creativity and empathy may be relevant in the typical dispute resolution process.24 One could make the case that empathy or social skills is relevant throughout the entire dispute resolution process. But in a contest between “cheap and quick” and “affable”, I would put my bet on the former. For the importance of affability, see: Cameron Ford, The 4As of Courting Corporate Counsel: http://v1.lawgazette.com.sg/2017-08/1920.htm The field of dispute resolution is broad, encompassing a $1,000 claim for the repayment of debt to a multi-billion dollars cross-jurisdictional trusts tussle. But nonetheless one may identify six broad stages common to many disputes.
The first is the taking of instructions from the client. This is essentially a data gathering and data input exercise: you are taking data from the client, assessing relevance, extracting details of the relevant, and trying to move the client on from the irrelevant. Save for the most routine of cases, this stage is likely to remain human-centred for the near future.
The second is identifying and researching the applicable legal principles. Essentially, the aim here is to predict the success of the available legal arguments. This area is primed for AI disruption. If the legal principles are established and the facts straightforward (as it is in many cases), machine prediction will outperform humans.25 Dispute resolution, interestingly, is where the Anna Karenina principle — that every unhappy family is unhappy in its own way — does not seem to apply. The obvious exceptions would be where one seeks to develop the law, or when there is a novel factual matrix where an application of existing law would lead to an unpalatable (to humans) outcome.
The third is the crafting of the case theory and pleadings. This is an area that requires creativity — to create a sensible and workable narrative based on the data available — and prediction — assessing the case theory’s chances of success. The strong element of creativity provides reasonable protection from AI disruption. The exceptions may be the kinds of cases where there is no necessity to create a compelling narrative: eg, the typical traffic accidents/personal injury, drug-trafficking, progress payments cases.
The fourth stage is alternative dispute resolution, eg, mediation. It is true that mediation involves the art of negotiation. But the space for negotiation to a large extent depends on the respective strengths of each side’s case. As machine predictions get more accurate, one would expect the scope for negotiation to get smaller. Say an AI reliably predicts that a claim for $10,000,000 has a 70% chance of success if it goes to court. It would be quite unrealistic to expect the Plaintiff to agree to settle for anything much less than $7,000,000. So in this sense, the process of mediation might soon be heavily influenced by machine predictions.
The fifth stage is discovery and document review. This involves predicting the relevance of a particular document. AI is already colonising this stage, albeit in conjunction with human supervision and review. Any lawyer who has spent the last few years specializing in document review should start getting worried.
The sixth is the hearing. This is an exercise in oral advocacy (persuasion), empathy (reading the court, the witness, opponents), and prediction (predicting what the witness will say under cross-examination, how persuasive the judge will find the evidence and submissions). There is unlikely to be major AI disruption at this stage for the foreseeable future. AI may however play a larger supportive role: eg. assessing from past decisions how favourably a judge may view an argument or a particular witness. And as we move to “documents-only” hearings, or hearings where the focus is on written advocacy, the room for human-only skills will gradually contract.
Hence the importance of creativity/empathy at each stage of the dispute resolution process will depend largely on the nature or type of the dispute. On a side note, it is interesting that the quantum of the sum in dispute per se is not a relevant consideration.26 Which is something to keep in mind when discussing the relationship between scaled fees and the quantum of dispute.
We are now in a position to create a working impression of the different kinds of dispute resolution cases that are at risk from AI disruption.27 This chart is adapted from AI Superpowers at pp 155–156.
It is important to state that this chart is impressionistic. The aim is to identify the areas of dispute resolution work that are most vulnerable to AI disruption. It is not comprehensive. It is necessarily simplified. The “positions” on the chart are impressions based on the typical or paradigm type of case, not the hard or difficult cases.
The chart is divided by two axes. The horizontal axis expresses the scale between disputes work with tasks that can be optimised and those that require creative input. Tasks can be optimised if they involve searching for an optimal outcome that can be derived from data.
The bottom-left quadrant is the “High Risk” zone. Much of the work here can be optimised with relatively little need for human input or empathy. There is accordingly a high risk of AI disruption, with machine predictions taking over many of the human tasks.
The top-left quadrant can be called the “Human Veneer” zone.28 I adopt the helpful terminology set out in AI Superpowers at p 157. Many of the tasks found here are suited for AI optimization, but because the nature of the work requires some level of empathy / social interaction, human involvement will probably remain necessary in the medium-term, although a large part of the work may be split with AI. Newish roles for lawyers may be created. For example, even if machine prediction replaces human assessment on the division of matrimonial assets, matrimonial clients may still want “Counsellors” to provide emotional and legal support. So family lawyers may find that aspect of their work increasing in significance.
The bottom-right quadrant is the “Slow Creep” zone. The work here does not rely on empathy or social skills, but more on creativity and strategy. The ability to create (particularly in the absence of rich data) is likely to remain a substantial hurdle for AI, but we should expect incremental advances.29AI Superpowers at p 157. Note that already AI is displaying signs of what may be described as “creativity”: https://www.washingtonpost.com/news/innovations/wp/2016/03/15/what-alphagos-sly-move-says-about-machine-creativity/?noredirect=on&utm_term=.f8e27a0af2f2. Work in the “Slow Creep” zone is also vulnerable to an unexpected breakthrough in AI technology.
The top-right quadrant is the “Safe Zone”. The requirements of creativity and empathy mean that these kinds of work are likely to be insulated from drastic AI disruption for the foreseeable future. For example, “Complex Civil Cases”, ie, those that may result in landmark judgments, will require lawyers to have creativity (because there will be little direct precedent) and some empathy (an innate sense of fairness). These cases will also require a fair amount of oral advocacy, and so will resist the trend towards document-only hearings. Similarly, prosecuting or defending crimes of passion (murder, etc) will generally require some understanding of the human condition, the ability to construct a compelling narrative, and the skill of interrogating what may be a unique set of facts.
We can expect AI to bring challenges and opportunities. But the costs of AI disruption may fall more heavily on those who are not well-placed to take advantage of the opportunities. The one-lawyer firm who has over decades built a practice in debt recovery will find it challenging to transition to “higher value” work or to incorporate AI into his workflow. Smaller firms, some of which lack WestLaw and/or LexisNexis subscriptions because of the costs, may balk at the fees for AI. Larger local firms have the resources to adopt AI, but history shows that incumbents are often casualties of disruption. Another concern is that the prevalence of US, China, and UK-centric AI programs will give their law firms first-mover advantage in AI adoption over Singapore law firms. This concern is particularly acute if Singapore firms are to compete for international work.
On a brighter note, it is only the start of 2019. There is still time to prepare.
|↑1||The title and verse were randomly derived from www.totopoetry.com, which uses AI to generate “poetry”|
|↑2||For a summary see https://deepmind.com/blog/alphazero-shedding-new-light-grand-games-chess-shogi-and-go/|
|↑6||Say hello to Ross: https://rossintelligence.com|
|↑7||This term and the following analysis are derived from Agrawal et al, Prediction Machines: the simple economics of Artificial Intelligence (HBR Press 2018).|
|↑8||See Kahneman, Thinking, Fast and Slow|
|↑9||Eg, AI is outperforming the average radiologist in identifying tumors etc: Aerts et al, Artificial Intelligence in Radiology, Nature Reviews Cancer 18, 500-510 (2018); https://www.economist.com/leaders/2018/06/07/ai-radiology-and-the-future-of-work|
|↑10||Daniel Martin Katz et al, A General Approach for Predicting the Behavior of the Supreme Court of the United States, https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0174698|
|↑12||https://premonition.ai/; https://home.ravellaw.com; https://lexmachina.com/what-we-do/|
|↑16||Malaysia’s 14th general election.|
|↑17||Prediction Machines at p 61.|
|↑19||Eg, extension of the duty of care in tort|
|↑20||Eg, most public law cases.|
|↑21||Prediction Machines at p 76.|
|↑22||See Kai-Fu Lee, AI Super-Powers: China, Silicon Valley, and the New World Order (Intl Ed) at chapter 8|
|↑23||The importance of empathy should not be underestimated. See e.g. doctors and the importance of “bedside manner”: https://medschool.ucla.edu/body.cfm?id=1158&action=detail&ref=699|
|↑24||One could make the case that empathy or social skills is relevant throughout the entire dispute resolution process. But in a contest between “cheap and quick” and “affable”, I would put my bet on the former. For the importance of affability, see: Cameron Ford, The 4As of Courting Corporate Counsel: http://v1.lawgazette.com.sg/2017-08/1920.htm|
|↑25||Dispute resolution, interestingly, is where the Anna Karenina principle — that every unhappy family is unhappy in its own way — does not seem to apply.|
|↑26||Which is something to keep in mind when discussing the relationship between scaled fees and the quantum of dispute.|
|↑27||This chart is adapted from AI Superpowers at pp 155–156.|
|↑28||I adopt the helpful terminology set out in AI Superpowers at p 157.|
|↑29||AI Superpowers at p 157. Note that already AI is displaying signs of what may be described as “creativity”: https://www.washingtonpost.com/news/innovations/wp/2016/03/15/what-alphagos-sly-move-says-about-machine-creativity/?noredirect=on&utm_term=.f8e27a0af2f2. Work in the “Slow Creep” zone is also vulnerable to an unexpected breakthrough in AI technology.|