Key Contacts: Clare Cashin – Partner | Paddy Mockler – Associate
Introduction
In recent years, there have been significant developments in the capabilities of artificial intelligence (AI), and particularly the ability to generate high-quality content through the use of so-called ‘generative AI’.
This article examines aspects of construction disputes which could potentially be simplified and improved by a particularly prominent form of generative AI, ‘large language models’ (LLMs) which are capable of analysing and producing large quantities of text. LLMs are ‘trained’ using vast sets of data which allow them to understand and predict the next word (or series of words) in a sequence.
On 2 February 2024, the House of Lords Communications and Digital Committee published its report into “Large language models and generative AI” which warned against the risk of failing to capitalise on the opportunities afforded by emerging AI and falling behind competitors through an overly narrow focus on AI safety. To offer a comparison on the scale of these opportunities, this report indicates that “Large language models (LLMs) will have impacts comparable to the invention of the internet”.
Can construction disputes potentially benefit from the use of such emerging LLM technology? It would certainly seem so, and the task at hand is to balance the risks appropriately with the benefits.
The Role of AI in Construction Disputes
The varied and flexible nature of AI tools are well suited to disputes arising under construction contracts, as this is typically an environment in which each case will stand heavily on its own facts. Taking a complicated set of agreed contract data, AI may even facilitate the early detection of potential risk areas and provide stakeholders with the opportunity to address issues before they escalate to a formal dispute procedure.
Dispute Avoidance
As modern projects have become increasingly sophisticated, so too have construction contracts become increasingly complex. A clearer understanding of the critical provisions, obligations and potential areas of contention between the parties is one area in which the market will seek to leverage AI. Similarly, AI can streamline the review process by carrying out automatic analysis of volumes of contract documents to identify potential inconsistencies.
Dispute Resolution
The potential benefits of AI are not limited to the avoidance of disputes.
Where it becomes necessary to engage in a formal dispute resolution process, there are a number of key AI tools available to assist. The use of LLMs in the context of live disputes carries a significant health warning – it would not be advisable to rely solely on the output from such models (the risks associated with such reliance have been demonstrated in practice as addressed below). However, a cautious approach to these tools can present a decisive advantage for astute parties to a construction dispute.
Of particular value could be the ability of AI to analyse lengthy legal submissions and identify potential weakness or inconsistency in the arguments presented. While it is certainly inadvisable to replace the human role in such a review, AI can offer an advantageous starting point and may expedite the parties or legal practitioners consideration of aspects which may not have been immediately apparent. Unlike its human counterparts, AI is capable of analysing large volumes of legal argument and the relevant contract documents simultaneously and comparing the relevant portions of each in real time. Such capabilities will only become more advanced and widely available in the coming years, which is likely to change the manner in which parties approach document heavy disputes.
Assessment of Merits
While the AI tools which are currently available on the market remain in their infancy, one feature of particular interest down the line will be the use of such technology to evaluate and predict the outcome of a live dispute process. Such a prediction will be reliant on historical case data and the factual evidence and submissions of the case at hand and so it goes without saying that the dataset which the model has access to will be critical. However, if this process could be adopted effectively, then it could facilitate an efficient assessment of the merits settlement in many cases which would otherwise incur the costs of being heard in full.
Pitfalls of using AI in legal Disputes
Hallucinations
For better or for worse, the risks associated with the use of AI in a legal context are well known. The New York case of Mata v Avianca 22-cv-1461 (PKC) (2023) ought to have been of limited local interest, but gained global prominence when it transpired that legal authorities had been cited by the Respondent’s legal counsel which were fictional and had been generated as a result of the submissions being drafted by AI. This revelation came to light by degree, as the suspicions were played out through exchanges between the parties and the judge which led to the attorney admitting that AI had generated the decisions and legal authorities relied upon; he was “operating under the false perception that this website could not possibly be fabricating cases on its own.”
In sanctioning the attorneys involved, Judge Castel assessed the fake ‘decisions’ which had been presented, and identified a number of attributes would give rise to a suspicion that they were illegitimate (including legal analysis classified as “gibberish” and internal flaws such as one decision citing itself as a precedent).
The wide publicization of the infamous decision in Mata v Avianca did not put an end to the matter. The issue resurfaced in the recent case of Harber v Commissioners for His Majesty’s Revenue and Customs UKFTT 1007 (TC) which was heard by the UK First-Tier Tribunal (FTT) Tax Chamber. In that case, a litigant in person presented a number of purported FTT decisions which appeared to support the appeal, but which could not be located on the FTT website or any other source. Despite the litigant’s claims that the decisions had been obtained from “a friend in a solicitor’s office”, the FTT concluded that they were not genuine and had been generated by an AI system.
Interestingly, in reaching this conclusion the FTT also relied on the principles outlined by Judge Castel in Mata v Avianca and endorsed his comments that the submission of fake authorities is particularly harmful as it leads to wasted time for the opposing side to expose the deception, consumes limited court resources and promotes cynicism about the legal system.The anomalies which occurred in the above cases are known as ‘hallucinations’ which are occasionally produced by LLMs where false information is presented, often in a convincing manner. Such hallucinations can be tricky to detect; indeed in Mata v Avianca the AI software at issue (when questioned as to the authenticity of the material) specifically reaffirmed the position that the cases were genuine and that they were available “in reputable legal databases”.
These are cautionary tales. Liability in construction disputes can often turn on fine details, and so the fact that snippets of information within the text produced by LLMs may appear genuine, but in fact are a hallucination, will justifiably cause concern to any party seeking to adopt these models.
A critical drawback in the capabilities of AI, is that it will typically be ill-equipped to deal with the wider context of a potential dispute. For example, although a party to construction contract will be acutely aware of the factors which could prove relevant to a dispute (the relationship between the parties, cashflow and supply chain management for example) an LLM will not be aware of any of these factors as it relies on the dataset on which it has been trained. The more nuanced approach which is adopted in practice, could appear illogical from the perspective of AI.
Confidentiality
Of particular concern to stakeholders will be that the openly available AI solutions could potentially carry significant consequences for confidential or commercially sensitive information. The extent to which data is used and retained will vary depending on the particular model and provider which is selected, however it is essential to be keenly aware of the risk. This issue came to a head in April 2023 when employees of Samsung inadvertently leaked confidential information to the popular ChatGPT platform which could then be used to ‘train’ the model and become publicly available to other users. This incident, which led to a decision on behalf of Samsung to temporarily ban the use of generative AI tools, should be seen as another cautionary tale given the significant level of commercially sensitive information involved in construction disputes.
Recognising this, there are a number of ‘private’ AI tools available on the market which are aimed at creating a closed-off model within internal organisations. This is a rapidly developing area and organisations will need to ensure that any such offering is procured from a trustworthy source (and that the terms of the arrangement are clearly understood before it is deployed). Indeed, this is a key development which will be required before clients and their advisors can have confidence in the use of such tools.
Statutory Adjudication and AI
Parties to construction contracts and practitioners alike will be familiar with the Irish Construction Contracts Act 2013 which provides for the right to refer a payment dispute to statutory adjudication at any time.
The potential benefits to be derived from the use of AI are readily apparent in relation to such adjudication, which is carried out within a short timeframe (28-days from referral to decision subject to limited exceptions). In spite of these tight deadlines, the volume of material which may be the subject of a dispute in adjudication can be extensive given the wide-ranging entitlement of the referring party to define its payment dispute, and the obligation of the adjudicator to consider any defence raised by the responding party.
With this in mind, there are potential advantages to be obtained by employing AI, and specifically LLMs, to assist with the more administrative work such as document review and proof-reading. When the period of time to produce multiple rounds of pleadings is limited, one can see a great benefit to resources being engaged in legal analysis, leaving AI to assist with the administrative tasks.
Although statutory adjudication does not require a discovery process in the same manner as a dispute heard before the courts or in arbitration, the volume of documentation which is exhibited to the pleadings delivered can often be challenging to navigate in the time periods available.
Needless to say, any submission of a claim or defence in adjudication will require careful and comprehensive consideration by the parties to the dispute. However, the ability of AI to distil voluminous pleadings and supporting material into the core points advanced could prove to be an invaluable starting point. In a process where the deadlines are measured in days as opposed to weeks and months, fine margins of this kind could prove significant to the outcome.
Conclusion
The general benefits of this emerging AI technology are evident in the scope to reduce time and legal fees incurred during a dispute resolution process. Ultimately however, the gains in efficiency must be viewed alongside the potential risks. It is clear from the examples of ‘hallucinations’ listed above and the well-established limitations of LLMs which are commercially available, that it is of paramount importance to ensure that any party must review and be fully satisfied with the material which is to form the basis of its position.
The pace of AI development is moving rapidly, and the report of the House of Lords on “Large language models and generative AI” warns that an excess of caution in the early stages could later be viewed in a similar light to the so-called ‘Red Flag Act’ of 1865, which required a person to walk in front of the new motorcars waving a red flag!
The risks must be balanced against the substantial benefits which can be offered by AI solutions, and the increasing availability of these tools. Technological advancements of this nature will be quickly adopted by the market and in a competitive construction environment, it would be ill-advised to ignore them.
For further information in relation to this article please contact Clare Cashin or Paddy Mockler.