Impact of AI on Handling Disputes and e-Discovery

Tom Whittaker and Ravi Tandon outline how AI is impacting litigation, before looking in depth at its impact on eDiscovery.

1 – Trends

AI is becoming increasingly relevant not just in what disputes are about but also in how they are handled.

On the substance, there are still few reported cases. However, anecdotally, AI-related disputes are on the rise. To assist with the future caseloads, the UK Jurisdiction Taskforce – chaired by the Master of the Rolls, Sir Geoffrey Vos – has an on-going project to produce a legal statement on harms caused by AI. 

AI is starting to transform the way disputes are conducted, though there are both risks and opportunities. While there have been highly publicised failures, like citing fake cases in Court, this should not obscure the potential benefits that are available:

• Legal research: Some tools can search real cases, legislation and practitioners’ texts. The starting point is that users must approach the results with caution and critical thinking. But AI tools can help accelerate identification of a “launch pad” case – i.e. a key recent authority that summarises the law and the basis for finding important cases and linked commentaries.
• Disclosure & case analysis: Technology produces voluminous documentation for review, but struggles to filter them effectively. AI tools are making progress in helping lawyers identify what matters. These issues are considered in more detail in the second part of this article below.
• Submissions and drafting: Lawyers aren’t out of a job yet. Indeed, where some litigants use AI to cut corners (like drafting a letter in 10 seconds flat), there is particular value in a considered response. Still, AI tools have clear uses, for example, in editing or rephrasing for clarity and conciseness, and in producing first draft chronologies.
• Decision-making or facilitating: Given time and cost can make traditional litigation non-viable, it is unsurprising that AI-assisted resolutions (like online dispute resolution) are being explored. 

If lawyers become cautious, informed adopters of AI tools, there’s scope for greater efficiency and better case outcomes. Knowledge of the tools’ shortcomings will also be invaluable when fighting or defending cases, including those about when AI causes harm. This article will now spotlight the vital role AI plays in relation to e-discovery.

2 – How can GenAI be used in eDiscovery?

Discovery (or disclosure) is process by which parties identify, collect, review and produce documents to another party or a legal forum, such as court. eDiscovery (or eDisclosure) is the common way of referring to electronic discovery, reflecting that most discovery involves electronic data or use of electronic systems.

However, eDiscovery can be challenging. In this article, we explore why and how GenAI could help and set out some practical points for it to be used effectively.

eDiscovery is getting more challenging

Data volumes have always been an issue, whether for the review of hard copies or electronic documents. The ability to search consistently and effectively could be difficult, especially depending on how data had been scanned or processed. The complexity of identifying where potentially relevant data may be stored and how to collect it grew as a challenge as organisations and individuals started to use multiple electronic systems. The consequence of these issues, and more, is that the time and cost involved with a Discovery exercise risked being disproportionate and unreasonable relative to the proceedings, risking the procedural timetable, and ultimately not enabling the parties and court to understand the evidence. 

These challenges appear to be growing in scale and complexity in recent years. Take for example a set of proceedings concerning what the board of an organisation – let’s call it BoardCo – discussed and decided from 2015 to 2025:

• 2015 – Shared drives and email dominated. Version control was manual, and notes were taken physically.
• 2020 – Hybrid working introduced Teams, document management systems, and more communication channels.
• Today – AI bots record meetings, SharePoint automates version control, and encrypted apps supplement email. Pen-and-paper notes still exist, but outdated governance leaves retention and integrity issues unresolved.

To add to this, BoardCo may not have updated its data governance practices in advance, affecting what documents were created, for how long they would be retained and how data integrity was maintained. This fragmented data landscape makes identification, collection and review far harder than a decade ago.

Use of predictive AI in eDiscovery

The courts have, to an extent, sought to keep apace. The procedural rules for Discovery in the Business and Property Courts in England and Wales were subject to a working group in 2019 and updated (now known as PD57AD). However, other aspects relevant to Discovery (and eDiscovery) have not seen updates. Many High Court cases use the procedural rules in place prior to PD57AD, namely CPR31, which have not been updated. Judicial reviews are subject to the duty of candour, but the Treasury Solicitor Guidance on duty of candour remains that from 2010. The TECBAR¹ ‘eDisclosure’ protocol is from 2015. This is not to say that they are all outdated. However, each is to an extent; they do not fully anticipate modern eDiscovery challenges or expectations.

Of course, forms of ‘AI’ in eDiscovery are not new. Predictive coding, that is the use of supervised machine learning models trained on human coding to identify other documents likely (or not likely) to be relevant (known as Technology Assisted Review or Continuous Active Learning), has been accepted by courts in England & Wales since 2016 (see Pyrrho Investment Limited and others v MWB Property Limited and others [2016] EWHC 256 (Ch)). PD57AD states that parties should explain why they do not use technology assisted review, especially for reviews of more than 50,000 documents and where large volumes will therefore require manual review.

Many practitioners will also have experience of using unsupervised machine learning, for example, to search by concept to identify conceptually similar documents even where they do not contain the same words, clustering by concepts to help with high-level analysis, and expanding keywords to identify conceptually similar terms found in the dataset.

GenAI is the next stage in the evolution of technology in eDiscovery. 

The capabilities of GenAI promise the potential to analyse data faster (at lower cost), more accurately and consistently, and uncovering greater insights (at a document and a dataset level) than had been realistically feasible in the past. 

GenAI tools also permit more expressive queries, such as, “find all documents and tag them responsive where John and Jane discussed transactions with respect to kickbacks”.

In providing these potential advantages, GenAI is different to predictive coding. For example:

1. GenAI relies on foundation models that are non-deterministic, meaning the same inputs may not always result in the same outputs. They are often described as a “black box”. This makes perceived error correction, explainability, as well as ability to control the outputs of the model hard, although some tools are available that may limit the variability in output. In contrast, predictive coding is deterministic: for example, a statistical model predicting whether a document is relevant (or not) to a production request may have clearer parameters on the confidence interval within which it can generate and classify the data.
2. Foundation models are designed to analyse large volumes of data. They are not designed specifically for eDiscovery use cases in mind, like predictive coding models.
3. GenAI is likely to rely on third party models, fine-tuned by AI system providers, adding a further reason for lack of explainability. This may place a greater emphasis on testing workflows and validating results.  Predictive coding algorithms are, however, more widely understood and accessible.

Procedural approaches

Most procedural rules are agnostic about technology used in eDiscovery; they don’t specify that a form of technology must, or must not, be used.   

Instead, procedural rules often do a couple of things. First, they provide overriding objectives or principles. These can directly or indirectly affect why technology is used or how. For example, in the case of the Civil Procedure Rules, which govern a range of civil litigation: the Overriding Objectives (CPR1) includes the objective of dealing with cases justly and at proportionate cost; CPR31 includes principles for eDiscovery; PD57AD specifies principles to be applied and duties for parties and their lawyers. Second, sometimes the procedural rules actively require consideration (but not use) of technology to assist review, such as PD57AD.

GenAI use cases in eDiscovery

What are the potential use cases of GenAI in eDiscovery? If BoardCo were involved in a dispute regarding their decision-making, examples include:

• Classification: early case assessment (flagging helpful or harmful documents for case risk assessment), relevance or issue classification (with tagging to issues and relevant text extracts), and identification of privileged material;
• Document summarisation: including chronology creation (such as summarising BoardCo’s board papers and creating a chronology or specific issues);
• Media processing: such as deciphering handwritten notes, transcribing audio and video (finding the few seconds where a topic was discussed out of years of discussion perhaps), or detecting objects in images (for example, a specific person).
• Comparison and consistency checks:  including comparing statements, categorising documents (e.g. having a list of similar documents, removing redundancies and tracking lineage), and detecting anomalies and inconsistencies (e.g. flagging board comments that are incompatible for further review).

These can be powerful tools, but they are not a ‘silver bullet’. They may be best deployed for specific tasks as part of a broader eDiscovery process, guided by legal and technical expertise, as well as knowledge of the facts.

What to consider

First, view Discovery holistically using a framework like the Electronic Discovery Reference Model (EDRM). It includes stages such as: considering BoardCo’s information governance to understand where documents may be stored; identifying which custodians and document repositories are likely (or unlikely) to hold potentially relevant documents; taking steps to preserve such documents or repositories, such as the board papers on the shared computer drive or the emails of the board members; collecting and processing data so that it can then effectively be analysed (at a high-level) and reviewed (at a document-level) so that it can then be produced to the other party.

The EDRM is a useful framework for a range of types of matters – from High Court litigation to judicial reviews, regulatory investigations, inquiries, and more.

Viewing Discovery holistically is necessary because it emphasises how each stage has legal, technological and practical issues specific to that stage. Further, it lets us view eDiscovery as an iterative process. The steps taken on the left-hand side – such as understanding BoardCo’s information governance to identify its data sources and how they can be accessed are foundational to subsequent stages, such as identifying which of BoardCo’s sources do (and do not) require preservation and collection. Conversely, issues may be identified at later stages which identify potential problems with earlier stages that require revisiting, for example, an apparent gap in data collected from BoardCo which requires investigation. Finally, what is implied by the EDRM, but is useful to state explicitly, is that the Discovery process is for specific legal objectives, such as to identify documents relevant to specified issues, and subject to legal parameters, such as limitations to the scope of search and costs budgeting. Each stage of the process, and the overall process, needs to be calibrated to the context accordingly.

What this means in practical terms for GenAI is that the tools are unlikely to solve all the challenges of Discovery. Instead, they are likely to be used for a specific task at a specific eDiscovery stage, where anticipated use of GenAI will help; they will continue to require legal experts to determine the objectives, strategy and parameters of the exercise. 

Second, what GenAI is used for is subject to multiple factors. For example:

1. Legal context – What is the issue to be addressed? As shown above, BoardCo could use GenAI for a variety of purposes. However, just because GenAI could be used does not mean it should be. For example, for small datasets, human or predictive review may suffice.
2. Available technology – What tools are available and at what cost? There are different pricing models, and the level of testing and validation required will impact on lawyer and technical costs.
3. Data quality – Some datasets may not be suitable for current GenAI tools.
4. Client requirements and jurisdictional limits – While AI tools increasingly have established confidentiality and data protections, some clients, such as those in the public or defence sectors, may require document review on a server (rather than cloud), limiting what AI tools are available. The relevant jurisdiction may also affect which foundation models can be used and, consequently, which tools can be used.
5. Standards of accuracy – How good is “good enough”? This will depend on why the GenAI tool is being used and in what circumstances. Generally, it is not a question of how good the AI tool is alone, but rather how good the AI and humans are when working together.

It can be helpful for both parties to agree an approach in advance to reduce ambiguity, improve fairness and prevent later disputes. Indeed, for High Court actions under PD57AD, a high degree cooperation is expected, and unilateral action is deprecated (see AAH Pharmaceuticals Ltd v Jhoots Healthcare Ltd [2020] EWHC 2524 (Comm)). Key considerations include:

1. Model Selection, Training and Bias Testing – Parties should be prepared to understand and potentially explain some of the key characteristics of the proposed underlying models, particularly if standard models are not used. There are now established systems and benchmarks that help reveal various biases inherent in different models such as those developed by OpenAI, Anthropic, and Google’s Gemini.
2. Model Architecture – Many large language models use multi-stage architectures to reach final outputs. These designs can influence how and why certain outputs are generated, particularly when the model draws from distinct expert pathways for different domains. Even a basic understanding of these architectural elements can be helpful in explaining variability, identifying edge cases or troubleshooting unexpected behaviour.
3. Prompt Disclosures – Prompt engineering plays a critical role in shaping GenAI outputs. Even minor substitutions or rewordings in prompts can lead to materially different results. As such, best practices should include logging and sharing prompts when GenAI outputs are being used for regulated or contested processes. Clear documentation of the prompt history ensures traceability and reproducibility.
4. Algorithm Disclosure and Workflow Boundaries – If the algorithm itself is not publicly available, it can become especially important for both parties to understand and agree upon the parameters and operational workflow used in conjunction with the model. There is considerable flexibility that can be built into preprocessing, context selection or retrieval steps, and this can unintentionally favour one party over another. Establishing a clear and strict framework to define the boundaries of what the model and associated systems can and cannot do can significantly enhance trust and potential bring a degree of determinism to otherwise probabilistic processes.

We are still at relatively early stages. There is clearly interest and use. There are multiple GenAI tools on the market and there is widespread report of their use including in High Court litigation including for helping conduct court-ordered Discovery. However, it is difficult to get a comprehensive picture of how GenAI is being used in Discovery and, whilst judicial guidance on the potential use of AI notes that that there are many potential uses in litigation, we have not yet had a court judgment or guidance specifically about the use of GenAI in the Discovery context.  However, with the way things are moving, it is only a matter of time that we see wider spread use of GenAI in Discovery. What that will look like and how will it work? Time will tell but it may be relatively soon.

1. Technology and Construction Solicitors’ Association ↩︎

This article is also available in the special AI issue of Computers & Law, which is available to download here.