AI in Practice

By: Benjamin Kennedy, Director, Epiq

Artificial Intelligence (“AI”) is bringing unprecedented changes to how legal professionals search for the truth within data. Where armies of lawyers might trawl through documents in search of material, computer algorithms can now effectively make human decisions—helping get to the smoking gun faster with significantly less resources.

The holy grail of AI is when computers are able to learn by themselves. However, AI is not there, yet. AI cannot function without human interaction nor understand on its own. AI is unable to reason and understand causation, yet it’s very good at discerning patterns within text, which can be extremely helpful to better understand the content of documents.

AI software must be trained by knowledgeable users in order for the computer to determine what information is relevant to a particular inquiry. The software computes scores against electronic documents indicating a proximity to a known pattern. The user reviews the results and tells the system whether it yielded any relevant data. The system will then retain what it was taught and be able to identify data patterns and relevant concepts in future inquiries.

Technology-Assisted Review

One way AI is used in the litigation setting is through technology-assisted review (TAR). TAR has been used successfully for years and has been accepted by courts around the world.[1] TAR requires a subject matter expert to review a sample set of random documents to train the system on what is relevant. Once the system is trained, it scores each document on how likely that document is to be relevant to the case.

Instead of needing eyes on every document, TAR enables lawyers to get to the most relevant documents first, and not spend time reviewing documents that the system rates as having low relevancy scores. Not only is using TAR a more efficient way to get through large data sets, many studies showed that TAR is far more reliable than human reviewers.[2] While human eyes can tire and make mistakes, computer systems do not. In addition, machines do not have opinions like humans. Two reviewers may come to alternate conclusions about the relevancy of a document. TAR consistently tags all documents in accordance to the training it was provided.

TAR 2.0

Practitioners are turning to a new variant of TAR, Continuous Active Learning (“CAL”), or TAR 2.0. Unlike TAR 1.0, where the reviewers need to train the system when more documents are added, CAL continuously trains the system as documents are coded. CAL usually involves specifying an initial seed set of documents to be reviewed and coded. Once coded, these documents are used to train the learning algorithm, which scores the documents based on likelihood of responsiveness. The learning is ongoing as more documents are reviewed and coded, the CAL results are continually updated against the entire corpus.

The benefits of using CAL cannot be understated. Vast, sprawling document sets – which are familiar to anyone involved in litigation – can now be approached very differently. CAL supports a natural train of enquiry that continually learns from the human review. This aligns more closely with how humans explore their cases and increases their understanding of the issues as they continue to review. CAL maintains focus on the relevant aspects of documents and highlights other documents with similar features.

Using AI to Learn From the Past

Historical data reviewed in prior matters can hold great value. Using AI to learn from the past allows for a more efficient future. The logic is that it is impossible for a person or team to understand every document that has been reviewed and produced. By training AI models on the review performed in previous cases, AI can be used to identify if there are documents in a new set of data with characteristics similar to documents reviewed and produced in previous matters. The historic data effectively ‘bootstraps’ the AI learning for a new matter and relevant, irrelevant and unknown characteristics are revealed before human review.

Legal teams with data from historic cases can test the proposition that AI trained on previous matters has some value when reviewing new data. Without too much effort, legal teams can train the AI on documents produced in litigation as relevant and those not produced as irrelevant documents. The teams can withhold one or more cases from the training. The AI can then score the ‘unknown’ documents in the cases withheld from training. Did the model make any correct predictions by ranking relevant and irrelevant material? It won’t be perfect, but do the results provide better visibility into an unknown data set and will that help drive a more efficient and consistent review?

There are many potential applications for AI developed on historical information tailored to a client or type of investigation. AI can help with another dispute to immediately identify known material or assist the legal team to focus further training and review on unknown concepts. For corporations, AI built on historical data from matters run with different counsel could promote consistency across legal services panels and capture knowledge that is otherwise lost at the end of the engagement.

AI Beyond Litigation

AI has also proven useful in the mergers and acquisitions space. The technology is reducing the human effort with assessing areas of risk in agreements, as well as documenting a book of record for a transaction. The savings this provides from both a time and cost perspective are significant[3]

Aside from being effectively deployed in the legal arena, AI can be used in other corporate settings. For example, AI can identify language that may reveal potential risks to your organisation, such as employee theft or sexual harassment. This would allow an organisation to jump ahead of a whistle-blower or prevent a trade secret theft by gathering and evaluating electronic intelligence to predict issues.

Organisations can also use AI to determine when policies and protocols aren’t working as intended and then revamp company communications to ensure compliance. Placing AI closer to the point of information creation and communication provides an avenue to identify and address known risks as an event occurs. Monitoring email and network traffic in real-time for threats is a common practice, and AI monitoring of information creation and use is a logical next step.

AI will increasingly be part of the practice of law. Machines can’t yet learn for themselves, though we can begin to prepare the information we create today for AI to learn from. If legal teams want to embrace an AI future, they should consider what they do best, what data they maintain, and how it might be valuable in training algorithms to help them reach better and more consistent outcomes sooner.

About the author: Benjamin Kennedy brings fifteen years' experience consulting within government and private sectors on litigation support. His technical skills, passion for new technology and legal knowledge are valued by clients when developing cost-effective solutions for information review exercises of any size. Kennedy's primary focus is overseeing eDiscovery in Australia and New Zealand. He and his team assist in a variety of matters including, construction and commercial disputes, class actions, and forensic investigations that involve all manner of information sources. Kennedy regularly provides educational seminars for lawyers on eDiscovery, and speaks on advancements in the field.