How to Use Human-Centered AI in Legal Document Review

Cat Casey
Cat Casey

How to Use Human-Centered AI in Legal Document Review

Unless you have been living under a rock, you already know that the AI Renaissance has finally reached legal. Between the constant articles and posts about ChatGPT and the new cases popping up about generative AI like Dalle & Lensa, AI is everywhere. But what does that mean for legal practitioners today and how do you get beyond the hype of “AI-powered” everything?

Today Human-Centric AI is being used to conduct research, review documents, expedite digital discovery, aid due diligence and more all while keeping the Legal professional in the driver's seat. This blog digs into practical ways you can use AI today to make Document Review for eDiscovery less of a nightmare!

Email Threading

One of the most basic flavors of eDiscovery analytics, email threading identifies and organizes all the emails or comments in a single conversation to review them in order. This can accelerate review greatly, reducing time necessary to review and code similar documents. Grouping the conversation into a single unit improves review speed and consistency across a data set. This allows the human reviewing the data to more quickly connect the dots on long conversation stings.  

Conceptual Clustering

A data visualization that uses unsupervised machine learning (algorithms developed without human input), concept clustering presents groups of similar topics to a human reviewer who can then decide how to proceed. While the visualization is generated automatically using unsupervised learning, the human is in the driver's seat in terms of interpreting the data and determining what to do from a workflow stand point.

Legal teams can use the concept clustering visualization to better understand the key issues on case, to prioritize what topics get reviewed first or to validate the accuracy of proposed search terms. Can be used on its own or in concert with other AI tools or visualization as part of ECA, the Review or QC process of a document review.


Social Network Analysis

Another data visualization that uses unsupervised Machine learning, Social Network Analysis presents who is communicating to who with what frequency. Case teams can use this visualization to prioritize which custodians are reviewed, determine if any key people are missing from a data collection and/or exclude custodians who are not communicating with the relevant custodians of a matter.

Social Network Analysis can be used on its own or in concert with other AI tools or visualization as part of ECA, the Review or QC process of a document review. I have seen cases where social network analysis was used to prioritize custodians for large-scale reviews, outright eliminate custodians from scope who were not in communication with the key people of interest or uncover a previously unknown subject that ought to be included in scope.


Technology Assisted Review

Technology Assisted Review (TAR) is a type of supervised machine learning where human input refines an algorithm based on coding decisions. This type of Human-centered AI relies on a human to both build the algorithm and determine what to do with the output. 

The algorithm them promotes similar documents for review based on the human feedback. Legal professionals can use this to prioritize and accelerate a review (practitioners coding similar document can review faster) or upon reaching a level of statistical certainty can end a review without placing eyeballs on every document.

AI Models

AI models in eDiscovery document review are built out of a combination of algorithms that are supervised and unsupervised to identify things like specified issues, behaviors or even privilege. The model is trained using data sets and coding decisions and then learns throughout the new case based on the input from the case team. Like TAR these Human centric AI models are built with human guidance and subject to human determination on what is done with the output.  The models can be based on linguistic structure, emotional signals, behavioral intelligence, and case team insights from active learning.

    • Portable AI Models - Some AI models can be taken from one case to future cases to take the insights previously learned to accelerate time to insight.
    • AI Model Library - Some models are prebuilt using data scientists and data sets and the case team can apply it to their case to get a jump start.
    • Bespoke AI Models - Build AI models to identify specific issues as part of a complete AI driven solution. Then retain and reuse these models applying what you’ve learned (your decision IP) to other similar types of cases increasing your speed to valuable insights.
    • Model Layering - Multiple ai models can be turned on at a time with the system learning on them all in concert.

Entity Recognition/ Extraction

Another supervised learning process where the system learns to identify things like names of people, places, or companies, dollar amounts, job titles, account numbers, case/matter numbers, etc.

Entity Recognition is based on numerical patterns, linguistic patterns, and other AI models, this enables practitioners to refine the documents they review and prioritize information relevant to a case. Can be especially helpful in dealing with Personally identifiable information subject to a DSAR or post cyber breach notification.

Computer Vision

Much in the way that AI aimed to teach a computer to think like a human, Computer Vision aims to make a computer see like one. Used in image recognition and search, legal teams can use this function to identify specific potentially relevant images. or types of images that could be relevant. If there is an exemplar image you can use that to find similar images. Perhaps regard for a construction litigation case or human skin in an EEOC matter. The Human Centered AI surfaces likely visual content and specific types of images to empower the Human reviewer to make more informed decisions quickly. 

Conceptual Search

Concept search is another Human-centered analytic process that allows legal practitioners to search by an idea, theme, or example material. Concept search differs from concept clustering in that it is guided by an image, document, or string of text that a legal practitioner uses to search for other conceptually similar material.

Like concept clustering, this analytic tool can help practitioners better understand what is. going on in a data set, uncover key facts of a case and prioritize or exclude material based on the search results.

If the custodians used keywords or euphemisms that are not immediately apparent, you can uncover them with concept search. Concept searching is also extremely helpful if you are early on in a matter and not exactly certain about the specific language the organization or people of interest used but have a general idea of the type of concept you want to search.

Sentiment Analysis

A type of Natural Language Processing (NLP), Sentiment analysis uses signals in the language to uncover the feeling behind a document. The AI identifies signals of anger, excitement, optimism, or pressure. “I hate this company” vs “the pasta is lovely.” This tool is especially helpful if you are looking for hostile behavior, harassment, etc.

Emotionally charged language can be highly impactful in a variety of litigation and investigation contexts. From identifying language that involves pressure or coercion for a fraud case to emotional language indicating compliance or employment issues, this context building analysis helps surface potentially impactful communication quickly. The sentiment analysis empowers human reviewers to prioritize specific types of communication based on the type of matter they are handling. 

Machine Translation & Language Detection

Many eDiscovery tools now have AI powered Language identification and translation built in. Automated detection of language can help identify foreign language(s) within a document so you can shift to fluent members of the case team. Machine translation can be used to get an idea of the subjects covered in a document. This is great for prioritizing or eliminating documents, I would still advise having a human reviewer as a final pass for more potentially relevant documents because machine learning is not yet perfect.

The Artificial Intelligence Spectrum

Artificial intelligence helps streamline and accelerate all aspects of eDiscovery document review and can be applied to cases large or small.  Human-centered AI for eDiscovery workflows can help law firms and in house counsel gain insights from day one of a document review. New technology with ai capabilities is entering the market at lightning speed. The legal industry is increasingly adopting AI workflows to address the need to do more with less.

While some of the AI-powered legal tech is more advanced, there is a spectrum of AI solutions that can be incorporated into the review process. One or multiple AI powered legal tech solutions can be used to make eDiscovery document review less time consuming. Legal Departments and professionals can directly engage with a software provider that has an AI-powered Review platform like Reveal or gain access to the ai powered legal tech through legal service providers.

At the end of the day, legal professionals have a wide variety of Human-centered AI solutions for eDiscovery to make legal work easier and to find key insights faster.