Article Law360

Why You Should Leverage AI For Privilege Review

In this article for Law360, partner Michelle Six and Of Counsel Laura Riff discuss using AI for privilege logging and analysis.

Privilege logging and analysis can be one of the most time-consuming, expensive and challenging components of electronic document review.

Although the 2008 amendments to Federal Rule of Evidence 502 reduced the risk of broad subject matter waiver, the unintentional discovery of privileged content and the sensitivity of potentially privileged materials continues to inspire concern about the accuracy of privilege assertions.

Civil litigation can involve vast amounts of electronically stored information, and the cost of an eyes-on attorney privilege review can be enormous. Depending on the nature of the case and the number of agreed-upon custodians, the volume of data requiring privilege analysis can number in the terabytes.

And the time required for such a review can be exacerbated by the complexity of the files, such as handwritten materials, audio and video materials, and unstructured data.

Most manual privilege logging review projects leverage simple, familiar features.

For example, large-scale privilege review document platforms often employ persistent highlighting, electronically flagging — typically in some bold and attention-seeking color — certain names and terms that are more likely than others to occur in a privileged document.

Those privilege search terms can include both common words and phrases — for example, "direction of counsel," "attorney work product," and "attorney-client privileged." They can also include specific names of interest provided by the client — for example, attorneys and law firms the client has previously retained.

But the utility of persistent highlighting is limited. For one thing, a list of privilege search terms that includes attorneys with common names is likely to be overinclusive, highlighting names that belong to nonattorneys. And certain keywords may frequently appear in boilerplate language, like email footers and disclaimers, and so they are not always reliable indicia of privileged materials.

Additionally, privileged content can certainly be found in emails without any attorney present, or in draft documents that may not have been accurately labeled, so persistent highlighting alone cannot be relied on to flag all potentially privileged language.

To quickly cull large data sets for privileged documents and communications, more advanced technological tools are often required. Indeed, for the attorney willing to experiment with them, the tools can prove to be quite valuable and may even identify previously unknown attorney names or surface unexpected privileged content.

Yet the thought of relying on artificial intelligence for privilege logging and review still makes many lawyers nervous. Despite the availability of robust clawback agreements under Rule 502(d) and the various ethical and rules-based obligations regarding the disclosure of privileged files, nothing makes attorneys and their clients more anxious than the production of a privileged document to an opposing party.

For better or worse, many lawyers still believe that a traditional, document-by-document review with multiple rounds of additional attorney quality control checks is the best way to avoid privileged material being included in production. But AI should not be feared — and in fact it can be mined and coupled with attorney review to further safeguard privilege and locate protected documents more consistently, often with improved accuracy and efficiency.

Certain AI tools can be leveraged to analyze document metadata as well as the context or timing of particular communications to assess hidden patterns — and anomalies in those patterns — that may occur repeatedly in a client's privileged materials. These tools tend to be most successful when they combine human inputs and computer learning to identify potentially privileged information.

The two primary objectives when using AI technology to scan for privilege are to (1) reduce the number of documents that need human review by prioritizing documents that are most likely to be privileged, and (2) eliminate from review those documents that are least likely to contain privileged material. While execution of search terms can help with both goals, machine learning models can refine the process and patch the gaps where search terms alone may be inaccurate or — more likely — incomplete.

In practice, there are numerous options for employing AI. One AI solution scores content to prioritize and shorten privilege review. This process uses AI features — which are often proprietary to the vendor — to evaluate the text and metadata of each document and assign it a score. Attorneys can use the scores to prioritize documents for review, ending review when the likelihood that the remaining documents contain privileged material is below a threshold percentage.

These AI features often analyze context and patterns in document metadata, rather than simply identifying hits on privilege search terms. The technology can thereby provide a low score to a document that contains a privilege search term, but for which the context indicates that it is unlikely to be privileged. And conversely, it can detect documents that are likely to be privileged more accurately than simple search terms can.

For example, certain AI features can recognize out-of-office alerts, documents with enormous distribution lists — like mass marketing emails — and communications in which privilege search terms appear solely in boilerplate content like company disclaimers.

Such documents can be assigned a low score or even eliminated from review, thereby reducing the total amount of review time and decreasing the likelihood that a reviewer will inaccurately code such documents as privileged.

Other features for scoring documents can help ensure that documents involving attorneys and law firms are spotted. These features analyze the email text and metadata to identify all the ways an attorney might appear on a document — like nicknames and other name variations, and email addresses and entity affiliations, including distribution groups.

Similar technology can assess the data to determine the roles of people involved in a communication in order to find emails where paralegals or legal assistants are communicating with the client on behalf of the attorney. In so doing, the AI can help identify potentially privileged content that may not have been apparent as such to a reviewer.

Another AI solution that may improve privilege review is known as iterative active learning. This process uses modeling to identify and prioritize potentially privileged content. Manual review of the prioritized content allows the model to continuously learn from reviewer coding.

In other words, as reviewers code the prioritized material, the model improves based on that coding and uses that information to reprioritize documents, constantly directing the documents that are most likely to be privileged to the top of the review. The model can even help the attorneys determine when it is appropriate to stop reviewing.

Certain advanced AI technology can also make use of privilege coding on reviews from previous litigation for the same client, analyzing that coding to reduce the amount of attorney input needed from the beginning. Some companies have prior privilege coding on hundreds of thousands, if not millions, of documents in old databases. That information can be used as a starting point for training a privilege model, or even to eliminate from review those documents that had been coded in the previous litigation.

In sum, while no technology is ever likely to make the privilege logging process particularly easy or fun, when faced with a massive privilege review, attorneys should consider the utility of AI. Traditional software, search terms and even the most experienced of attorney reviewers can mistake or ignore privileged content that can defensibly be withheld from discovery.

Investigating and experimenting with advanced tools and models may be the most effective way to shore up weaknesses in the traditional review process, and leveraging these advances in technology may well make privilege reviews more efficient, more accurate and less costly.