Legal Alert: The Tax Court Approves the Use of Predictive Coding

On September 17, the U.S. Tax Court, in Dynamo Holdings LP v. Commissioner, 143 T.C. No. 9 (Sept. 17, 2014), held that a taxpayer could use predictive coding, over the objection of the Internal Revenue Service (IRS), to identify relevant electronically stored information (ESI) for production. This is the first Tax Court case to address the use of predictive coding in response to a discovery request.

Background – What Is Predictive Coding?

Predictive coding, also referred to as Technology Assisted Review, is a method to expedite the review of large volumes of electronic documents or information. At the outset of the review, the predictive coding platform identifies the distinct qualities of each document in the data set. The platform then generates a random, statistically sound sample or seed set of 1,000 documents for review. A senior-level reviewer codes the “seed set” of documents for relevance, and the reviewer’s decisions are fed back into the platform. The platform’s algorithm “learns” from the coding decision made by the reviewer to separate relevant documents from irrelevant documents based on the identifying characteristics of the documents. The algorithm codes the unreviewed documents as responsive or non-responsive or scores them numerically based on the predicted level of responsiveness. Tests are performed to ensure to a statistical certainty that the documents were accurately coded. The predictive coding platform is then asked to generate additional sets for review using what it has cumulatively learned from the reviewer’s coding decisions. When the algorithm generates a set of documents that achieves satisfactory rates of recall and precision, the manual reviewing stops, and the platform applies its algorithm to the greater set of documents. The resulting set is then ready for production (unless there are attorney-client privilege or work product concerns).

If there are privilege or work product concerns, human reviewers examine the resulting set to determine which documents may be withheld. This method greatly reduces the number of documents the human reviewer must examine, thereby decreasing costs. An alternative method is to use keywords to generate the first seed set. Some say that doing so expedites the iterative process described above, although speed might be achieved at the risk of missing categories of documents that were not caught by the keyword search.

Background – The Legal Context

The Tax Court rules were amended in 2010 to include ESI provisions essentially similar to the Federal Rules of Civil Procedure. See Tax Court Rule 70(a). The IRS has provided guidance to its attorneys regarding the preservation and discovery of ESI. See, e.g., Chief Counsel Notice 2012-17 (Sept. 13, 2012). Neither Tax Court nor IRS guidance addresses predictive coding.

The principal substantive issue involved in Dynamo was the tax treatment of transfers of property to the taxpayer from Beekman Vista, Inc., a corporation wholly owned by a Canadian citizen, Delia Moog. The taxpayer characterized these transfers as loans; the IRS characterized them as disguised gifts from Ms. Moog to her daughter and nephew, who indirectly owned the taxpayer. The case has a long procedural history, including a summons enforcement proceeding which originated in the Southern District of Florida and ultimately reached the U.S. Supreme Court. United States v. Clarke, 573 U.S. ___ (June 19, 2014). The Court in Clarke held that the taxpayer was entitled to an evidentiary hearing to determine whether the summons against its chief financial officer should be enforced. It does not appear that any further proceedings have occurred in the District Court after the Supreme Court’s decision.

The IRS filed a motion to compel production of ESI contained on two specified backup storage tapes or, alternatively, the tapes themselves (either originals or copies). The taxpayer requested that it be permitted to use agreed search criteria in combination with predictive coding to identify relevant documents on the two tapes. The taxpayer proposed to use human reviewers to identify privileged and protected confidential information (for example, health information), and then to produce relevant, nonprivileged ESI. The taxpayer estimated that, without predictive coding, the cost of reviewing 3.5 million to 7 million documents on the two tapes would be more than $500,000. The use of predictive coding would lower the number of documents reviewed to 200,000 to 400,000 and the cost of review to $85,000.

The IRS objected to the taxpayer’s request to use predictive coding, arguing that the taxpayer should instead produce the tapes in their entirety, subject to a “clawback” agreement under which the disclosure of any privileged or confidential information on the tapes would not be a waiver of any privilege. The IRS asserted that it needed the backup tapes to review metadata and verify the dates on which documents were created, as well as to ascertain all transfers to the taxpayer by Beekman. The IRS also asserted that predictive coding was an “unproven technology” which would not reliably identify relevant documents. On March 11, 2014, the Tax Court held a hearing at which the parties presented expert witnesses on the use of predictive coding.

The Tax Court’s Decision

The court noted that neither party appeared to contest the fact that relevant information existed on the backup tapes. The court did not consider it appropriate to order the taxpayer to give the tapes to the IRS in their entirety subject to a clawback agreement. Although the parties could consent to such an agreement, the court was not willing to compel the taxpayer to do so.

Having decided that only relevant ESI needed to be produced, the court sought to find a “happy medium” that would require the taxpayer to produce relevant nonprivileged ESI but would reduce the costs of doing so by reducing the universe of documents that had to be manually reviewed for privilege. The court relied on a leading article on predictive coding by Magistrate Judge Andrew Peck, the substance of which was adopted by the court in Moore. Moore v. Publicis Groupe SA, 2012 WL 1446534 (S.D.N.Y. 2012). Other courts have followed the example from Moore, recommending predictive coding as a method of producing responsive ESI in discovery. In F.D.I.C., the court ordered the parties to consider predictive coding. F.D.I.C. v. Bowden, CV413-245, 2014 WL 2548137 (S.D. Ga. June 6, 2014). See also, In re Actos (Pioglitazone) Products Liab. Litig., 6:11-MD-2299, 2012 WL 7861249 (W.D. La. July 27, 2012) (approving the use of predictive coding); Hinterberger v. Catholic Health Sys., Inc., 08-CV-380S F, 2013 WL 2250603 (W.D.N.Y. May 21, 2013) (recommending predictive coding to parties who showed a lack of progress in resolving issues pertaining to ESI).

The court disagreed with the IRS assertion that predictive coding was “unproven technology,” noting that the technology industry considered predictive coding to be widely accepted for both limiting e-discovery to relevant documents and producing ESI without undue burden. Because the taxpayer had represented that it would retain electronic discovery experts to meet with IRS counsel and to devise a search acceptable to the IRS, the court saw no reason why the taxpayer should not be able to use predictive coding. The court also stated that, in the event the IRS could demonstrate that the taxpayer’s approach did not produce all relevant documents, the IRS could renew its motion to compel.