FTC Issues Rigorous Requirements on Using AI-Assisted Review or Technology-Assisted Review to Respond to Investigational Subpoenas

Recently, the Federal Trade Commission (FTC) issued investigational subpoenas — in both a Civil Investigative Demand and a merger Second Request investigation — that include far more rigorous requirements for using either technology-assisted review (TAR) or artificial intelligence (AI) to identify documents responsive to them. The instructions define AI-assisted review for the first time — a recognition of its increasing use by parties. But the significant conditions placed on the use of AI-assisted review are likely to discourage parties from using it, at least if parties are required to follow the new requirements to the letter. The new requirements will also make it much more cumbersome for parties to use TAR than was the case previously. The new requirements include:

• Expansive Up-Front Disclosures: While FTC subpoenas previously included instructions requiring submission of a written description of any technology used to identify or eliminate responsive documents, the new instructions require significantly expanded disclosures which must be made at least one week prior to the start of any model training or review. These disclosures include a detailed description of the technology and review workflow used and identification of all Subject Matter Experts (SMEs) who will be involved in training, each of whom must be a full-time employee of the firm managing the response familiar with all relevant issues in the investigation. Parties must also disclose reports on the types of documents that will be run through TAR or AI and the file types that are excluded for special handling, with specific requirements applicable to ephemeral and text messaging files and linked (modern) attachments. If AI-assisted review is used, parties must also disclose the tool, version, provider, and underlying model, as well as “AI vs. human agreement rates” for a statistically valid sample of documents, broken down by classification category.
• Prescriptive Metrics That Must Be Achieved: As part of the initial disclosures, parties must also identify the statistical metrics that will be available during training and validation. The FTC goes further, however, and requires the parties meet certain thresholds for those metrics. In addition to requiring a recall of at least 75% (with a margin of error of plus-or-minus 3% at a 95% confidence level), the FTC requires the production to meet a precision level of at least 75% — the first time the FTC has required a set precision and is a departure from agency practice, which has previously focused on recall. The cutoff score used for responsive documents must also be disclosed for the first time, and must be no less than 0.50. These metrics may be difficult to meet because precision is highly dependent on the specifics of the document collection and its richness, and there is a tradeoff between precision and recall.
• Requirement That Collections Must Be Complete Before Starting: “Only once collection from all custodians is complete” and the FTC has a chance to review and confirm the required disclosures, will parties be permitted to begin training and run the predictive coding or AI technology. This suggests that training cannot begin until agreement is reached on a full custodian list and all custodians are collected, which will likely delay parties’ ability to begin using TAR or AI. This requirement seems aimed at discouraging TAR 1.0 in favor of TAR 2.0/Continuous Active Learning which is not dependent on having a substantial corpus of custodial documents collected before beginning training.
• Disclosure of Prompts, Relevance Scores, and Rationales: If AI-assisted review is used, parties are required to disclose all prompts or review protocols provided to the AI. Relevance scores and relevance rationales for each document produced must also be provided, if available. Parties may be reluctant to produce these materials, with some parties having argued that such materials are protected by attorney work product protection because they reveal attorney legal strategies about the case.
• Requirements for Model Validation: Parties are required to identify all control sets and training sets to the FTC, including a custodian and file type breakdown of any seed or training sets. The instructions also require a pre-production validation process utilizing an elusion sample that must be provided to the FTC (the sample itself, not only summary statistics) at least two weeks prior to compliance. While parties typically offer the FTC the opportunity to review a “null set” for validation when using TAR, the updated instructions now formalize that process. The instructions also impose an express duty to retrain if the FTC determines that validation reveals deficiencies, and a duty to make supplemental productions if the review process changes as a result of the FTC’s review of the validation sample.
• No Pre-Culling Without FTC Consent: Parties are not permitted to use analytics (including email threading or search terms) to reduce the population of documents prior to deduplication or the application of TAR or AI, absent agreement with the FTC on the specific process to be used.
• Required Calendar of Productions: Parties must provide a calendar of “anticipated productions” within a week of the start of TAR or AI-assisted review training, identifying dates and volumes of rolling productions. This is a new requirement involving the FTC in details of the parties’ production scheduling, which was not the case previously and is something often difficult to predict.
• No Supplemental Responsiveness Review: The updated instructions require that there be no use of manual review or search terms to eliminate documents identified as responsive by predictive coding, “except to the extent required to identify privileged information, including PII, SPII, and SHI.” This formalizes a position that the FTC (and DOJ) have historically taken, which parties often resist due to their desire to weed out sensitive documents that are in fact non-responsive that TAR or AI identified incorrectly as responsive.

These updated instructions insert the FTC into the detailed inner workings of parties’ review and production processes, which many parties will be resistant to do. If parties follow these instructions without negotiated modifications it would likely increase the time and burden required to use either AI or TAR and may prolong the compliance process. This will also give the FTC much more control of compliance timing, allowing it to delay the start of review through the required disclosures and approval process, which cannot begin until collections are complete.

To avoid these requirements, parties may revert to using search terms and manual review (instead of TAR or AI-assisted review), at least for smaller review populations. Parties may also negotiate with the FTC and try to reach acceptable compromises on these requirements, or they may decide to proceed under a protocol they think is reasonable while providing a statement of reasons for noncompliance. That approach carries risk, however. In a Second Request, the FTC may dispute that the party has substantially complied, which could delay the parties’ ability to close the transaction. It is not clear whether a federal court would mandate compliance with these requirements should the issue be litigated as long as the party followed a reasonable search and review protocol.

© Arnold & Porter Kaye Scholer LLP 2026 All Rights Reserved. This Blog post is intended to be a general summary of the law and does not constitute legal advice. You should consult with counsel to determine applicable legal requirements in a specific fact situation.