A federal jury indicts on insider trading without a tipster, cooperating witness, or wiretap. The case began when the SEC’s Market Abuse Unit used algorithms to sift billions of trading records and identify suspicious patterns across 12 corporate transactions over six years. In another courtroom, police searched a murder defendant’s apartment based on an AI-powered facial-recognition match generated by software whose error rates, training data, and methodology remain shielded as proprietary. These are not hypotheticals, and they mark a new evidentiary and constitutional battleground for criminal and white collar defense lawyers.
The federal government’s use of artificial intelligence (AI) in criminal investigations is expanding faster than the legal framework for challenging the evidence these tools produce. On Feb. 11, 2026, FedScoop reported that the Department of Justice’s (DOJ) 2025 AI Use Case Inventory listed 315 entries—up from four two years earlier.
These entries included 114 “high-impact” uses whose output may materially affect rights or safety. Defense counsel who do not understand these tools risk litigating with one hand tied behind their backs.
The Regulatory Landscape: From Mandate to Acceleration
The federal framework has shifted from caution to acceleration. Executive Order 13960 first required agencies to inventory AI applications in December 2020.
President Joe Biden’s Executive Order 14110 later imposed safety-testing and risk-management obligations, including for AI in the criminal justice system. President Donald Trump revoked that order on Jan. 20, 2025, through Executive Order 14148, and soon issued Executive Order 14179, “Removing Barriers to American Leadership in Artificial Intelligence.” OMB Memorandum M-25-21 implemented that approach and defines “high-impact” AI as a system whose output serves as a principal basis for decisions with legal or safety effects, even with human oversight.
Current OMB guidance requires annual AI use inventories, public reporting of releasable uses, reporting of certain commercial AI tools, individual reporting of high-impact use cases, and disclosure of waiver information.
The inventories are important, but incomplete. M-25-21 permits agencies to exclude AI uses whose disclosure would be “inconsistent with applicable law and government-wide policy.” In March 2026, the Government Accountability Office (GAO) reported that that the Internal Revenue Service (IRS) inventory had omitted several AI-enabled tools contracted to help build criminal cases.
Published inventories therefore should be treated as a starting point, not the full map of the government’s investigative technology.
The Government’s AI Arsenal
Based on these reports and inventories released by various government agencies, the government’s AI arsenal already spans predictive analytics, facial recognition, pattern detection, financial-transaction analysis, and e-discovery. The 315 entries in DOJ’s 2025 AI Use Case Inventory represent a 30.7 percent increase from 2024 and include 188 active use cases and 114 “high-impact” uses. Law-enforcement applications account for roughly 62 percent of deployed
DOJ use cases. The FBI alone reported 50 AI use cases in 2025—up from 19 in 2024—including 27 supporting law-enforcement activities and nine classified as high-impact, involving functions such as identity verification, biometric matching across large datasets, language translation, and automatic vehicle identification.
DOJ’s inventory also describes tools used to accelerate triage, discovery, investigations, and reporting, including systems that identify patterns in financial transactions and cryptocurrency activity.
According to the GAO, while the IRS reported 126 active AI applications across audit selection, fraud detection, taxpayer services, and operations; 65 use cases were deemed either too sensitive for public reporting or exempt research-and-development efforts.
The Securities and Exchange Commission (SEC) provides the clearest white-collar example. Its 2025 inventory lists 60 AI use cases, more than double the prior year. ARTEMIS—the Advanced Relational Trading Enforcement Metric Investigation System—analyzes billions of equity and options trades across issuers and traders.
In May 2026, the SEC charged 21 defendants in an alleged insider-trading scheme spanning 12 corporate transactions over six years, an investigation in which the Market Abuse Unit used data-analysis tools, not a tip or cooperating witness, to generate its leads.
In any federal investigation, defense counsel should demand identification of all AI systems used, seek discovery into algorithmic tools at each stage, and challenge vendor trade-secret objections through protective orders or expert-only access.
The Evidentiary Battleground: Proposed Rule 707
The admissibility framework is still catching up. On June 10, 2025, the Committee on Rules of Practice and Procedure approved proposed Federal Rule of Evidence 707. If adopted on schedule, it will take effect Dec. 1, 2027.
The rule would allow machine-generated evidence offered without an expert only if it satisfies Rule 702(a)-(d), while excluding “simple scientific instruments.” Until then, AI-generated evidence will continue entering federal courts without a dedicated rule.
Rule 707 would be useful but limited. It applies only when the proponent acknowledges machine-generated evidence, contains no mechanism to compel disclosure of concealed AI provenance or deepfakes, and does not solve problems of training-data integrity.
Nor does it eliminate bias simply because a human expert presents algorithmic results. And because the rule is federal and incorporates Daubert through Rule 702, it does not directly govern Frye jurisdictions such as New York.
New York practitioners should be ready to fill that gap. The New York court system’s Advisory Committee on Artificial Intelligence and the courts recommended in 2025 that courts consider whether Frye should be modified for AI-generated evidence. Until then, counsel should argue that the government must show general acceptance of the specific AI tool and methodology—not merely of artificial intelligence in the abstract.
Recent scholarship and practice reinforce the need for gatekeeping. Researchers Pulvino, Sutton, and Naddeo reported in the Columbia Science and Technology Law Review that ChatGPT displayed a “consistent prosecutorial default bias,” recommending prosecution even in weak cases or where constitutional violations appeared. Law360 likewise reported in June 2026 that False Claims Act subpoenas had been issued based on misinterpreted data-mining results.
Defense counsel should press for reliability hearings requiring validation data, error rates, and bias testing before machine-generated evidence reaches the jury.
Fourth Amendment and Brady Obligations
AI surveillance also raises Fourth Amendment questions. In Carpenter v. United States, 585 U.S. 296 (2018), the Supreme Court held that acquiring comprehensive digital records revealing the “privacies of life” can be a search even when third parties hold the data. AI-enhanced license-plate readers, facial-recognition systems, and predictive-policing tools amplify that concern by enabling the government to aggregate, identify, and target at scale.
The Congressional Research Service identified two emerging issues in May 2026: whether AI-enhanced surveillance triggers Carpenter-style protection, and whether AI-generated identifications can establish probable cause. Courts are beginning to confront both. In State v. Tolbert, 2025-Ohio-4469 (8th Dist.), the court suppressed evidence from an apartment search after finding that the warrant rested on an unreliable AI facial-recognition identification.
AI filtering also complicates Brady. If an algorithm decides which materials prosecutors review or produce, it may suppress exculpatory information without anyone learning of the omission. A 2026 Columbia Science and Technology Law Review study by Hartline, Wexler, Shan, and Sun found that technology-assisted review can hide Brady material and recommended separate searches for inculpatory and exculpatory evidence.
Counsel should move to suppress AI-powered surveillance where aggregating data operates as a warrantless search; seek disclosure of the tool used and the circumstances of any identification; and demand error rates, training-data composition, validation methods, and reliability information under Brady and Rule 16(a)(1)(F). Where prosecutors resist, counsel should argue that opaque algorithmic systems implicate due process and confrontation rights.
The Practitioner’s Playbook
AI literacy is now a core defense skill. Counsel must understand how systems like ARTEMIS analyze trading data, how machine-learning models generate risk scores, and how predictive analytics identify enforcement targets. Technical experts should be retained early enough to shape discovery, admissibility motions, and constitutional challenges.
The first battleground is discovery. Defense counsel should seek the data, error rates, validation methods, and e-discovery configurations needed to test reliability. Trade-secret objections should be met with protective orders, in camera review, or expert-only access rather than allowing commercial confidentiality to override confrontation and due process rights.
The second battleground is admissibility. In federal court, counsel should invoke Rules 702 and 707 (once adopted) and seek hearings on machine-generated evidence. In New York and other Frye jurisdictions, counsel should require proof that the specific tool and methodology have gained general acceptance in the relevant scientific community.
Counsel should also look beyond the immediate forum. State restrictions on facial recognition and algorithmic impact assessments, along with the European Union AI Act’s treatment of certain law-enforcement AI as “high risk,” may support challenges in cross-border matters where evidence was obtained through tools that failed to meet emerging governance standards.
The government’s AI arsenal is expanding faster than the rules governing it. Practitioners who master the technology, demand transparency, and force courts to test algorithmic reliability will shape how AI evidence is treated for years to come. The algorithm may be the witness, but it is not beyond challenge.
"The Algorithm Is the Witness: How to Challenge AI-Generated Evidence in Criminal Investigations," by Bradley L. Henry, Rebecca L. Orel, and Layla S. Najjar was published in the New York Law Journal on July 13, 2026. Reprinted with permission.