What Changed in Federal Surveillance?
Federal intelligence agencies have integrated artificial intelligence into foreign intelligence collection over the past eighteen months without updating the statutory framework that governs compliance reporting under the Foreign Intelligence Surveillance Act. Two officials familiar with the matter said the expansion accelerated after the April 2024 reauthorization of Section 702, which added auditing requirements but omitted machine driven query tools. The result is a growing gap between surveillance capability and oversight capacity. That gap should worry anyone who values both security and the Fourth Amendment.
The Office of the Director of National Intelligence reported in April 2026 that U.S. intelligence agencies conducted approximately 204,000 Section 702 queries of U.S. person identifiers in calendar year 2025. That number is down from the peak years of 2021 and 2022, but the complexity of each query has increased. A senior official, speaking on condition of anonymity, confirmed that some agencies now use natural language models to prioritize and triage intercepted communications before human analysts review them. The official emphasized that human review remains part of the process. The concern is that the review standard itself was written for a manual workflow.
The Foreign Intelligence Surveillance Court, in an April 2026 opinion declassified in redacted form, noted that the FBI had conducted queries for purposes beyond foreign intelligence on multiple occasions and required the bureau to report corrective measures. The court did not rule on AI tools specifically. A Justice Department official with knowledge of the case said the opinion highlighted recurring compliance problems rather than a single catastrophic failure. But recurring problems become systemic problems when software multiplies the speed at which queries can be run. A flawed process at human speed becomes a mass violation at machine speed.
Why Does AI Complicate Oversight?
Artificial intelligence introduces three distinct oversight challenges that existing surveillance statutes do not address, including machine generated inferences, opaque model training data, and automated triage decisions that happen before human review. Those gaps mean that a compliance violation at human speed can become a mass violation at machine speed. Inspectors general are now asked to evaluate systems they cannot fully inspect. That is not oversight. It is after the fact documentation of choices made elsewhere.
A former Senate Intelligence Committee staffer said current reporting requirements focus on the number of queries and the number of compliance incidents, not on the design of the systems that generate them. That is the wrong metric for an environment in which a single poorly calibrated model could produce thousands of questionable hits in an hour. The staffer, who worked on the 2024 reauthorization negotiations, noted that lawmakers had discussed AI language but removed it from the final bill due to time constraints and disagreements over scope. The omission was not intentional sabotage. It was a legislative blind spot.
The Privacy and Civil Liberties Oversight Board issued a report in March 2026 warning that agencies had not produced uniform standards for testing AI driven analytic tools before deployment. The board found that some components within the intelligence community had conducted internal reviews, while others relied on vendor assurances. That inconsistency makes it impossible for congressional overseers to compare risk across agencies. It also makes it harder for the Foreign Intelligence Surveillance Court to evaluate whether querying procedures meet the reasonableness standard required by law.
Civil liberties groups have raised additional concerns about data retention. Under current procedures, incidentally collected U.S. person information can be retained under certain conditions for up to five years. If AI systems are used to re-scan archived intercepts against new identifiers or behavioral patterns, the practical effect is to extend the useful life of collected data far beyond what the original retention rules contemplated. A senior official, speaking on condition of anonymity, acknowledged that agencies are reviewing retention policies but said no formal changes had been approved as of late May 2026. The lag between technological reality and policy clarity is the problem.
What Should Congress Do Now?
Congress should require the intelligence community to submit an annual public report on AI use in foreign intelligence collection, including the categories of systems deployed, the data they process, and the compliance safeguards built into their design. Two officials familiar with the matter said similar reporting language had been drafted during the 2024 reauthorization but was stripped in conference. Reintroducing it now would not impede legitimate collection. It would simply restore the balance between capability and accountability that Congress intended when it created the current oversight structure.
Lawmakers should also direct the Government Accountability Office to conduct a cross agency audit of AI surveillance tools within 180 days. The audit should examine whether vendor contracts include sufficient access rights for inspectors general, whether agencies have documented the accuracy rates and error patterns of deployed models, and whether training data includes information that would be illegal to collect directly. A Justice Department official with knowledge of the case said the department would cooperate with such an audit if requested. That is a useful starting point for legislation.
The Foreign Intelligence Surveillance Court should receive additional technical support so it can evaluate AI systems without relying entirely on government representations. The court currently employs a small staff of legal professionals and security cleared clerks. It does not have independent technical advisers capable of assessing algorithmic bias, model drift, or prompt engineering risks. Providing those resources would strengthen judicial review without expanding the court's jurisdiction beyond its statutory mandate. Judges should understand the tools they are asked to approve.
None of these reforms require banning AI from intelligence work. That would be foolish and counterproductive. American adversaries are using the same technologies. The question is whether American surveillance will operate under rules written for the current century or under rules written when analysts flipped through paper transcripts. Congress has an opportunity to update the law before a major abuse occurs. It should take that opportunity. The country is safer when its spies are both capable and constrained by clear law.
