Pentagon Centralizes AI – Who Checks It?

Soldier in tactical gear holding rifle with futuristic HUD overlay — Photo: K_E_N / Shutterstock

A sweeping Pentagon order just centralized wartime artificial intelligence under one office, raising hard questions about who is watching the watchers.

Story Highlights

A Pentagon memo made Palantir’s Maven an official program and shifted oversight to a single AI office.
Military leaders admit integrating AI into real targeting is “tricky,” even as deployments expand.
A new AI pilot could become the top authority for classifying secrets within months.
Law now requires annual reports to Congress on lethal autonomy through 2029, confirming ongoing oversight duties.

Pentagon consolidates AI power under one office

Reuters reported that Deputy Secretary of Defense Steve Feinberg ordered Palantir’s Maven to become a formal program of record. The same directive moved oversight from the National Geospatial-Intelligence Agency to the Pentagon’s Chief Digital and Artificial Intelligence Office within 30 days. This shift concentrates authority for targeting software and data pipelines. Supporters say speed wins wars. Skeptics see a risk: fewer independent checks when algorithms touch life-and-death calls.

Officials also described a fast-expanding footprint for Maven in real operations. The Pentagon’s chief digital and AI officer said the first full rollout, during Operation Epic Fury, pushed classified computing to its limit. That kind of surge suggests rapid scale. It also signals a testing challenge. Bigger deployments mean more chances for error. Good governance needs strong logs, audits, and human review. Central control without visible counterweights can let problems grow unseen.

Commanders warn adoption is hard while targets cannot wait

Army Brigadier General John Coghill, a senior leader at the 18th Airborne Corps, said adopting AI into targeting is “tricky”. That simple word tells the truth. Real combat is messy. Data is noisy. A model trained on one fight can stumble in another. When tools flag threats faster than humans can verify them, leaders can feel pressure to trust the screen. That is why chain-of-command accountability and clear stop rules still matter.

Supporters inside the department say humans remain in charge of critical decisions. Standing policy and doctrine assign commanders responsibility for priorities, timing, and law-of-war compliance. Guidance also warns against over-reliance on AI without human oversight. Those are sound principles. But principles on paper are not performance in the field. The question for Congress and the public is whether real units have time, training, and data access to check the machine before acting.

AI for classifying secrets moves on a rush timeline

A report on an internal October 5, 2026 memo says Feinberg ordered an “ACME” classification capability within 180 days. The plan envisions making it the single, digital source for original classification decisions. That would hand a core gatekeeping job to software on an accelerated schedule. The memo itself is not publicly available, and testing standards are not in view. Speed can save money. But classification mistakes can hide failures or expose sources. This needs sunlight before scale.

Congress already acted to keep a spotlight on lethal autonomy. The Congressional Research Service says the law now requires the Secretary of Defense to send annual reports on approval and deployment of lethal autonomous weapon systems through December 31, 2029. That duty assumes continuing oversight, not one-time approval. It is a floor, not a ceiling. Lawmakers should demand readable reports, unit-level audits, and red-team findings so oversight is real, not a box-check.

Guardrails clash: vendors, law, and mission need alignment

Reuters reported a clash between the Pentagon and Anthropic over usage limits. The company warned its tools could aid targeting or domestic spying without enough human oversight. Pentagon officials argued that lawful use takes priority over vendor policies and moved to sideline the firm as a supply-chain risk during the dispute. That fight shows the stakes. If providers who raise safety flags get frozen out, fewer dissenting views will reach commanders when it counts.

At the same time, the Pentagon says responsible AI will be tested, traceable, and governable across its life cycle, with clear uses and the ability to deactivate systems that behave badly. That is the right promise. Now it needs proof. Conservatives should insist on three things: human-in-the-loop authority that is meaningful; audit trails that an inspector general can follow; and competition that favors safe, effective tools, not cozy contracts. Warfighters deserve gear that works and rules that protect them.

What to watch: proof, not press releases

Here is the bottom line for readers. First, ask for the ACME memo, model cards, and error rates before it becomes the nation’s arbiter of secrets. Second, require after-action reports that show when Maven changed a target decision and how a commander validated it. Third, use the annual lethal-autonomy reports to check that approvals match deployments in the field. Strong oversight is not anti-military. It is pro-mission, pro-troop, and pro-Constitution.