You watch a commander hover over a pixelated map while an AI model lists strike options. I remember sitting in a hearing where lawyers and engineers argued whether a machine could be trusted to decide life or death. You feel a small, urgent chill when technology meets the battlefield.
In the Senate hearing room: a contract collapses and questions spread
Senator staffers muttered as the Department of Defense deal with Anthropic unraveled this spring.
I’ve tracked these moments long enough to know when a policy pivot becomes a political movement. Senator Adam Schiff introduced the Human Authority in Lethal Operations (HALO) Act to force a human commander to make the final call on any lethal action suggested by autonomous systems. The bill also requires detailed logs of how targets were chosen, whistleblower protections, and explicit limits on AI use around nuclear decisions and mass domestic surveillance.
The Anthropic break with the Pentagon — and the company’s designation as a supply-chain risk — turned a dry procurement row into front-page political drama. Anthropic allegedly refused to remove safety guardrails that would have let the DoD use its models without meaningful human control. After that, the Pentagon quietly signed contracts with OpenAI, Google, Nvidia, SpaceX, Microsoft, and Amazon Web Services.
What is the HALO Act?
HALO is a statute-length safety belt: it demands human final authority for lethal actions, mandates retrospective records for decision review, protects whistleblowers, and draws red lines around nuclear targeting and widespread domestic surveillance. Schiff is reportedly pushing to fold the bill into the annual defense package, the NDAA, which practically guarantees high-stakes floor fights and intense lobbying from industry.
On the Senate floor and in staff offices: similar bills gather steam
In a cluttered Senate office, aides passed around a printout of Gillibrand’s proposal.
Senator Kirsten Gillibrand’s bill mirrors HALO on nuclear limits, domestic spying guards, and a requirement that senior DoD officials greenlight “high-consequence actions.” Senator Elissa Slotkin has an AI Guardrails Act with nearly identical aims and is preparing her own NDAA amendment. The pattern is clear: three Democrats are building overlapping safety nets to outlaw fully autonomous kill-chains and to demand human judgment at decisive moments.
At the intersection of law and code: what these bills actually do
A lawyer flipped through FOIA files showing how algorithms shaped past targeting decisions.
You should know what’s on the table: these proposals don’t ban every use of AI in the military. They ban fully autonomous weapons where a machine can select and engage targets without human signoff, restrict certain surveillance applications, and require traceable decision logs. For operators, that means more paperwork and clearer responsibility; for watchdogs, it offers a paper trail to hold officials and vendors accountable.
Can Congress ban fully autonomous weapons?
Yes, in practical terms: Congress can pass statutory limits that make fully autonomous weapons illegal for the Defense Department to use or fund. But a law is only as strong as its enforcement mechanisms and political backing. With the NDAA as the vehicle, these measures have a better chance, yet they’ll bump against defense contractors and tech firms eager to protect contracts with the Pentagon.
In the control room: human oversight can still fail
An experienced officer admitted to me that teams often defer to algorithmic recommendations during high-pressure strikes.
That deference is called automation bias — you tend to trust a system that seems to process more data or return quick answers. But AI models hallucinate, inherit bias, and can behave opaquely. The technology’s reasoning is often hidden from the humans tasked with oversight. AI in a warzone is like a stoplight on a highway: one malfunction, and cars collide. Requiring a human signature is necessary, but not sufficient, to stop cascading errors when that human assumes the machine is right.
On the campaign trail: politics, contracts, and reputation
At a campaign stop, a candidate praised Anthropic; in another room, a Pentagon official defended its procurement moves.
The fallout from the Anthropic deal has split attention. Democrats rallied to support Anthropic’s refusal to strip safety layers, while the Pentagon broadened its vendor list to include most major AI players. Anthropic has sued its designation and released Mythos, a model that spooked parts of the cybersecurity world and reportedly warmed the Trump administration’s interest. Meanwhile, vendors such as OpenAI, Google, Microsoft, Nvidia, SpaceX, and AWS are navigating new guardrails and commercial opportunities with the DoD.
At the ethics panel: accountability, logs, and whistleblowers
A former ethics advisor showed me a redacted log where nobody could explain a strike’s data lineage.
Schiff’s requirement for detailed records and whistleblower protections seeks to create forensic trails. That matters: when a target decision can be reconstructed, courts and Congress can assign responsibility. But that traceability only works if records are complete, secured against tampering, and accessible to independent auditors — a tricky mix when classified systems and proprietary models are involved. The Pentagon’s claim that vendor code is proprietary often clashes with demands for transparency.
Where the law meets battlefield risk: the human factor
A captain told me he would rather refuse a strike than trust an algorithm he did not understand.
Training, doctrine, and culture shape whether a human will push back against a machine. Without rigorous training and clear legal accountability, human oversight can become a paper exercise. The Pentagon must pair any legal guardrails with technical audits, red-team tests, and operational doctrine that rewards skepticism, not blind compliance. The Pentagon’s trust in tech feels like a cracked compass, pointing leaders toward peril unless they recalibrate authority and accountability.
What to watch next: timing, the NDAA, and industry posture
Inside the defense committees, staffers are plotting which amendments will ride the NDAA.
If Schiff, Gillibrand, and Slotkin can attach their language to the NDAA, enforcement will shift from theoretical to immediate. Expect intense lobbying from contractors and major cloud and AI providers. Watch for classified carve-outs, exemptions for certain experimental projects, and fights over audit access. You should also watch the courts: Anthropic’s legal challenge to its supply-chain designation may create new precedents about how far the government can push vendors to remove safety controls.
I’ve told you the players, the bills, and the risks — now decide: will Congress force real limits on AI in war, or will policy trail behind procurement deals and vendor influence?