Claude's Military Use: Do Principles Hold in Contracts?

The English draft matches the Korean source accurately in meaning, structure, numbers, links, and headings. No corrections needed.

What caught my eye in the news about the Iran strikes was Claude’s military use. The fact that an AI I’ve used myself is also being deployed for military intelligence analysis and operational support made me reconsider who actually decides the boundaries of its use.

On March 3, 2026, CBS reported, citing sources, that the U.S. military used Claude in its strikes on Iran. The Department of Defense hasn’t disclosed the specifics of how it was used. CBS report

This report shouldn’t be read as meaning the AI picked targets and struck on its own. We need to distinguish between using AI in military operations and letting it decide on an attack without human involvement. That distinction matters in the friction between Anthropic and the U.S. government, too.

From Video Analysis to Broader Military Operations: AI’s Expanding Role

Project Maven, launched by the U.S. Department of Defense in 2017, focused on using computer vision to identify objects of interest across massive volumes of footage. It was designed to ease the burden of having humans review every frame by hand.

Google took part in the project too, but decided not to pursue follow-on contracts after employee opposition and controversy in 2018. At the time, Google’s official explanation was that it would fulfill its existing contractual obligations without signing new ones. Alongside a principle of not supporting AI for weaponized systems, it said it would continue other military-related work, such as cybersecurity and productivity tools. Google’s 2018 statement

Military use of AI continued after that. In May 2024, the U.S. Department of Defense awarded Palantir a contract for a Maven Smart System prototype. That November, Anthropic and Palantir announced a partnership to deliver Claude to U.S. intelligence and defense agencies through AWS. Claude was integrated into Palantir’s AIP so it could be used for complex data processing and decision support. May 2024 contract notice, November joint announcement

In July 2025, Anthropic, Google, OpenAI, and xAI each received contracts with ceilings of up to $200 million. A contract ceiling isn’t the same as revenue actually paid out. What this announcement confirms is that multiple AI companies were already involved in development and deployment within the defense sector. U.S. DoD CDAO announcement

The Two Uses Anthropic Wanted to Exclude

Anthropic isn’t a company that opposes military use across the board. In its statement of February 26, 2026, it listed intelligence analysis, modeling and simulation, operational planning, and cyber operations support as areas where it’s willing to cooperate.

What the company sought to restrict was large-scale domestic surveillance within the US and fully autonomous weapons. The statement explains that neither use was included in the existing Department of Defense contract. Based on what’s been made public, it’s hard to conclude that new conditions were suddenly attached after the system was already deployed with the military. Anthropic’s February 26 statement

The reasoning behind these two restrictions also deserves separate scrutiny. For domestic surveillance in the US, the company cited democratic values and privacy concerns. For fully autonomous weapons, it explained that current AI reliability and oversight mechanisms aren’t yet sufficient. The company also stated that it supports partially autonomous weapons and legitimate foreign intelligence activities.

So this position can’t be read as opposition to all wars or all weapons use. Nor does it mean the company unconditionally permits all surveillance conducted abroad. You have to read closely what scope of use the company has publicly permitted, and why it drew the limits where it did.

The Department of Defense took the position that the technology should be usable for all lawful purposes. The dispute wasn’t only about who makes the individual decisions in military operations — it was about whether a supplier can use a contract to restrict the scope of use, on top of what law and military policy already govern.

An announcement to stop use and actual withdrawal are different processes

On February 27th, President Trump ordered federal agencies to stop using Anthropic, and Defense Secretary Pete Hegseth announced plans to designate the company as a supply chain risk. A transition period was proposed for existing military use.

In a statement on March 5th, Anthropic said it had received notice of the designation the day before, on March 4th. While disputing the legality of the designation, the company said it would continue providing its models and technical support for as long as permitted, so that the military’s operations wouldn’t be disrupted during the transition. Reports that the model was still being used in operations even after the stop-use order was announced need to be understood in this context of a transition process. Anthropic’s March 5th statement

Replacing a model that’s already been integrated requires re-verifying the new model’s outputs, its connections to existing software, and users’ work procedures. That said, based on publicly available information alone, it’s impossible to say exactly how many months this would take, or which functions would be hardest to replace.

Other companies’ commitments deserve the same close reading

OpenAI announced on February 27th that it had reached an agreement to deploy AI in classified environments. The company said the deal restricts large-scale domestic surveillance, autonomous weapons command, and the automation of high-risk decisions that require human sign-off.

The contract language OpenAI disclosed includes a clause permitting lawful uses, alongside conditions for human control under law, regulation, and Department of War policy. The company also pointed to cloud deployment, its own safeguards, and the involvement of security-cleared personnel as protective measures. On March 2nd, it said it had added language prohibiting domestic surveillance of Americans — including through the use of commercially acquired personal data. OpenAI’s explanation of the agreement and the March 2nd addition

This is the contract content and explanation the company has chosen to disclose. How those safeguards actually function and are overseen in real operational settings is a separate matter to verify. It would be excessive either to claim the deal grants unlimited use, or, conversely, to declare that safety has been fully guaranteed.

When Google revised its AI principles in February 2025, it cited a more complex geopolitical environment and cooperation in national security. The emphasis shifted from the explicit exclusions on weapons and certain surveillance uses that it had laid out in 2018, toward principles that weigh benefits against risks. The wording from the two points in time is worth comparing, but it can’t be read as a single, unified shift across all of Silicon Valley. 2018 principles, 2025 revision explanation

xAI is also among the companies participating in the 2025 CDAO contract. But that fact alone doesn’t let us conclude that the specific terms of its classified agreement are looser than other companies’, or that it accepts every government request unconditionally.

Oswarld’s Lens

I found it hard to simply take one side in this dispute. Having actually used Claude, I’ll admit I lean a bit more generously toward the company—but I want to keep that product experience separate from my judgment on the military contract itself.

From a business standpoint, I think it matters to clearly agree on permitted uses and the procedure for changing terms before systems become deeply integrated. How a conflict is handled publicly can also affect the negotiation itself. That said, since we can’t know the full internal history of this particular negotiation, we can’t simply assume things would have turned out differently had it been settled quietly. In particular, treating a usage restriction that had reportedly existed all along as a belated change in terms gets the judgment wrong.

The question that stays with me is how human control actually works in practice. If AI gathers information and proposes a prioritized list of target candidates, the final approver should be able to examine the underlying evidence and its uncertainties. The mere fact that a human presses the approval button doesn’t tell us whether adequate review actually took place.

I think the same standard needs to apply when discussing military manpower and unmanned systems here in Korea. Alongside deciding how far AI should be used, we need to determine who reviews the outcomes, how things get halted when errors are found, and who bears responsibility for the judgment.

Acknowledging a military need and dismissing every concern a supplier raises are two different things. A company’s stated principles, too, should be verified through contracts, technical controls, and actual usage records—not just PR language. Watching this dispute unfold made me think again about how difficult that verification is, and precisely because of that, how necessary it is. More than anything, I hope the harm from this war doesn’t grow any further and that it ends soon.

If you’d like to read more on the legal issues, see Lawfare’s analysis of the supply-chain risk designation. It’s a legal analysis critical of the designation, and should be read as distinct from any actual court ruling.

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.