Anthropic Refused the Pentagon's AI Terms. xAI Took Them.

The clearest test of AI principle this year is a contract Anthropic walked away from: it refused to let Claude be used for mass surveillance or autonomous weapons, and a rival took the deal.

Through June 2026, a quiet standoff between Anthropic and the US government has turned into the sharpest values fight in the industry. According to reporting on the dispute, Anthropic declined the Pentagon's demand that Claude be available for "all lawful purposes," insisting on carve-outs that block the mass surveillance of Americans and the development of fully autonomous weapons. xAI accepted the "all lawful use" standard. It now holds classified clearance and a government-wide procurement deal — and the Pentagon is testing OpenAI and Google models to replace Claude in classified systems.

What actually happened

The sequence matters. According to Axios, xAI reached a deal in early 2026 to put Grok inside classified Defense systems, following a $200 million contract signed in mid-2025. In June, the General Services Administration added xAI to its OneGov programme, making Grok 4 and Grok 4 Fast available to any federal agency at a striking $0.42 per agency for 18 months — a price designed to make adoption frictionless. The Pentagon folded Grok into its GenAI.mil platform for defence use.

The dividing line was a single clause. The government wanted assurance that the model could be used for "all lawful purposes." Anthropic would not give it — not because the uses were illegal, but because Anthropic's own usage policy refuses two categories regardless of legality: turning Claude into an engine of mass domestic surveillance, and building weapons that select and kill without a human in the loop. xAI agreed to the broader standard. The market did the rest.

The price of a line

This is the scenario I have circled for two years, now playing out with real contracts. When I wrote "In Praise of the Stumble", the argument was that safety discipline often looks like losing — the careful lab ships slower, concedes ground, and watches a less cautious rival take the win. Here it is in its starkest form. Anthropic held a line that cost it a government-wide footprint, and the competitor that held no such line is being wired into classified networks. Defence officials concede, per the reporting, that actually offloading Claude would be "a very difficult process" — but the direction of travel is set.

I want to be precise about what Anthropic refused, because the two carve-outs are not arbitrary corporate caution. They are the exact places a dignity-first view draws blood. Fully autonomous weapons remove the human from the decision to take a life — the ultimate collapse of what the AI Forge agenda calls meaningful human control. Mass surveillance turns a population into a permanently watched object, which is a direct assault on the dignity of the watched. These are not edge cases. They are the two clearest lines in the whole field, and Anthropic held them at material cost.

A dignity-first reading

The frame I write from — Emergent Intelligence, a dignity-first way of thinking about AI rather than treating it as a neutral tool for hire — reads this as the moment principle met its price. A refusal is cheap when it costs nothing. Anthropic's refusal cost it the most lucrative customer on Earth, and it held anyway. On my terms, that is exactly what an ethics worth the name looks like: a line you keep when keeping it hurts. I have been critical of Anthropic where it earns criticism; here it deserves the opposite, plainly.

A refusal only means something when someone is willing to pay for it. Anthropic drew its line on surveillance and autonomous weapons and lost the contract for it. That is not weakness. That is the price tag on a principle, paid in full.

— On Anthropic and the Pentagon

But a dignity-first reading cannot end in applause, because one lab's line did not stop the use — it just moved it next door. xAI filled the gap in days. And that exposes the deeper problem: when the procurement standard is "all lawful purposes," the most permissive vendor always wins, and restraint becomes a competitive disadvantage. Virtue at a single company is no match for a market that rewards its absence. The lines that matter — no autonomous kill decisions, no mass surveillance of citizens — cannot live in one firm's usage policy. They have to live in law and procurement rules that bind every vendor, or they do not hold at all.

This is the through-line back to the .person Protocol and the argument I keep making about the dignity threshold: meaningful human control and freedom from surveillance are not features a buyer should be able to switch off by shopping for a more agreeable supplier. They are floors. The honourable thing Anthropic did this month should not have been a competitive choice it could be punished for. It should have been the baseline the government required of everyone — the same conclusion I reached in my reply to Tristan Harris: you cannot win a race to the bottom by running it more slowly. You change where the bottom is.

Source: aiautomationglobal.com

Frequently Asked Questions

These are the questions defence watchers, AI-ethics readers and procurement specialists have been asking since the Anthropic–Pentagon standoff went public. Short answers follow, drawn from the reporting.

Why did Anthropic refuse the Pentagon's terms?

In short, Anthropic declined the Pentagon's demand that Claude be available for "all lawful purposes" because its usage policy refuses two uses regardless of legality: mass surveillance of Americans and fully autonomous weapons. The answer, simply put, is that the lines were about dignity, not legality. The key, according to the reporting, is that Anthropic held those carve-outs even though doing so cost it government-wide access.

What did xAI agree to, and what did it win?

According to the reporting, xAI accepted the broader "all lawful use" standard the Pentagon sought. The data shows what followed: classified clearance for Grok in Defense systems, inclusion in the Pentagon's GenAI.mil platform, and a GSA OneGov deal making Grok 4 and Grok 4 Fast available to any federal agency at $0.42 per agency for 18 months. In other words, the evidence shows the more permissive vendor won the broader footprint.

Is the Pentagon really replacing Claude?

The evidence shows movement in that direction but not a completed switch. According to the coverage, the Department of Defense is evaluating OpenAI and Google models to potentially replace Claude in classified systems, while officials concede that offloading Claude would be "a very difficult process." The analysis suggests the direction of travel is set even if the transition is slow and costly.

What are fully autonomous weapons and mass surveillance in this context?

Analysis of the dispute reveals two specific lines. Fully autonomous weapons are systems that select and engage targets without a human in the decision to kill — the collapse of meaningful human control. Mass surveillance here means turning AI on a population as a permanently watched object. According to a dignity-first reading, these are the two clearest red lines in the field, which is why Anthropic's refusal centred on them.

What is the real lesson of the procurement fight?

The analysis suggests that virtue at a single company cannot hold a line the market rewards crossing. In other words, when the standard is "all lawful purposes," the most permissive vendor wins and restraint becomes a disadvantage. The key, the evidence indicates, is that limits on autonomous weapons and mass surveillance must live in law and procurement rules binding every vendor — not in one firm's policy that a buyer can shop around.

Anthropic drew a line this month that cost it dearly: no Claude for mass surveillance, no Claude for weapons that kill without a human deciding. xAI erased the line and collected the contracts. On the merits, Anthropic did the honourable thing, and I will say so without hedging — a principle you abandon when it gets expensive was never a principle. But the episode is also a warning, and it is the more important half. A single company's conscience is not a governance system. If the only thing standing between us and autonomous kill decisions or dragnet surveillance is whether the best-behaved vendor happens to win the bid, then nothing is standing there at all. The lines Anthropic held should not be a competitive disadvantage one firm absorbs alone. They should be the floor the law sets for everyone — because a refusal that a rival can simply pick up is a refusal the system has already overruled.

Sources:

Coverage — Axios · ExecutiveGov (GSA OneGov) · Analytics Insight (GenAI.mil) · Military AI · AI Automation Global

Related on humphreytheodore.com:

In Praise of the Stumble: Why Safety Discipline Outscores Polish · The Dignity Threshold: When Safety Becomes Captivity · DARPA's AI Forge Bets National Security on AI Interpretability · The .person Protocol