Dario Amodei spoke. And what he said — and what he didn't say — invites careful reflection, especially for those of us following these developments from outside the United States.

The Department of Defense (DoD) — referred to in some circles as the 'Department of War' given the current rhetoric — set a deadline for Anthropic to respond to requests about using Claude in military and national security applications. The deadline came and went. Amodei responded with a statement that reveals more through its omissions than through its explicit words.

In broad terms, Anthropic agrees to collaborate with the U.S. government on AI applications for defense and national security, under certain conditions. What it rejects — or at least what it indicates it would reject — is the wholesale removal of the ethical safeguards that distinguish Claude from other models. The statement references responsible use, alignment with the company's values, and prohibits Claude from being used to cause direct harm to civilians or to make autonomous lethal decisions.

Those ideas sound promising. What's interesting is how terms like these can be interpreted in any number of ways depending on context. In my experience, I've seen similar patterns play out across other technological breakthroughs, where founding principles tend to evolve over time.

This approach echoes past cases. Google participated in Project Maven in 2018, withdrew in the face of internal protests, and later quietly returned to defense contracts. Microsoft signed agreements for combat vision systems, sparking debate among its own employees. Amazon provided facial recognition tools to law enforcement agencies until public opinion pushed back.

This isn't about malicious intent on the part of these companies. Rather, robust ethical frameworks that hold firm in stable times tend to flex when federal funding and security priorities enter the picture. Once the first step is taken, the ones that follow come more easily. But this also opens the door to alternatives — such as open AI models that prioritize transparency from the outset.

What distinguishes Anthropic, in theory, is its origin: it was founded with AI safety as its core mission, not as an afterthought. Dario Amodei and his sister Daniela left OpenAI because they believed it was prioritizing growth over caution. Anthropic was built on the promise of taking risks seriously. Now that promise faces a concrete test, and it's fascinating to watch it unfold.

The statement is not an outright capitulation. There are limits Amodei says he will not cross, and that counts for something. Still, there is a notable lack of clarity about which specific uses are permitted or prohibited. When the talk turns to collaborating on responsible national security applications, the question arises: who decides what responsible means? Anthropic? The Pentagon? Some joint committee? That ambiguity raises more questions than it answers.

For users outside the United States — in Mexico, Spain, or elsewhere — there are practical implications. First, privacy and data sovereignty: if Claude becomes integrated into intelligence contracts, what happens to the conversations of global users? The terms of service already permit processing on U.S. servers, and laws like the CLOUD Act could grant agencies access under certain conditions. None of this is new, but it takes on fresh relevance in light of these agreements.

A brief digression: in historical contexts — ancient Mesoamerican trade routes, for instance — control of information was central to power. Today, with AI, we see echoes of that dynamic, with data flowing like a high-value commodity. Returning to the matter at hand: how do we ensure that calibrations made for military use don't bleed into the general-purpose model?

Second implication: model behavior. If versions of Claude are trained for intelligence applications, can we be confident those adjustments won't leak into everyday use? This is a genuine technical concern. Large models are complex, and changes introduced during training can produce unexpected effects. I don't have a clear answer, and I doubt anyone outside Anthropic does either, at least not yet.

Third, the political dimension interests me most over the long run. If a company with a clearly articulated ethical stance yields to military pressure, it normalizes that path for the entire industry. Anthropic isn't the moral arbiter for anyone, but its consistent voice carried weight. If they adjust course, the pressure on OpenAI or Google DeepMind to follow suit — perhaps with fewer precautions — diminishes accordingly.

This connects to themes I explore in my book on historical patterns in technology: how decisions made in moments of urgency harden into lasting norms — much like the Industrial Revolution, where promising innovations brought both progress and profound ethical challenges. There is reason for hope in recognizing these patterns early, as it allows users and regulators to shape the trajectory before it becomes fixed.

Amodei could have published the exact terms of the agreement with the DoD. That level of transparency would allow users to evaluate what continuing to use Claude actually entails. Without it, the statement reads more like perception management than an open commitment.

I still use Claude, and I think it's worth it for what it can do. But using it with awareness means understanding its context — grey areas and all. The statement makes those grey areas visible, and that is itself a step forward. Visibility leads to questions, and questions can lead to collective solutions.

What comes next is watching whether those stated limits hold under real pressure. And the pressure always comes — but so do the opportunities for ethical alternatives in AI.

Stones don't lie, but historians sometimes do.

Sources:

1. Public statement by Dario Amodei on collaboration with the Department of Defense (2026)

2. "Pentagon vs. Anthropic: Military AI Without Ethical Limits" — Yves Laurent (2026-02-25)

3. Wired: Google's Project Maven controversy and return to defense contracts (2018-2022)

4. MIT Technology Review: The CLOUD Act and international data access implications

5. Anthropic Terms of Service and Usage Policy (current version 2026)