Yuval Shany is the Hersch Lauterpacht Chair in International Law and former Dean of the Law Faculty of the Hebrew University of Jerusalem.

In recent weeks, Anthropic — one of the world’s leading AI companies — has made headlines in two security-related contexts, both of which test its commitment to appropriate legal and ethical standards.

The first is its confrontation with the US Department of Defense (DoD) (referred to as the “Department of War” by the Trump Administration), over limitations it placed on the use of Claude, its suite of general-purpose large language models. The second is the US military’s reported use of Claude targeting decisions in the ongoing war with Iran.

Notably, shortly before these developments, on January 21, 2026, Anthropic published its new “Constitution” for Claude, which articulates relevant legal and ethical standards. The 84-page Constitution is presented as a “foundational document that both expresses and shapes who Claude is.” It enumerates actions that Claude should refrain from undertaking (“hard constraints”), and identifies considerations the system should weigh when deciding whether to perform certain actions.

This convergence of normative developments and real-world deployments highlights the importance of subjecting AI systems to robust human rights safeguards and enables a critical evaluation of the extent to which the new Claude Constitution achieves that goal.

Where are human rights?

According to the Constitution, Claude should be guided by four sets of values, applied in the following hierarchical order: safety, ethics, compliance with Anthropic guidelines, and helpfulness. In other words, Claude should strive to assist users unless instructed otherwise by Anthropic or if it deems a request unethical or unsafe.

The Constitution also introduces a number of “hard constraints” — specific prohibitions that must never be violated, including attempts “to kill or disempower the vast majority of humanity or the human species as a whole” or assisting “any individual or group with an attempt to seize unprecedented and illegitimate degrees of absolute societal, military, or economic control.”

While some ethical standards set out in the Constitution overlap with human rights, e.g., privacy, protection from harm, the rule of law, equal treatment, access to information and political freedom, the document does not explicitly reference the concept of “human rights.”

This marks a shift from earlier iterations; the 2023 version of Claude’s Constitution referred to the UN’s Universal Declaration of Human Rights.

As a result, many important human rights protections that could be relevant to the operation of Claude — including personal liberty, freedom of religion and intellectual property rights — are not clearly integrated into its normative framework.

Given the global reach of systems like Claude and the limited capacity or willingness of states to regulate their use, this omission is significant. This is especially so given the growing interest in developing formal and informal international human rights standards for AI systems, such as the Council of Europe’s 2024 Framework Convention on AI and Human Rights, Democracy and the Rule of Law and the 2025 Oxford White Paper on an International AI Bill of Human Rights.

Anthropic vs DoD

Soon after the publication of the Constitution came the real-world tests noted above, both underscoring the importance of effective normative safeguards.

First, the DoD designated Anthropic a supply-chain risk due to its refusal to allow Claude to be used for mass domestic surveillance purposes or for operating lethal autonomous weapon systems. Instead, the DOD entered into a contract with OpenAI for the provision of alternative AI systems.

As Dr. Brianna Rosen of Oxford University’s Blavatnik School of Government explained, the DoD’s insistence on being able to use AI systems for “any lawful use” leaves in place a governance gap, since neither US law nor international law clearly prohibits, in all circumstances, mass surveillance or the use of autonomous weapon systems.

Defining the permissible scope of such extraordinary capabilities through contractual negotiations between the government and a private AI company appears to provide weaker human rights protections than embedding universally accepted safeguards directly into the AI system itself, through clear language in its constitution or a comparable normative framework governing its actual operations. This is particularly the case given the difficulties of monitoring and enforcing state compliance in sensitive domains such as national security.

Dr. Rosen is also right to point out, in this regard, that Anthropic’s negotiating position on mass surveillance, which focuses only on restricting domestic surveillance, may already fall short of international human rights standards in the field, which extend to foreign surveillance too.

Second, it has been widely reported that Claude systems, still in use by the US military, have been employed in the war in Iran for target selection. It has also been speculated, albeit without hard evidence, that the use of AI systems by the Pentagon may have contributed to a high-profile operational mistake (the targeting of an Iranian school located next to a military base) due to reliance on out-of-date maps of the area.

Here again, questions arise as to whether Claude’s Constitution, as currently drafted, contains sufficient safeguards against the use of AI in contexts involving lethal consequences.

Arguably, a robust human rights-oriented approach would include an explicit requirement that any use of the AI system in armed conflict comply with the core principles of international humanitarian law (which operationalize human rights principles in wartime), including flagging precautionary obligations such as real-time target verification before attacks are recommended.

In this sensitive policy space, reliance on AI systems may not only result in operational errors but also risk perpetuating accountability gaps, allowing human actors to deflect responsibility onto the technology.

In such cases, embedding human rights “by design” within a system’s constitutional framework could provide a more effective safeguard against violations of basic individual rights. If the events of the last few weeks have shown us anything, it’s that Anthropic and other leading AI companies that do business with governments and militaries would be wise to do so.