As artificial intelligence asserts itself in sensitive domains like defense, a central question emerges: who truly controls the machine?
Beyond the technical debates, the case pitting Anthropic against the U.S. government reveals an unprecedented tension between state sovereignty and algorithmic control. Thus, the alignment of AI systems surfaces as a major new legal challenge.
AI is taking an ever-growing place in sovereign sectors such as defense and public security, and this rapid rise is accompanied by delegating certain tasks traditionally performed by humans to complex and sometimes opaque algorithmic systems.
Yet AI technologies carry a high risk of producing biased or erroneous results, which has led AI system designers to develop so‑called alignment mechanisms to regulate the behavior of these systems.
Hence, it seems legitimate to ask to what extent AI alignment, conceived as a tool for ethical and technical governance of autonomous systems, constitutes a right enforceable against the State, or a constraint on sovereign prerogatives, when human oversight collides with national defense and public safety imperatives.
AI alignment as a mechanism for the technical governance of autonomous systems
Alignment refers to the set of techniques aimed at embedding human values (legal, ethical, or security-related) directly into the operation and architecture of AI systems. In practice, it involves tuning these systems so they deliver reliable responses while avoiding uses deemed dangerous. The objective is to prevent the machine from straying from fundamental principles, even if a user attempts to manipulate it.
Alignment should be viewed as a default compliance guarantee. A flagship example is Anthropic’s “Constitutional AI framework,” which imposes intangible limits on the capabilities of this AI tool. The aim is to prevent the risk of divergence between the objective set by the user and the fundamental principles of safety and
In the Anthropic vs. the U.S. government case, human control is no longer merely ex post oversight but an ex ante restriction embedded in the algorithm of the AI system “Claude,” allowing the protection of certain fundamental principles. By refusing to lift its “guardrails” for autonomous weapons and mass surveillance, Anthropic exercises a technical withdrawal right grounded in alignment, and argues that the loss of this architectural control over AI would render the AI system unpredictable.
In this particularly sensitive case across the Atlantic, a federal judge in California issued a temporary injunction on March 26, 2026, effective April 2, 2026, suspending the Pentagon’s decision to classify Anthropic as a “supply chain risk” to national security and to bar federal agencies from using its technology, finding that these measures are likely illegal and arbitrary as well as infringing on freedom of expression. In this battle, it is up to the U.S. government to decide whether to appeal or not the decision, which has not yet been disclosed publicly.
Human control over AI thus lies at the heart of a major legal and political confrontation.
The AI market players who have agreed to collaborate with the American authorities, including in the defense sector, principally argue that cooperation with the state falls within a framework of democratic legitimacy where public institutions responsible for national security must be able to access the most advanced technologies.
However, these actors do not claim an absence of limits but argue that guardrails can be maintained, notably through the system’s technical architecture with integrated constraints. This position is sometimes accompanied by a relaxation of ethical principles, particularly in military or surveillance applications, which does not signify abandoning human control but rather reconfiguring it—shifting from strict ex ante control by the vendor to a shared, negotiated control, partly transferred to the State.
The technological conscience objection tested against raison d’État and defense imperatives
This disagreement goes beyond a purely technical frame and highlights a clash between two logics:
On one side, the State invokes its sovereign prerogatives, notably national defense and public safety. In this context, access to advanced technologies can be regarded as strategic.
On the other side, companies defend their freedom to undertake and their responsibility in designing safe tools; refusing certain uses thus becomes a form of technological conscience objection.
Human control thus becomes a battleground: should it be exercised by the State in the name of the public interest, or by designers in the name of system safety?
The legal necessity of effective human oversight over AI systems as a condition of legitimacy and safety
Beyond this conflict, a fundamental question remains about the use of AI systems without real human oversight. A major legal issue arises from the absence of alignment or supervision, since AI by its nature is irresponsible. Thus, in the event of damage, it would be difficult to identify a responsible party if no clear rule had been embedded upstream.
This refers back to the core rules laid out by the AI Act. Human control is a principle of architecture for AI systems, not merely a post hoc capability, and it directly contributes to alignment.
The “Human-in-the-loop” principle could be partially challenged by the Digital Omnibus Package currently being discussed within European Union institutions. Indeed, while this project does not remove human control, it reveals a tension between regulatory simplification and maintaining the structural conditions for alignment, the outcome of which will depend on the final compromise among the institutions.
Maintaining human principles at the heart of AI system algorithms is not a brake on innovation but, on the contrary, a guarantee of legal safety that is necessary to prevent the automation of potentially dangerous actions from proceeding without real human supervision.
* Corinne Thiérache is a partner at the law firm ALERION and Aya LYAZIDI, a Master 2 student in Digital Law at Université Paris Panthéon-Assas
