Anthropic CEO Dario Amodei: Models Cannot Be Shut Down After Delivery to Military
Anthropic stated in a court filing that once its AI models are deployed to the U.S. Department of Defense, it will not have the capability to remotely control or shut down these models. This statement directly addresses the issue of control boundaries for AI systems in military contexts.
The document reveals the background of legal disputes surrounding AI safety and accountability. Axios cited relevant materials indicating that Anthropic's position means that once the models are delivered, they are under the complete control of the user, which creates a clear tension with the current public narrative of "controllable AI" by tech companies.
Similar issues have previously emerged in collaborations between various AI companies and the government, including the division of permissions in cloud deployments between OpenAI and Microsoft, and the control architecture design in defense data systems by companies like Palantir, all involving the core question of "who has the final switch."
Source: Public Information
ABAB AI Insight
This information's key point lies not in "technical limitations" but in "transfer of control." Anthropic's statement essentially confirms at a legal level that once AI enters the national machinery, its governance no longer belongs to the technology provider but to sovereign institutions. This marks a shift from "technical products" to "infrastructure capabilities," similar to how early cloud computing transitioned from corporate IT outsourcing to national critical systems.
This structural change indicates that AI is entering the power framework of the traditional military-industrial complex. Historically, whether it be nuclear technology, satellite systems, or internet infrastructure, once integrated into the national security system, the control logic shifts from "commercially controllable" to "strategically irreversible." Anthropic's statement is, in fact, a preemptive delineation of responsibility boundaries to avoid future legal and ethical liabilities in uncontrollable usage scenarios.
A deeper issue lies in the institutionalization of "alignment failure." The current AI industry emphasizes alignment, but this premise assumes that models remain under company control. Once control is transferred, the alignment mechanisms will no longer be led by model developers but defined by the users. This means that the behavioral boundaries of AI will shift from "technical specifications" to "will of power," which is particularly evident in military systems.
From a financial and industrial structure perspective, such statements are also reshaping the risk pricing models of AI companies. Uncontrollable deployments mean that potential legal tail risks are transferred, while also implying that government clients will gain a higher level of "sovereign-grade AI capabilities." This will accelerate the concentration of the AI supply chain towards the state, reinforcing the "quasi-military contractor" attributes of leading model companies rather than traditional software companies.