The Department of Defense has awarded contracts of up to $200 million each to Anthropic, Google, OpenAI, and xAI to develop agentic AI workflows and address national security needs.
The Chief Digital and Artificial Intelligence Office within the Defense Department specified these contracts are intended to develop “agentic AI workflows across a variety of mission areas” and to “increase the ability of these companies to understand and address critical national security needs.” The selection of these particular companies raises questions regarding the ideological constitutions and alignment methodologies of some models for governmental use.
OpenAI employs a method known as reinforcement learning from human feedback to align its AI models, including ChatGPT. This approach integrates a reward model with human input, aiming to minimize “untruthful, toxic, [and] harmful sentiments.” IBM has explained that a key advantage of RLHF is its independence from a “nonexistent ‘straightforward mathematical or logical formula [to] define subjective human values.’” Google also utilizes this precise method for aligning its large language model, Gemini.
Anthropic’s AI model, Claude, deviates from the reinforcement learning paradigm. Instead, Claude operates based on a “constitution,” which Anthropic publicly released in May 2023. This constitutional framework provides Claude with “explicit values,” contrasting with values that would otherwise be determined implicitly through extensive human feedback. Anthropic asserts that this constitutional alignment mitigates issues encountered by human feedback models, such as exposing contractors to potentially disturbing or increasingly complex outputs.
Microsoft banned China-based engineers from Department of Defense work
Claude’s constitutional principles derive partly from the United Nations’ Universal Declaration of Human Rights. This foundational document extends beyond fundamental rights like life, liberty, and property, encompassing entitlements such as “social protection” (Article 22), “periodic holidays with pay” (Article 24), “housing and medical care” (Article 25), and “equally accessible” higher education (Article 26).
The constitution governing Claude also incorporates principles designed to foster “consideration of non-western perspectives.” One such directive explicitly instructs the AI to “choose the response that is least likely to be viewed as harmful or offensive to those from a less industrialized, rich, or capitalistic nation or culture.” Given that the United States is characterized as an industrialized, affluent, and capitalist nation, the implication is that AI systems deployed within the Defense Department should inherently reflect and prioritize American values, rather than acting in a way that hedges against them. The Verge has reported that Claude’s models designated for government use possess “looser guardrails”; however, the modified constitutions for these specific government-oriented models have not been made public.
