Anthropic’s prolonged dispute with the Department of Defense unfolded through social media posts, critical public pronouncements, and direct quotes from anonymous Pentagon officials to news outlets. Yet, the destiny of the $380 billion AI startup hinges on a mere three-word phrase: “any lawful use.” These fresh provisions, to which OpenAI and xAI have purportedly consented, would grant the US military unfettered permission to employ these services for extensive surveillance and lethal autonomous weapon systems – artificial intelligence possessing complete authority to identify and eliminate targets without any human involvement in the decision-making process.
The discussions have deteriorated, with Emil Michael, the Pentagon’s CTO and a former senior executive at the ride-sharing company Uber, reportedly spearheading the government’s threats to classify Anthropic as a “supply chain risk,” according to two individuals familiar with the ongoing talks. Such a categorization is typically reserved for national security dangers, encompassing malevolent foreign influence or cyber conflict. Anthropic’s CEO, Dario Amodei, is purportedly slated to confer with Secretary Pete Hegseth on Tuesday at the Pentagon; an undisclosed Defense official characterized this as a “do-or-die encounter.”
The Defense Department’s act of presenting this warning to an American corporation is genuinely unparalleled. However, the Pentagon *openly* voicing this threat makes the situation even more peculiar.
In the interest of safety, the Pentagon refrains from openly revealing which companies appear on such lists, let alone publicly menacing those firms whose perspectives diverge. Indeed, Geoffrey Gertz, a prominent researcher at the Center for a New American Security (CNAS), conveyed to *The Verge* that under existing federal statutes, the Pentagon possessed the ability to categorize Anthropic as a risk without any public notification or explanation of its rationale. “The additional measure of explicitly branding them a national security risk, and striving to prevent other enterprises from engaging in commerce with Anthropic, truly exceeds the norm here.”
The disagreement centers on Anthropic’s adherence to its “appropriate usage guidelines”
Should this designation be formalized, it would terminate Anthropic’s $200 million agreement with the Pentagon, yet its wider implications for Anthropic’s overall financial standing would be far more profound. Leading defense contractors and technology firms, such as AWS, Palantir, and Anduril, utilize Anthropic’s Claude for their projects with the Pentagon, primarily because it was the inaugural AI model authorized for handling sensitive data. To state it more directly: should Anthropic be categorized as a “supply chain risk,” any company presently collaborating with the military or aspiring to secure a military contract would be compelled to discontinue using Anthropic’s AI systems, which are widely considered among the finest in the industry. (On the eve of Amodei’s planned meeting with Hegseth, the Pentagon verified its signing of a pact to deploy Grok, the contentious AI model developed by Elon Musk’s xAI, within sensitive systems. The Pentagon provided no prompt reply following an inquiry for comment.)
This could be enacted in a highly specific application, or, conversely, in a much more expansive manner. Gertz stated, “I anticipate the more plausible rationale would involve a more restricted interpretation, meaning Anthropic cannot be employed within a particular project scope for the Pentagon.” “However, considering some of the accounts and the endeavor to portray this as a retaliatory action against Anthropic, it’s worthwhile considering both of these possibilities.”
Despite the Pentagon and its media supporters having launched an initiative to characterize Anthropic as “woke,” they have thus far presented no genuine allegations concerning security weaknesses or the risk of spying. Rather, the dispute relates to Anthropic’s application of its “appropriate usage policy,” according to individuals conversant with the internal deliberations.
An informant acquainted with the circumstances, who desired not to be identified due to the delicate character of the negotiations, conveyed to *The Verge* that Anthropic has unmistakably expressed its boundaries to the government, highlighting two restricted matters the company will not assent to: self-governing offensive actions and widespread internal monitoring. The reason for the latter, the informant explained, is that “legislation hasn’t kept pace with AI’s capabilities,” and it could potentially encroach upon American civil liberties. Regarding the former — deadly self-governing armaments — the informant stated that the capability “is not yet developed for fully independent armaments lacking human oversight.”
Hamza Chaudhry, the principal for AI and national security at the Future of Life Institute, a neutral research organization centered on AI governance, observed that Anthropic’s stated boundaries already mirrored existing government mandates that have not been revoked.
“DoD Directive 3000.09 stipulates that all autonomous weapon systems must be constructed to enable commanders and operators to ‘apply suitable levels of human discretion over the application of power,’ and the Political Declaration on Military Use of AI, initiated by the US Government and supported by 50 states, upholds this fundamental tenet,” he conveyed to *The Verge* via message. “Furthermore, DoD Directive 5240.01, strengthened by stipulations within the FY2017 NDAA and the Trump-era Responsible AI Implementation Pathway, forbids intelligence components from gathering data on U.S. persons unless operating under specific legal frameworks like FISA or Title 50.”
“Anthropic’s appropriate usage policy mirrors these very boundaries, and until the Pentagon formally disavows, elucidates, or revises these policy positions, the significant query remains whether the company can be forced to abandon a policy that the government itself has theoretically adhered to.”
Michael, an individual appointed during the Trump administration who serves as the Undersecretary of Defense for research and engineering—a role frequently likened to the Pentagon’s primary technology executive—is representing the Pentagon in discussions. A [first source] characterized Michael, known for cultivating a formidable image as Uber’s head of commercial operations and having previously boasted of gathering detrimental information on journalists, as an “unyielding bargainer.” (Michael departed from Uber in 2017, following an inquiry by the company’s governing body into its culture of inappropriate conduct, which was triggered by his and several senior staff’s visit to a South Korean escort bar.)
“This is genuinely an issue of conviction for Emil,” stated a second individual acquainted with the situation, indicating Michael’s displeasure that a commercial entity sought to restrict the state’s deployment of its innovations. It remains uncertain whether the White House or David Sacks, the venture capitalist and influential AI and cryptocurrency authority, had sanctioned Michael’s aggressive strategies beforehand.
Presently, Anthropic’s “acceptable use policy” is integrated into an agreement valued at $200 million it concluded with the Department of Defense the previous July. During its public statement, the company referenced “ethical AI” on five occasions. “Central to this endeavor is our belief that the most potent innovations entail the utmost accountability,” they penned, asserting that within a governmental framework, “where choices impact millions and the repercussions are immense,” accountability was “vital” to guarantee that “AI progress reinforces democratic principles worldwide by preserving technological dominance to safeguard against autocratic exploitation.”
“The stipulation would compel each military vendor pursuing public contracts to attest to the removal of all Anthropic’s technological components from their platforms”
However, in January, Hegseth issued a memorandum declaring that the department would transform into “an ‘AI-first’ combat entity across all its branches” and that the phrase “any legitimate application” ought to be integrated into any acquisition agreement for AI services within six months, encompassing current directives.
Within Hegseth’s memorandum, he emphasized multiple times that the department would give precedence to swiftness regardless of expense, stating that the nation needs to “remove impediments to information exchange … [and] address compromises in risk, ‘fairness,’ and other discretionary matters as though engaged in conflict.” He further mentioned that concerning the creation and testing of AI agents, the department would incorporate them “from strategic campaign formulation to operational elimination sequences,” as well as convert “intelligence into armaments within a few hours.”
Hegseth consistently favored swiftness above security and prospective inaccuracies: “It is imperative to acknowledge that the perils of insufficient rapidity supersede the hazards of flawed concordance.” He reiterated his stance further on in the memorandum, stating that “ethical AI” would undergo significant transformations within the agency, both in combat zones and among armed forces personnel. “Variety, Fairness, and Integration, alongside societal viewpoints, are unsuitable for the DoW,” he penned, further stating that the department “is also required to employ models unburdened by usage restrictions that might impede legitimate military uses.” Echoing Trump’s directive against “politically correct AI,” Hegseth announced that standards for model impartiality would serve as a novel chief acquisition metric for AI offerings.
OpenAI, xAI, and Google promptly re-evaluated their respective $200 million agreements with the Pentagon to conform to Hegseth’s memorandum. However, none of the models from those firms possess an Impact Level 6 security clearance, signifying that ChatGPT, Grok, and Gemini would be unable to instantly substitute Claude if Anthropic were to be banned—a reliance on one provider that could disadvantage the Pentagon.
“Claude is the sole cutting-edge AI model functioning on highly sensitive Pentagon systems, implemented via Palantir’s AI Platform and Amazon’s Covert Cloud, implying its central role in processes inaccessible to most other models currently,” Chaudhry observed. “The stipulation would compel each military vendor pursuing public contracts to attest to the removal of all Anthropic’s technological components from their platforms.”
This has provided Anthropic an advantage in its disputes with the Pentagon, which have become more severe following the company’s alleged discovery that its AI tools aided in the apprehension of Venezuelan President Nicolás Maduro, breaching their existing accord.
Anthropic legally cannot endeavor to collaborate or unite with the other AI development firms presented with the updated conditions, even if they were amenable to acceptance, as such actions would contravene federal acquisition regulations. However, given that this conflict is unfolding publicly, technology professionals, AI staff, and individuals presently or previously employed in the tech sector have voiced annoyance that rival firms are not contending for identical conditions as Anthropic. Some speculated it was merely a question of time until Anthropic capitulated.
“It would be a particularly opportune moment for [other labs] to ask, ‘Hold on, how is our technology being utilized?’” stated William Fitzgerald, an ex-Google staffer who currently operates a lobbying company called The Worker Agency. “Individuals at these AI research facilities wield considerable influence. They consist of smaller groups, and are still largely defining their identity… I genuinely believe they can validate their worth without engaging in military projects. Alternative methods exist to operate a venture without incorporating lethal outcomes into its operational framework.”
{content}
Source: {feed_title}