The free world has found itself a new hero: 43-year-old American, Dario Amodei. Amodei is a cofounder and the CEO of Anthropic, the main rival to OpenAI in the artificial intelligence (AI) market. According to international media, he embodies the resistance to the US administration’s drift towards fascism. After all, isn’t Anthropic ‘that AI startup that dares to contradict Donald Trump’ (Le Monde, 11 February 2026) by refusing to submit to the Pentagon’s diktats? The company whose ‘principled stand’ made Trump’s team ‘livid’ (Fortune, 21 February 2026)? Since his return to the White House in January 2025, Trump’s increasingly overt alliance with tech industry leaders has caused some unease in a reputedly liberal sector. Anthropic has supposedly ridden to the rescue, however, and restored the sector’s reputation by laying down its red lines for the state and military.
But let’s have a look at the facts. In July 2025 Anthropic signed a $200m two-year contract with the US Department of Defense, becoming the first company to use a large language model (LLM) on the Pentagon’s classified networks. Its flagship product, Claude, has been integrated with Palantir’s big data platform, and the whole system is hosted on top-secret infrastructure built by Amazon (1). From its creation in 2021, Anthropic has trumpeted its commitment to ethics, so it imposed two restrictions on its AI models which seem to have been accepted at the time by the military: not to use them for mass surveillance of Americans or to control fully autonomous weapons.
Shortly afterwards, however, Trump issued an executive order aimed at ‘preventing woke AI in the federal government’. This was followed on 11 December 2025 by a memorandum calling on executive departments and agencies to review existing contracts between the federal government and LLM providers, to ensure these models contained no ‘ideological bias’ (2). In February the Pentagon publicly demanded that Anthropic lift its contractual restrictions to allow any application the military deemed lawful – something it refused to do.
The deadline set by the Pentagon of 17.01 on 27 February passed without an agreement. Trump then ordered federal agencies to stop using Anthropic technology, while defense secretary Pete Hegseth designated it a ‘supply-chain risk’, a status previously only applied to a handful of foreign companies such as Huawei, and aimed at depriving Anthropic of lucrative contracts.
Where’s Anthropic’s humanism?
The dispute has all the drama of a soap opera: a Silicon Valley startup that refuses any misuse of its AI model, a CEO who steps up to courageously defend his position, an indignant president and his cronies. Yet, despite appearances, Anthropic’s actions are in stark contrast to its declared humanism.
Its partnership with Palantir, working on behalf of the Pentagon, formalised barely a few days after Trump’s reelection in November 2024, provides a striking illustration (3). Founded in 2003 and initially supported by In-Q-Tel, the venture capital arm of the CIA, Palantir has made mass surveillance its core business. For example, the company provides US Immigration and Customs Enforcement (ICE) with the technology necessary to scale up its deportations. When thousands of Google employees protested in 2018 at the company’s involvement in Project Maven, the Pentagon initiative integrating machine learning into military systems, Palantir, untroubled by such scruples, stepped in to pick up the massive contract.
Its cofounder and CEO, Alexander Karp, presents himself as a progressive but believes in the superiority of Western values and advocates an imperialist approach to international relations as well as the development of autonomous weapons. As for the other cofounder, Peter Thiel – Silicon Valley’s venture capital oracle – his open disdain for democracy and early support for Trump are well known. So there is a fundamental contradiction between Anthropic’s principled objections to the Pentagon, and its collaboration with a company that embodies the violation of those principles.
Once Claude was integrated into software developed by Palantir for the US army, it appears to have quickly established itself as one of the military command’s preferred interfaces. Thanks to Anthropic’s AI, analysts can now use natural language queries to probe vast data sets drawn from images captured by satellites and drones, electromagnetic intelligence, web data and a wide variety of reports. Claude generates summaries, offers analyses and makes recommendations. But it can also identify military targets and rank them according to strategic importance, suggest the most appropriate weapons systems and generate strike sequences almost in real time, even compiling legal arguments along the way to justify them.
In January the press reported rumours that Claude was used in the raid that culminated in Nicolás Maduro’s abduction in Venezuela. But it was the war against Iran launched on 28 February – one day after the Pentagon’s ultimatum to Anthropic expired – that, for the first time, allowed the Pentagon to rely heavily on Anthropic’s chatbot to intensify its bombing campaign, pending a replacement system. Like the AI systems developed by the Israeli army, used to bomb Gaza, or those deployed by China and other countries’ armed forces, the integration of Claude with Palantir’s platforms speeds up the ‘kill chain’, in military jargon (4).
Legal and ethical questions
At a conference on 12 March, Chad Wahlquist, a Palantir executive, cited statistics about the war in Iran: ‘Normally we would have 2,000 intelligence officers, actually trying to do targeting and look at stuff. Now that’s 20 and they’re doing it in rapid succession as well’ (5).’ The use of AI to automatically analyse aerial images has reportedly allowed the number of designated targets to increase from 100 to 1,000 a day; incorporating LLM into these systems would now allow that figure to rise to around 5,000 a day (6). Anthropic sets no guardrails for its partners in this area. Beyond the issue of autonomous weapons, the use of AI to intensify military operations poses huge legal, ethical and political questions, which military leaders and their contractors have so far largely dodged.
The same applies to the other major safeguard that Amodei has demanded of the Pentagon: a ban on using Claude to analyse US citizens’ personal data bought by the agencies from data brokers. Whether it involves accessing geolocation or browsing history, health data or financial information, smartphone applications effectively play a key part in state surveillance. But Anthropic’s apparently brave refusal, upon closer inspection, proves as limited as it is selective. The company does not seem to be concerned by the use of these systems to refine the mass surveillance of non-Americans, contrary to international law, nor the repurposing of the NSA’s foreign surveillance programmes to target US residents – nearly 200,000 such requests are made every year.
After Anthropic was sidelined and excluded from government contracts, its competitor OpenAI announced a deal with the Pentagon. Responding to the outcry, its CEO Sam Altman specified that OpenAI’s systems ‘shall not be intentionally used for domestic surveillance of US persons and nationals,’ particularly in cases where such surveillance would rely on ‘the procurement or use of commercially acquired personal or identifiable information’ (7). This wording may suit the Pentagon better than that required by Anthropic, but it is hard to spot the difference.
Normally we would have 2,000 intelligence officers, actually trying to do targeting and look at stuff. Now that’s 20 and they’re doing it in rapid succession as well Chad Wahlquist
In reality this standoff, supposedly meant to demonstrate an opposition between Silicon Valley and the government, recalls the role-playing that took place after whistleblower Edward Snowden’s revelations in 2013 about the NSA’s digital surveillance. Caught between their cooperation with the security state on the one hand, and the reputational cost of such an association on the other, businesses like Google, Apple and Microsoft put on a show of resistance. They used encryption to better protect users’ privacy, and promoted digital sovereignty to their major foreign clients, pretending to shield them from the NSA. They funded campaigns in Washington for better oversight of intelligence operations.
But behind the apparent concern for human rights, these moves were primarily about reassuring their employees and users, disturbed by the evidence of a symbiotic relationship between digital technology and the US security apparatus. In doing so, they largely succeeded in calming their staff, preserving their collaboration with the security services and consolidating their share of the international market – and helping to reinforce the surveillance infrastructure that underpins US power.
The Trump administration and Anthropic now seem to be singing from the same hymn sheet. By playing the ‘good cop’ who upholds the values of freedom, Anthropic has made itself part of US soft power – which has declined significantly in the past few months – and strengthened its business position. Eighty per cent of the company’s revenue comes from the private sector, and its clients, like the general public, are not displeased by the tremendous publicity Claude has had during this conflict, in which company sources interviewed by the press have not stopped praising the model’s superiority over its competitors. During February Anthropic’s application rose from the bottom of download rankings to first place, and in the first quarter of 2026 its annualised turnover more than tripled, to over $30bn. Next to this, the loss of a $200m contract with the Pentagon seems almost insignificant.
The company’s reputation among tech workers has also been bolstered at a time when there is huge competition for the best AI researchers. Investors do not seem particularly alarmed either: GIC (Singapore), MGX (United Arab Emirates) and the Founders Fund (led by Thiel) contributed to Anthropic’s historic fundraising round of $30bn in February, precisely when its dispute with the Pentagon was at its height. This is healthy portfolio diversification at a time when both Anthropic and OpenAI are preparing for initial public offerings.
The US government has, for the moment, given up on using more coercive legal tools, such as the Defense Protection Act, to force Anthropic to cooperate. What counts, in the end, is that US models are still taking the lion’s share of world markets, managing vast data reservoirs hosted on Amazon, Google and Microsoft infrastructure, and guaranteeing Uncle Sam’s hegemony.
