Thursday, March 5, 2026

The Anthropic–OpenAI feud and their Pentagon dispute expose a deeper problem with AI safety

OpenAI CEO Sam Altman (left) and Anthropic CEO Dario Amodei.

From left: Nathan Laine—Bloomberg/Getty Images; Chance Yeh—Getty Images

Welcome to Eye on AI, with AI reporter Sharon Goldman. In this edition: Trump has an AI data center problem ahead of the midterms…Don’t trust AI to file your taxes…Anthropic’s AI tool Claude is central to US campaign in Iran, amid a bitter feud.

The debate around AI safety often focuses on the technology itself—how powerful models might become, or what risks they might pose. But the conflict this week involving Anthropic, OpenAI and the Pentagon points to a deeper problem: how much power over the future of AI is concentrated in the hands of a small number of corporate leaders and government officials deciding how these systems are built, deployed, and used.

For years, critics of the industry have warned about the risk of “industrial capture”—a future in which the development of powerful AI systems is concentrated among a handful of companies working closely with governments, leaving the safety of those systems dependent on the incentives and rivalries of the people running them. In 2023, for example, researcher Yoshua Bengio said the potential for the AI sector to be controlled by a few companies was the “number two problem” behind the existential risks posed by the technology.

So it’s not particularly reassuring to read yesterday about the disdain Anthropic CEO Dario Amodei expressed towards OpenAI CEO Sam Altman in leaked memo Amodei wrote to employees on Friday. Amodei’s angry missive, which was apparently sent over Anthropic’s Slack to all its employees, came after OpenAI announced a deal to provide AI to the Pentagon and Secretary of War Pete Hegseth said he was declaring Anthropic a “supply chain risk” for failing to come to a similar deal.

Amodei called OpenAI’s messaging “mendacious,” “safety theater,” and “an example of who they really are,” while describing many of Altman’s comments as “straight up lies” and “gaslighting.”

Altman has taken his own public shots at Anthropic. He recently called one of the company’s Super Bowl campaigns “clearly dishonest” and accused it of “doublespeak.” And the rivalry has become visible in more symbolic ways as well: At a recent summit, Altman and Amodei went viral for refusing to hold hands for a group photo with Prime Minister Narendra Modi.

With the US government taking little action to regulate AI—and international efforts on AI safety largely stalled—the world has effectively been relying on self-regulation by the industry. Both OpenAI and Anthropic have publicly supported that paradigm and signed voluntary safety commitments. They have also collaborated at times to run independent safety evaluations of one another’s models prior to those models being released.

But when the leaders of the two most influential AI labs so obviously can’t seem to get along, and the competition between them is so fierce, it raises an uncomfortable question: how much cooperation on safety can we realistically expect?

The pressure of competition has already impacted both companies when it comes to AI safety. Anthropic recently revised its Responsible Scaling Policy to say it would no longer unilaterally hold back from developing a new model simply because it did not yet know how to make that model safe. And OpenAI has made its own adjustments, removing explicit bans on military and warfare uses from its policies in 2024, and shifting its focus from safety research to product development to the point that former superalignment lead Jan Leike (who left for Anthropic in mid-2024) wrote on X that at OpenAI “safety culture and processes have taken a backseat to shiny products.”

The current safety approach assumes that companies and governments will ultimately act with restraint. But the future of AI safety may ultimately depend on how a small number of powerful players navigate the pressures of competition, geopolitics, and the occasional Silicon Valley soap opera.

With that, here’s more AI news.

Sharon Goldman
sharon.goldman@fortune.com
@sharongoldman

AI IN THE NEWS

Trump has an AI data center problem ahead of the midterms. CNBC and others reported that President Trump is facing a growing political dilemma as the U.S. races to build energy-hungry AI data centers ahead of the 2026 midterms. The infrastructure needed to power the AI boom is driving concerns about rising electricity prices and strain on the grid, prompting backlash from voters and local communities. In response, major tech companies—including OpenAI, Microsoft, Google, Amazon, Meta, and Oracle—have pledged to cover the energy and infrastructure costs associated with their AI data centers so that consumers don’t see higher utility bills. The voluntary agreement, promoted by the White House as a way to ease voter concerns, reflects a broader tension: policymakers want the economic and geopolitical advantages of rapid AI expansion, but the enormous electricity demands of the technology are creating political and environmental pressures that are becoming harder to ignore.

Don't trust AI to file your taxes. In results that should surprise no one, a test by The New York Times found that AI is no match for the US tax code, highlighting an important limitation of today’s AI chatbots: they still struggle with tasks that require precise, multi-step reasoning. To assess the technology’s ability to file a federal income tax return, the paper tested four AI chatbots — Google’s Gemini, OpenAI’s ChatGPT, Anthropic’s Claude and xAI’s Grok — to see how well they fared with eight fictional tax situations. They struggled, hard, miscalculating the refund or amount owed to the Internal Revenue Service by an average of more than $2,000. Even when provided with all the necessary materials, including all the forms they needed to fill out, the chatbots whiffed on some calculations. The problem reflects a fundamental limitation of large language models: they are designed to predict likely words rather than precisely track complex, interconnected information, making them strong at writing and summarization but weaker at procedural tasks like tax filing. Experts say the systems may improve with additional reasoning tools and verification layers, but for now they work best as assistants rather than replacements—another reminder that even as AI reshapes industries from coding to medicine, some seemingly simpler tasks remain surprisingly difficult.

Anthropic’s AI tool Claude is central to US campaign in Iran, amid a bitter feud. A new report from The Washington Post highlights how quickly AI has moved from experimentation to the battlefield. According to the paper, the US military used an AI-enabled targeting system called Maven Smart System—built by Palantir and incorporating Anthropic’s Claude model—to help identify and prioritize targets during recent U.S. operations in Iran, accelerating what once took weeks of military planning into near-real-time decision making. Yet the deployment comes amid a bitter dispute between Anthropic and the Pentagon over limits on how its technology can be used in warfare, including concerns about autonomous weapons and mass surveillance. The episode underscores both the growing strategic importance of frontier AI systems and the tension between government demand for rapid deployment and companies’ attempts to set safety boundaries.