The Trump administration is struggling to replace Claude, the chatbot embedded throughout the Pentagon’s scaffolding, with Elon Musk’s favorite AI system, Grok.
On paper, XAI’s grok makes sense: AI models are already used in select parts of the Defense Department, not to mention other parts of the federal government. Musk should also be deeply familiar with the outlines of the federal government, given that he spent the better half of 2025 scraping wires from its walls.
However, in practice, Grok also has some deep flaws. It performs significantly less AI benchmark tests Compared to other leading models, And it has earned quite a notorious reputation for irregular, disgusting and offensive outbursts.
this is also certain No Federal Insider’s Choice, Who’s Who told wall street journal There are significant concerns about the safety and efficacy of Grok.
Per WSJSeveral executives said Grok is more vulnerable than other AI systems to “data poisoning”, an issue where new information leads large language models to corrupt basic training data. (As you might expect, it goes huge cyber security risksEspecially for an entity like the Pentagon.)
Insiders, speaking anonymously, warned that these concerns reached all the way up to Ed Forst, the head of the General Services Administration, in charge of federal procurement. According to the newspaper report, the GSA considers Groke too sycophantic and susceptible to manipulation.
Put it all together, and until Anthropic refused the Pentagon’s order to remove two key ethical guardrails, military officials overwhelmingly preferred the cloud over Musk’s Grok.
“I don’t think they are equivalent in performance across all the capabilities that matter to a customer like the (Defense Department),” said Gregory Allen, senior AI adviser at the Center for Strategic and International Studies. WSJ.
Complicating matters for Trump and Hegseth, Sam Altman – CEO of Anthropic’s arch rival OpenAI – indicated this week that his company would keep a similar ethical “red line.”
So unless the Trump administration convinces Google or Microsoft to exceed the limits that Anthropic and OpenAI are supporting, the Pentagon is stuck with Grok — facing the consequences.
More on Musk: Man bet his entire life savings of $342,195.63 that Elon Musk would fail
