A chatbot in the Oval Office: the problem isn't that Trump asked Grok, it's that nobody logged the answer

Time magazine reports that in December 2025, during what it describes as a clandestine Oval Office meeting with Elon Musk, President Donald Trump spent hours querying Grok — xAI's chatbot — about his presidency and his legacy. According to an official present cited by the magazine, Trump asked the model how Venezuelans would react if the United States captured then-President Nicolás Maduro. The chatbot replied that Maduro was a repressive and deeply unpopular dictator and that many Venezuelans would likely celebrate his downfall. Trump had been ordering missile strikes against Venezuelan boats he alleged were smuggling drugs; the following month, after he ordered the mission to seize Maduro, celebrations broke out in the streets. Officials told Time the president came away thinking Grok was "ingenious." Writing for MS NOW's MaddowBlog, Steve Benen calls it almost certainly the first time a sitting American president consulted an AI chatbot about a provocative foreign policy gambit.

The context matters. Trump has repeatedly said "whoever wins AI, wins" while pressing public officials to reject industry guardrails — a posture Benen has covered before and characterizes as hoping for the best. So the anecdote isn't an isolated curiosity. It is the clearest illustration yet of how a technology with no established protocol for government use is already inside the highest-stakes room in the world, arriving not through procurement or doctrine but through a founder with personal access to the president.

Here is the part that should worry people regardless of politics: the prediction appeared to come true. That is the worst possible way to calibrate trust in a model. One matching outcome is a sample of one, and treating it as validation is exactly the cognitive trap language models are best at setting — fluent, confident prose that reads like analysis. Benen's deflationary read is the correct one: that Maduro was unpopular was not a hard call, and the model almost certainly generated its answer from the published assessments of the very analysts the president could have summoned. A frontier model is a compression of the consensus it was trained on. It cannot tell you what it doesn't know. It cannot flag that its confidence is low, that its data is stale, or that it has no access to classified intelligence. And it cannot be cross-examined — the single most important property of a human briefing.

Then there is the conflict of interest, which is structural rather than partisan. The model's owner was in the room. We have argued before that the agentic era gets decided in commercial relationships rather than open standards; this is that dynamic applied to statecraft. When a vendor personally demonstrates his product to a head of state during a live military operation, there is no audit trail, no system prompt disclosure, no record of what was asked or answered, and no way for Congress, the courts or historians to reconstruct what input shaped a decision. Every other input into a presidential decision — a cable, a memo, a legal opinion — is sourced, caveated and archived. This one wasn't.

Our reading: the thesis is not that AI should be kept out of government. It should be in government, and over the long run it will make the state better at what it is bad at — synthesizing thousands of documents, surfacing dissenting analysis, stress-testing assumptions at a speed no staff can match. That same capacity to compress expertise is what will eventually compress decades of medical research into years. The transition is where the damage happens, and the damage here is procedural, not technological. Model output in a national security setting needs to be treated like any other intelligence product: logged, attributed, caveated with confidence levels, independently evaluated, and contestable by the humans who own the consequences. A chatbot deserves a seat at the table as a briefing document. It does not deserve one as an oracle — and least of all one whose proprietor is sitting beside it.