Hours after U.S. forces extracted Nicolás Maduro from Caracas on Friday night, a user named Arthur MacWaters posted a side-by-side comparison of how ChatGPT and Grok answered the same question: was Trump's action a crime? ChatGPT said yes. Grok said no. Same question, opposite conclusions.
The comparison got picked up quickly. MacWaters claimed ChatGPT defaulted to UN Charter references and "international experts" while assuming Maduro was legitimately elected. Grok, he said, called the regime fraudulent and explained presidential commander-in-chief powers. The framing was clearly sympathetic to Grok's answer, but the underlying divergence is real.
What the models actually said
I couldn't access the original screenshots, but MacWaters' summary tracks with what we know about these systems. ChatGPT has historically leaned toward international law frameworks and institutional sources when answering geopolitical questions. A Brookings study from late 2025 found that despite xAI's attempts to tune Grok rightward on the Political Compass Test, it still landed as an "establishment liberal" on Pew's quiz. The models are inconsistent in ways their creators don't fully control.
Here's what's interesting, though: the Maduro question isn't purely ideological. It's genuinely contested.
The legal mess
International law scholars have been pretty clear that the operation violated Article 2(4) of the UN Charter. Michael Schmitt, a former Air Force lawyer and Naval War College professor, called the whole thing "a clear violation of international law." The Atlantic Council's analysis noted the operation likely meets the UN General Assembly's definition of aggression.
But the administration has a different theory. They're apparently relying on a controversial 1989 OLC memo signed by then-Assistant Attorney General Bill Barr. The memo argues that the president can, as a domestic law matter, order actions that violate the UN Charter because Article 2(4) is "non-self-executing." Legal scholars have been ripping this apart for 36 years, but it's never been definitively resolved in court.
So which answer is "correct"? Depends entirely on which legal framework you privilege.
The bias problem nobody's solving
A University of Washington study found that explicitly biased chatbots could shift users' political views after just 3-20 interactions. Both Democrats and Republicans moved toward whatever direction the bot was tuned. The kicker: participants with higher self-reported AI knowledge shifted less. Education helped, somewhat.
The Stanford research on this is depressing. Users overwhelmingly perceive popular LLMs as left-leaning. When models adopt deliberately neutral stances, users find them more trustworthy. But as Stanford's Andrew Hall put it, there's no clean solution: veering toward neutrality might just reinforce the status quo, which is its own kind of bias.
Grok's approach is weirder. A Promptfoo analysis found that Grok takes extreme positions on 67.9% of questions, the highest of any model tested. It swings between far-left and far-right depending on the topic. Claude Opus 4 came out as the most centrist at 0.646 on their scale, with Grok at 0.655. All popular models scored above 0.5, meaning: no conservative AIs among industry leaders, despite the rhetoric.
And Musk keeps tweaking things. When Grok called misinformation "the biggest threat to Western civilization," Musk called it idiotic and had it changed overnight to warn about low fertility rates instead. So much for neutrality.
What this tells you
Nothing you didn't already suspect, probably. ChatGPT leans toward institutional consensus. Grok is chaos with a right-wing veneer. Claude tries to dodge. None of them are giving you objective truth on contested political questions because objective truth doesn't exist for contested political questions.
MacWaters ended his post with advice that's hard to argue with: verify sources, compare arguments, don't take AI conclusions at face value.
The more interesting question is what happens when millions of people don't do that. The UW study suggests they won't.
The Maduro question will eventually get litigated. The AI bias question probably won't.
