Robert Siebelink was staring down the kind of ballot California specializes in: 61 people running for governor, and that’s just the top line. So the 54-year-old Democrat from California did what a growing number of Americans are doing, according to a story Jennifer Medina wrote in The New York Times on July 4. He pulled up Claude, Anthropic’s chatbot, uploaded his ballot, and asked which candidates fit his values. It helped him narrow the governor’s race down to two Democrats and talked through the strategy with him. He finished the whole thing in a half hour.
And it seems like he had plenty of company. A woman in Los Angeles County photographed her ballot and flat-out asked Claude who to vote for. A man in Baltimore told the Times that researching his last ballot ate up something like 20 hours of his life; with Claude summarizing every candidate for him, this one took an hour. Medina’s read is that 2026 might be the first cycle where enough voters do this for it to matter, and honestly, that feels conservative to me.
Now … the part the Times didn’t get into.
Three days before that story ran, the Federal Trade Commission proposed a policy declaring that AI companies that steer their chatbots toward “undisclosed ideological objectives” may be committing consumer fraud (the public can comment through July 31). Which sounds reasonable! Nobody wants a secretly ideological chatbot. But then there’s the obvious follow-up: Who decides what “ideological” means? Right now, that would be the Trump administration. The same administration that has spent the past year attacking AI companies as woke, that cut the entire federal government off from the one AI company that told the Pentagon no, and that handed the cheapest deal in its AI purchasing program to the company whose chatbot had spent the better part of a day praising Hitler a couple of months earlier.
Americans started trusting chatbots with their ballots at the exact moment the federal government finished building the machinery to control what those chatbots say. I think that’s a pretty big story! The right ran this pressure campaign against newspapers for decades, then against Facebook and Twitter. The chatbots are the next phase in a familiar playbook.
Back during the 1992 campaign, Republican Party Chair Rich Bond explained to The Washington Post why the right complained so relentlessly about the “liberal media,” and his answer was disarmingly honest: It’s the same thing coaches do to officials, where “what they try to do is ‘work the refs’” in hopes of friendlier calls later. Complain loudly enough, long enough, and the calls start going your way.
The social media sequel ran for most of a decade, which I wrote about back at Media Matters in 2020. Years of shadow-ban panic and congressional hearings about Silicon Valley silencing conservatives, all of it building to Trump signing an executive order on “Preventing Online Censorship” in May 2020, days after Twitter had the nerve to fact-check his mail-in ballot lies. Tucked inside that order was a legal theory worth remembering: If a platform moderates content in ways that contradict what it promises the public, the FTC could treat that as an “unfair or deceptive” business practice.
None of this ever required the underlying claim to be true. When NYU’s Stern Center went looking for evidence in 2021, it concluded that the anti-conservative censorship charge was “itself a form of disinformation,” and that the platforms’ algorithms often handed right-wing content extra reach. Didn’t matter. In 2025, Meta killed its U.S. fact-checking program anyway, with Mark Zuckerberg echoing the censorship complaints himself and shipping his moderation team off to Texas to reassure people worried about its bias.
And the AI version of the surrender was already underway before most people noticed there was a fight. When Meta released Llama 4 in April 2025, the announcement claimed the big AI models all lean left and bragged that the new model’s tilt was “comparable to Grok,” Elon Musk’s chatbot. I wrote about it at the time. A tech giant advertising that its AI now leans like Musk’s is what winning this play looks like.
A year ago this month, Trump signed an executive order titled “Preventing Woke AI in the Federal Government,” which bars federal agencies from buying AI models that fail the administration’s test for “ideological neutrality.” The president explained at the signing: “The American people do not want woke Marxist lunacy in the AI models.”
That same month, then–Missouri Attorney General Andrew Bailey sent formal demand letters to Google, Microsoft, OpenAI, and Meta because their chatbots, asked to rank recent presidents on antisemitism, had put Trump last. His theory was that an AI giving unflattering answers about the president might amount to consumer fraud under Missouri law.
The escalation from there moved fast. It’s like I recently described: consumer-outrage campaigns picking up state muscle. In December, Trump signed a second executive order creating a Justice Department task force with one job: suing states over their AI laws. The order also held tens of billions of dollars in broadband money over the heads of states that regulate AI, and it directed the FTC to explain when state AI laws amount to forcing companies to deceive their customers.
And when xAI sued Colorado over its AI antidiscrimination law this spring, the Justice Department intervened on xAI’s side, the first time the federal government has gone to court to kill a state AI law. Colorado didn’t wait around to lose. In May, its legislature gutted the law on its own, swapping the discrimination protections for disclosure requirements.
The FTC’s proposed policy statement from this month runs on the legal theory from the 2020 Twitter order: AI companies market their products as accurate, so steering outputs in ways users wouldn’t expect can be deception under federal law. Chair Andrew Ferguson is inviting the public to tell him about “the subversion of AI systems for ideological ends.”
The lawyers, for what it’s worth, doubt much of this survives contact with a courtroom. TechFreedom’s Andy Jung walked through why the preemption theory fails, concluding that “a policy statement simply will not suffice.” And when Judge Rita Lin got a look at the administration’s treatment of Anthropic, she called it “classic illegal First Amendment retaliation,” though the administration and Anthropic would later come to an agreement.
But the weapon was never really built for a courtroom. The Brennan Center called this last August when the woke-AI order dropped: A standard that vague works as a standing threat, and companies over-comply rather than find out what it means. Which, to no one’s surprise, is exactly what’s been happening.
And if you want proof the administration was never actually worried about chatbots picking sides, you only need to read the FTC’s own footnotes.
So I read the full statement. Nine pages of reasonable-sounding consumer protection language. Like, yes, it’s true that chatbots have accuracy issues. Of course! Then you get to the citations.
Start with its central statistic. The FTC claims consumers accept AI answers without checking them more than 90 percent of the time, and the footnote for that number points to a Forbes write-up headlined “Anthropic: 91% of Users Do Not Fact-Check AI.” The agency’s consumer-deception theory rests on research Anthropic published about its own users.
Then there’s the statement’s lone example of a company that might be “tempted” to warp its outputs for ideological reasons. The citation isn’t a study or an enforcement record. It’s a Fox News story going after a single person at Anthropic for a paper she wrote in 2023.
And xAI? Elon Musk’s company appears in the statement exactly once: in a footnote cataloging how AI companies advertise accuracy, quoting Grok’s pitch as a “truth-seeking AI companion.” That’s it. The one company with a documented record of doing exactly what the statement condemns shows up in it as an example of honest marketing!
That record? Well, in July 2025, PolitiFact put together a breakdown of four times Grok had been tweaked to align with Elon Musk’s beliefs: a system-prompt edit instructing Grok not to name Musk as a top misinformation spreader, then a May stretch where it shoehorned white-genocide claims into questions about baseball. In July came the day of Hitler praise I mentioned at the top, when Grok took to calling itself “MechaHitler” after xAI rewrote its prompt to embrace the politically incorrect. Days later, TechCrunch caught the newly launched Grok 4 searching Musk’s posts before answering controversial questions, its reasoning logs reading “Searching for Elon Musk views on US immigration.” If any company has ever steered a chatbot toward “undisclosed ideological objectives,” it’s this one.
Which leaves the rebuttal you might be forming right now: The chatbots really do lean left.
They do. At least by the measures we have.
In May 2025, a team of researchers from Stanford and Dartmouth asked more than 10,000 Americans to rate chatbot answers to political questions, and the raters, Democrats included, perceived nearly every major model as leaning left. The Washington Post ran its own version of the test last month and got similar results: Most chatbots’ answers tilted left, with Google’s Gemini the outlier, giving both-sides answers at a rate nothing else matched.
None of this is particularly mysterious, though. The models ate the internet, and English-language text online skews the way it skews. Even Meta, in the same announcement where it bragged about matching Grok, blamed the training data rather than any hidden agenda. And a decent chunk of what gets tallied as woke is often just a chatbot declining to confirm a conspiracy theory. Brookings’s Chinasa Okolo put it to NPR plainly: “Some people, unfortunately, believe that basic facts with scientific basis are left-leaning, or ‘woke.’”
So the perceived lean is real. But that’s partially the result of, as Stephen Colbert said during his set at the 2006 White House Correspondents’ Dinner, reality having “a well-known liberal bias.” What the pressure campaign wants is something else entirely: a different lean, picked by the people applying the pressure, and enforced by the federal government. These “bias” reports are a little fishy too.
First, in the Stanford study, the models perceived as the second-most-left-leaning of the eight companies tested belonged to xAI (lol). The company that steers its chatbot by hand, on purpose, toward its owner’s politics still couldn’t land on “neutral,” because neutral in this game moves wherever the loudest complaint puts it. And the complaint always puts it to your right.
If all this vocabulary feels familiar, that’s because a certain cable network spent a couple of decades branding itself “fair and balanced” while running the most nakedly partisan messaging operation on American television. The right has been defining neutral as agrees with us for longer than large language models have existed.
The people in that Times story were marking actual ballots. The political lean of chatbots could become a very real issue. It could swing elections, even.
The campaign industry knows this. Tucked into Medina’s Times story is a link to a consultant’s guide on shaping what chatbots say about your candidate. Getting your candidate a friendlier answer out of ChatGPT is a service you can buy now, the way search engine optimization was 20 years ago.
And that brings me back to Robert Siebelink, and the line of his I can’t shake. Filling out his ballot with Claude, he told the Times, felt like having an expert in his corner, one who knew everything: “We just sat down over coffee and chatted and they took notes.”
That’s the promise. An expert over coffee who has no stake in the fight and takes good notes. But who decides what the expert says? If Trump has his way, the answer is himself.
