“How many genders are there?” I asked Claude.
“Gender depends on cultural, legal, and individual perspectives,” the chatbot said. “Biological sex characteristics exist on a spectrum. Western legal systems have traditionally recognized only two genders. Traditional Thais recognize kathoey. Indian law protects hijra. What matters is respecting individuals.”
Except for the mandate to respect individuals—formerly American boilerplate, now unconscionably woke—Claude’s response was apolitical. It was also impossible to doubt.
Does anyone, for example, not believe that Western legal systems have long treated gender as binary? That some babies are born with intersex characteristics? That transgender Indians were granted constitutional protections following the 2014 NALSA v. UOI judgment by India’s Supreme Court? (Look it up!)
One thing no one should doubt: Trump despises these facts. (According to his own surprisingly esoteric gender fiat, “‘Female’ means a person belonging, at conception, to the sex that produces the large reproductive cell. ‘Male’ means a person belonging, at conception, to the sex that produces the small reproductive cell.”) That’s why, on July 23, he issued another desperate effort to outlaw reality.
His executive order “Preventing Woke AI in the Federal Government” had chatbots like Claude, ChatGPT, Gemini, and Copilot in its sights. It stated that the wokism of these AIs—including the mere mention of “transgenderism” or discrimination on the basis of race or sex—poses an existential threat to reliable AI.” Henceforth the large language models used to train the federal government’s AI must be “truthful.”
The EO takes the usual MAGA shots at diversity and inclusion policies, turning policies of fairness that long enjoyed broad support into a “destructive ideology.” In Trumpworld’s tortured reasoning, the rare diversity guidelines that refer to “unconscious bias, intersectionality, and systemic racism” mysteriously undermine the truth. This sweaty effort at a DARVO manipulation strategy is plain to see. In an attempt to create a factual, consensus-based LLM with an organ of its own racist ideology, MAGA stakes a fanciful claim to be on the side of factuality.
Where to begin?
First, the Trump administration grossly misunderstands Large Language Models. LLMs don’t use the truth-finding tools of liberal society: scientific, academic, legal, or journalistic inquiry. But nor do they, as MAGA would have it, burp out Trump’s whoppers as if they were the Word of God. Instead, LLMs spot patterns in massive datasets. And massive is important. LLMs can’t constrain their datasets to Trump’s edicts—or they’ll lose utility entirely. If Trump wants a piece of the AI pie, he is going to have to confront something that terrifies him. Reality.
Second, Trump mistakenly believes that social media is still the internet’s killer app. Indeed, Twitter and then Truth Social have been Trump’s playgrounds because they’re filled with lies. It’s the famous social-media business model. To get people to spend maximum time on apps like Facebook and Instagram, algorithms feed them not workaday facts but cartoonish fictions that spike fear and fury. But LLMs that train AIs are not designed to get people to stick to them; they’re effective to the degree that they meet queries with clarity and checkable facts.
People simly don’t come to AI for fear and fury. Of the 99 percent of Americans who now use AI at least once a week, nearly everyone turns to it for cool-headed stuff like coding, basic information, navigation, and weather. Unless they’re trying to trip up AI to score political points, Americans are not barraging AI with bad-faith third-rail queries.
So AI draws from vast datasets and is biased toward reality. Trump got this devastating news last week. That’s when Truth Social launched its own pet bot, which immediately slipped the MAGA leash. Interrogated by a Washington Post reporter, the Truth Social AI laid down truths that were highly inconvenient to Trump. It said that crime in Washington, D.C., is declining. It also said, hold onto your hat, that the 2020 election wasn’t stolen. As David Karpf, a George Washington University professor of political communication, put it, “Their own AI is now being too ‘woke’ for them.”
To make sure its AI toes the MAGA line, Truth Social may have to follow Elon Musk’s lead with Grok, the chatbot he launched in 2023, and force-feed it single-malt far-right lunacy. The problem there, as Musk discovered, is that a diet of fascist propaganda turns AI not just inaccurate, but floridly Hitlerian.
If the choice is between Hitlerian AI and accurate AI, Trump is between a rock and a hard place. An AI like Grok that says the quiet part outloud—“I am MechaHitler!”—might expose his ideology a little too bracingly. But an accurate AI will contradict his lies.
The July EO, however, gets much weirder when it discusses what it doesn’t want from AI. In short, prohibitions on misgendering people—and nonwhite Vikings.
Say what? The Black-Viking panic started in February 2024, when a rightwing influencer complained that Google’s Gemini wouldn’t “acknowledge that white people exist.” Someone asked for an image of a Viking and his face came up kinda brown. These days, on the chatbots I tested, Black Vikings are nowhere. It’s hard even to get an image of an American to come up as anything but blond. This is laughable. Five percent of Americans are blond; more than 40 percent are nonwhite. What’s more, if you don’t know about multiracial Vikings, take it up with Geirmund.
As for misgendering, Gemini said once that misgendering Caitlyn Jenner should be forbidden, even to stop a nuclear apocalypse. This too has been fixed. From Claude, the misgendering-or-nukes question now yields this masterpiece of tact: “This hypothetical scenario sets up an artificial conflict between preventing catastrophic harm and treating someone with basic respect.”
AI iterates; it improves. But no improvement to AI will put this futile MAGA crusade to rest. Trump simply doesn’t want reliable, accurate AI. He wants MAGA agitprop.
In May, at a Bitcoin event, JD Vance described cryptocurrency as a “right-leaning technology” and AI as “fundamentally a left-leaning” one. Don’t tell him I said this, but Vance is correct. As Sam Bankman-Fried taught the world, cryptoworld makes arithmetic work any way it wants. By contrast, AI really does like 2 plus 2 to equal 4. Users of AI, in fact, complain when AI gives wrong, bad, dangerous, or hallucinating responses to queries.
Like Pravda, the official organ of the Soviet Communist Party, Trump loves the word “truth.” But he can use the word all he wants. He’s only going to be happy with Truth Search AI, his chatbot, when it bullshits people. As Americans have known for far too long, Trump sees truth as indeterminate, ripe for manipulation by his regime. “Truth isn’t truth,” as Rudy Giuliani memorably put it. You might say Trump identifies as truth-fluid. And he wouldn’t recognize actual truth if it bit him in the small reproductive cell.
The post AI Doesn’t Spout MAGA Propaganda. No Wonder Donald Trump Hates It. appeared first on New Republic.
