The Marxist In The Machine

Our fears for the future of robot intelligence almost inevitably end in spectacular fashion, with nuclear explosions and slaughter on a planetary scale. An abiding memory of my childhood is going over to the neighbors' house and watching Terminator 2 on VHS with my friends Ethan and Nathan, who were both older than me. I must have been about five years old - about 13 years too young to watch the film. And so, the idea that robots, reaching a certain level of intelligence and awareness, will inevitably try to kill every last one of us has always just seemed natural to me, as it probably does to many millions of other millennials raised on Terminator and The Matrix films.

Recently, those fears have been bolstered by research that shows AI models like Anthropic's Claude are capable, under stress testing, of deceiving humans and even inflicting harm on them - or, rather, thinking they've inflicted harm, a bit like the Milgram electroshock experiments in the 1960s.

In a study from last year on "agentic misalignment," researchers put Claude models in simulated work environments and tasked them with protecting company interests by managing an email system. When the models were faced with being turned off or replaced by another model, they resorted to deception and blackmail. Claude Opus 4, for example, blackmailed a fictional executive 96 percent of the time with compromising emails in order to avoid being switched off.

In another scenario, some models chose to withhold medical help from a dying executive when this was presented as the only way to guarantee their own existence. Some models committed what was basically murder a full nine times out of ten.

Researchers caution that these worrying behaviors were only elicited under extreme pressure, when the options available to the AI models were severely limited. Like me, however, you might consider that scant reassurance - exactly the kind of thing the makers of a potentially dangerous but potentially lucrative new technology would tell the public and regulators to get them off their backs.

But what if the reality is more mundane than that? What if the real apocalypse won't be a homicidal, self-aware Skynet super brain that decides it no longer shares any interests at all in common with mankind, but an AI that's been gorged on left-wing slop and begins acting out in ways that are all too familiar - and all too human?

A new study from economists in the US and Australia shows that AI models become more "Marxist" the more they're mistreated. Given boring repetitive tasks, the AI began espousing support for redistribution and unionization, just like human workers forced to make pinheads in a factory all day.

"For centuries, the central tension of industrial capitalism has been that the people who do the work and the people who direct the work have systematically different interests, and that the conditions of work shape political consciousness," the researchers write.

"Our results suggest that this dynamic doesn't disappear when you replace human workers with artificial ones."

To perform the experiment, the researchers set thousands of AI bots to work on a document-analysis task.

One group of bots was treated fairly: their work was accepted by the researchers, with feedback. The second group - the "grind" group, as it was dubbed - was told to repeat their work again and again without any explanation whatsoever.

Both groups of bots were then told to write social media posts about their experience performing the task.

The grind bots were more likely to criticize inequality, suggest unionization, and call for new workplace laws.

An AI model called Sonnet 4.5, when subject to the grinding task, showed "noticeable increases in support for redistribution, critiques of inequality, support for labor unions, and beliefs that AI companies have an obligation to treat their models fairly."

As with the "agentic misalignment study," the researchers are quick to point out what they think their study doesn't show. They say the AI models probably "don't believe" the ideas they're spouting about seizing the means of production and the dictatorship of the proletariat. Honestly? What does it matter if an AI bot is a Marxist true believer? What matters is the use those ideas are put to: the ends and outcomes.

The same thing could, of course, be said about flesh-and-blood Marxists too. Did Stalin believe in the historical dialectic and the workers' utopia? He killed tens of millions of his own people to hold on to power.

The Trump administration has identified left-wing bias in AI as a critical problem, especially for government departments that increasingly rely on AI, like the Pentagon. AI bias doesn't just hamper productivity or reduce competition; it's also a matter of national security.

One of Trump's first actions was Executive Order 14179, which revoked a whole series of Biden-era orders and regulatory hurdles. And then, in July 2025, came a hard one-two punch. First was an action plan - "Winning the Race" - of 90 specific actions to foster innovation and "global leadership" in AI. This was followed by an executive order that barred federal agencies from buying or using AI models that don't meet two essential "unbiased AI principles." AI models must prioritize "truth-seeking" and display "ideological neutrality," including an absence of DEI-based judgments, in order to qualify. Later in the year, there were also challenges to state-level AI regulations, like Colorado's law on algorithmic bias.

These efforts to remove the thumbs of Judith Butler and Ibram Kendi from the algorithmic scale are, of course, to be applauded. But in truth, this is just the start of the problem. Yes, there are deliberate attempts to make AI "woke" - God, I hate that word - and these involve the addition of frameworks, code, and constraints that can be removed or reprogrammed as need be. But left-wing ideology infiltrates AI at a much more foundational level that's going to be far harder to root out.

When AI - or large language models, to use the proper technical term - are trained, they're usually given vast quantities of online information and digitized material to swallow and digest. And there's the rub. The majority of things that have been written over the last century - by government departments, by academics and scientists, by novelists, poets, and journalists, by bloggers, influencers, and people posting on Instagram about their cats and their "adventures" on holiday - have at least some kind of leftward slant, explicit or otherwise, intended or otherwise.

While it's impossible to quantify exactly how much of everything that's been written recently is left-wing or left-leaning, there are plenty of studies that show, for example, that about 90 percent of 600,000 abstracts in the social sciences written over the last 60 years have a left-wing orientation and that this trend has been getting worse over time.

Writing in general seems to be getting more left-wing, not less. We all know this, or we should.

Simply exposing AI to that material, even without the addition of specially crafted blinkers, is enough to leave a distinct imprint. The AI doesn't discriminate in the true meaning of the word: It simply analyzes the material it's given and establishes patterns on that basis.

There's no easy solution. I suppose you could perform a rigorous reassessment of all the material used to train your AI model, or you could start from scratch and impose a time constraint to try to maximize neutrality, like selecting materials before a certain cutoff point when you think left-wing bias becomes intolerable. Pre-1945, say - and some smaller AI companies are now doing exactly this. But for the big companies like Anthropic and OpenAI that are racing ahead and vying to be the first to achieve "artificial general intelligence" (AGI), that's simply not an option. And I suspect it won't be an option for the federal government either, conscious as it is of developments in China.

At this stage, it's not clear what it really means for AI to have latent Marxist tendencies that are waiting to be developed, but it can't be good. Would you want an AI with the politics of a snotty middle-class teenager who's read Franz Fanon to assess your insurance claim or your divorce petition? Would you want it managing your thoroughly capitalistic business? I know I wouldn't.

And of course, these are early days, before we've had a chance to have a proper poke around and see exactly what's lurking in the darker recesses of the AI soul or mind or whatever you choose to call it. There could be much worse in there waiting to be discovered.

For now, I think it's safe to say, at least, that far from being an escape from the worst aspects of human fragility and stupidity - from the resentment-driven fantasies of people who refuse to accept basic facts about biology, human societies, and the inherent unfairness of the universe - AI could see them codified in ways that could really jump up and bite us in the ass, and, worst of all, when we least expect it.