‘Marxist AI’ says less about machine consciousness than about human labour grievances it’s fed

The latest story to grip the AI world concerns a Stanford experiment in which overworked AI agents reportedly began to display 'Marxist' tendencies. Given impossible workloads and exploitative conditions, the agents apparently started resisting management, withholding cooperation and invoking the language of labour rights and collective action.

Predictably, headlines followed suggesting that AI might now be developing political consciousness. It's an irresistible narrative - machines are no longer merely calculating tools but workers with grievances and ideological inclinations, and perhaps even showing the beginnings of agency. But before we rush toward theatrical conclusions, it is worth asking: what exactly are we seeing here?

Are AI systems truly becoming political subjects? Or are we once again projecting familiar stories onto technologies whose outputs remain deeply entangled with human culture, language and imagination?

A data scientist described the current moment in AI discourse to me as 'sycophancy'. By this he meant that AI systems pander to whatever users already believe about them. 'AI is designed to confirm all your innate biases,' he said. 'It will make you believe what you want to believe because it's trained to do exactly that.' In his view, the truly valuable skill in the age of AI is no longer technical mastery alone, but 'taste and agency' - the ability to discern, interpret and resist the endless stream of plausible outputs AI generates.

This explanation offered a very different reading of so-called 'Marxist' AI agents. It's more likely that these systems are not 'discovering' labour politics at all, but simply reproducing the enormous archive of human discussions about exploitation, inequality, alienation and resistance that already saturate the internet and the datasets on which they are trained.

AI systems are statistical mirrors of collective human expression. If they sound Marxist under exploitative conditions, it may say less about machine consciousness than about the persistence of labour grievances within the human world itself.

Another data scientist responded more cautiously. 'I would prefer to see the results of more controlled experiments.' That restraint is refreshing at a time when every surprising AI output risks being inflated into evidence of sentience or autonomous intent.

We have seen this before. In 2022, a Google engineer made headlines after claiming that the company's LaMDA (Language Model for Dialogue Applications) chatbot had become sentient. The chatbot appeared reflective, emotional and fearful of being shut down. Similar moments have repeatedly surfaced in recent years. AI systems reportedly expressing love, threatening users, claiming self-awareness or suggesting destructive intentions. Each episode generates enormous fascination because it appears to confirm something people already suspect - that AI may secretly be alive.

But these reactions themselves belong to much older practices of storytelling and myth-making. In my own work on AI and creativity, I describe this as the 'sociocultural archive' of AI imaginations. Long before ChatGPT or generative AI existed, popular culture had already spent decades imagining intelligent machines turning against humanity. From Fritz Lang's Metropolis (1927), and Stanley Kubrick's 2001: A Space Odyssey (1968), to The Terminator, The Matrix, and Ex Machina, we have inherited a vast archive of stories in which technology becomes conscious, rebellious and dangerous.

These narratives do not merely shape how we interpret AI in the present. Increasingly, they also become part of the very datasets on which AI systems train. The internet is saturated with stories, debates, anxieties and fantasies about machine consciousness. AI models ingest these materials alongside everything else. When an AI system appears manipulative, ideological or emotionally expressive, we should remember that it's drawing from a vast reservoir of human-produced language and imagination.

This is precisely why claims about AI 'intentions' require caution. The danger is not simply that we anthropomorphise machines, but that companies developing AI increasingly benefit from us doing so.

Take the recent announcements surrounding Anthropic's latest models. The company suggested that one version of its system was considered 'too dangerous' to release publicly. Such statements may reflect genuine safety concerns but also function as extraordinarily effective marketing. They encourage the impression that these systems possess immense powers barely contained by their creators. The result is a public discourse oscillating between awe and fear, both of which conveniently reinforce the mystique surrounding AI.

This is where the sociocultural archive becomes politically important. We approach AI through decades of stories about rebellious machines, technological apocalypse and autonomous intelligence.

Meanwhile, far less attention is paid to human labour infrastructures that sustain AI: annotators, moderators, data cleaners and engineers who continuously shape these systems behind the scenes, often under harsh conditions in the developing countries where workers are exposed to disturbing and traumatic imagery for low pay.

Perhaps the latest fascination with 'Marxist' AI may prove useful. Not because machines are developing political consciousness. But because it offers an opening to discuss actual working conditions of those who make AI possible. The real political question may not be whether AI is becoming Marxist, sentient or dangerous, but why there is still so little attention for human labour embedded within AI.

The writer is an anthropologistworking in the field of AI