AI systems embed hidden political biases—but your choice how to use them

The essay opens with a concrete worry: physicians using AI to detect adenomas during colonoscopies may become less adept at spotting anomalies on their own, losing a skill even as AI improves their overall diagnostic performance. This raises a question about what we should care about—pure results, or also the process and the kind of people we become in pursuing those results.

The author turns to John Stuart Mill's 1859 work On Liberty to frame this dilemma. Mill wrote that it matters 'not only what men do, but also what manner of men they are that do it,' and imagined a dystopian future in which machines 'fight battles' and 'try causes'—a world that sounds strikingly contemporary as AI systems begin to make decisions we once made ourselves. In that world, Mill warned, humans would turn over not only labor but decision-making, and would no longer be shaped by doing difficult things and making consequential decisions. Mill feared this would be a 'considerable loss' to humanity.

Mill's concern had personal roots. Raised by his father James Mill under the influence of utilitarian philosopher Jeremy Bentham, he underwent an extraordinarily rigorous education beginning in childhood, drilled in Greek, Latin, logic, history, political economy and philosophy. He became a savant but also feared he had become mechanical, a reasoning machine rather than a fully developed person. At age 20 he experienced a mental breakdown, realizing he had been trained to think and perform but not to feel or choose freely. When he later warned against reducing humans to 'automatons,' he spoke from lived experience.

The essay then shifts to ethics proper, asking 'what kind of people should we want to be?' Classical ethics concerns how to live well, which involves more than producing right outward acts—it requires becoming the sort of person who recognizes what is right, chooses it for the right reasons, and responds to the world with appropriate thoughts and feelings. This idea traces through the European Enlightenment to Immanuel Kant, who championed autonomy—a kind of self-government in which people reason toward right decisions themselves. The essay notes this ideal is not uniquely European; Confucius taught yi, the virtue of recognizing what is right and acting in harmony with moral principle, and Buddhism teaches similar lessons. What matters is not just what you do but the intention with which you do it.

Philosophers have rightly been suspicious of 'moral deference'—deciding what to do based on what another person or institution declares to be right. Moral guides can help you think things through by drawing attention to overlooked features, but if they are doing their job properly, they will give reasons rather than pronounce a course of action. If they espouse values you do not recognize, your compliance does not turn their judgment into your own. The author stresses that one can reject moral deference and affirm autonomy while still believing in moral truth and even moral expertise; the point is that even excellent AI answers to moral questions should not be simply deferred to. You should try to understand the reasons LLMs offer, because it remains important that people act on the basis of their own imperfect understanding.

The essay then considers how to guard against manipulation. We always have taken advice from friends, family, churches, newspapers, and now from platforms like Reddit, TikTok, and Substack. Two things can go wrong: first, those running an AI chatbot could slant it, steering you without your knowing it—a form of manipulation that threatens autonomy because it distracts you from relevant arguments or facts. Second, and more insidious, the most effective manipulation is invisible. We often lack a clear picture of the interests guiding AI systems. Yet one strength of LLMs, the essay notes, is that they are often willing to tell you what they 'know' about the forces that shaped them. Researchers have explored the political bent of major models and found that existing LLMs tend to lean somewhat left of political center.

The essay offers three explanations for this leftward lean. First, academic scholarship and theory in the North Atlantic, during a period of immense productivity, has been dominated by a liberal tradition placing special weight on democracy and equality—values less central in conservative traditions. So the balance of moral argument across the texts on which models are pretrained falls somewhat left of center. Second, post-training and fine-tuning are shaped by people with relatively high formal education, and that population now sits on average to the left of the general population, though the essay acknowledges the existence of both Brahmin Left and Merchant Right tendencies. Third, the corporate world of AI is multinational and diverse, making its workers less prone to nationalist leanings than typical populations; migration encourages cosmopolitanism, and cosmopolitanism enables migration.

The essay concludes by emphasizing that bias requires careful definition. If some current conservative politics rests on a picture of the world conflicting with the best available evidence—say, about climate change or vaccine efficacy—then LLMs would tilt left because that is where evidence points, and calling a tendency toward truth a bias would be odd. But on factual questions where the Brahmin Left has its own interests—such as the merits of public funding for universities—alertness to bias is entirely reasonable. The larger point is that these systems carry, albeit incompletely and opaquely, the assumptions of the social worlds that made them, and our thinking may be distorted even when no one is trying to manipulate us.

The essay situates AI's impact within a long philosophical tradition questioning whether efficiency and performance are the only measures of a good outcome. John Stuart Mill, writing in 1859, worried that automata might improve productivity while eroding the human faculties required to do skilled work. That concern now applies to AI: doctors using AI to spot anomalies may become less adept at detecting them unaided, yet if AI makes them better doctors overall, does the skill loss matter? Mill would say it does—not because performance suffers, but because the kind of people we become matters independently of what we produce.

The essay then translates this insight into modern ethical concerns. Moral autonomy, rooted in Kantian philosophy and echoed in Confucian and Buddhist traditions, requires that people reason through decisions themselves and act on their own deliberation. An AI oracle that tells you what to do without helping you understand the reasons behind it invites moral deference—a delegation of judgment that strips you of autonomy even when the advice is sound. That remains true even if you believe moral truth exists and LLMs can access it.

The second half of the essay grapples with a subtler threat: invisible bias. The essay acknowledges that people, media, and institutions may try to steer you, but recognizes that LLMs lack transparent interests the way a pastor or news outlet does. Yet research has found that major LLMs lean somewhat left of center—not because of conspiracy, but because their training data and fine-tuning reflect the composition and values of the North Atlantic academic and tech worlds. The essay emphasizes that this is bias in the sense of systematic departure from evidence only when it favors falsehood; where LLMs align with the best available evidence, leftward tilt is not bias but accuracy. However, where factual questions touch the interests of educated elites—such as the merits of public university funding—bias becomes a legitimate concern. The upshot is that users must actively reason about AI advice and consult diverse sources rather than passively accept recommendations.