The machine that matters most is not the one that writes poetry

The public argument over artificial intelligence has become a theatre of extremes. On one side are the prophets of abundance, promising a post-work era in which algorithms do the boring stuff and humans finally move on to more exalted pursuits. On the other are the doomsayers, warning that machine intelligence will soon exceed our own and escape our control. Both camps may be useful for selling books, funding start-ups, or entertaining panel discussions. Neither quite captures the more prosaic and more immediate transformation now underway.

The real power of AI is not that it thinks like a human, but that it does not. It does not tire, forget, hesitate, or feel moral discomfort. It absorbs extraordinary volumes of data, finds patterns too subtle for the eye, and applies them across work, commerce, policing, hiring, credit, advertising, and public administration. That makes it less a rival intelligence than a force multiplier for institutions that already know how to extract value from attention, labour, and behaviour. In the hands of democratic governments, it can improve services. In the hands of firms, it can sharpen profit. In the hands of illiberal states, it can thicken the machinery of control. In all three cases, the common denominator is surveillance.

That is why the most consequential effect of AI may be the least glamorous. It will not necessarily replace the worker in the dramatic sense of a robot taking a factory job, nor will it necessarily herald the extinction of human judgement. Instead it may quietly reorganize the conditions under which people are hired, monitored, disciplined, nudged, and punished. AI is often described as a productivity revolution. It may prove to be that. But it is also a visibility revolution: a way of seeing workers and citizens more closely than ever before, and of using that sight to shape behaviour.

Automation has always been a promise. Surveillance is the delivery mechanism.

Every technological epoch comes with a story about liberation. Mechanization was supposed to lighten toil. The personal computer would democratize knowledge. The internet would flatten hierarchies and amplify speech. Each of these things happened, and yet each also produced new concentrations of power. AI follows the same pattern, but with a darker edge, because its economic model increasingly depends on prediction. To predict well, you need data. To get data, you watch people. To watch people effectively, you monitor them continuously.

That is the logic critics of surveillance capitalism have been describing for years: when the dominant business model is based on extracting behavioural data and selling certainty about what people will do next, the incentives push relentlessly toward more invasive observation. AI does not create this logic, but it intensifies it. It lowers the cost of inference. It turns previously messy human signals into actionable predictions. It makes it easier to infer who is vulnerable, who is likely to churn, who might unionize, who may quit, who deserves a loan, who is worth advertising to, and who should be ignored.

This is why the more sophisticated AI firms become, the less they resemble product companies and the more they resemble intelligence agencies with shareholder reports. Their value lies not merely in the products they sell, but in the data exhaust they capture while selling them. Search, email, maps, social feeds, voice assistants, workplace software, and wearables all become listening posts. The consumer is told these systems are convenient. They are. But convenience is often the velvet glove on the iron fist of data collection.

“When the product is free, the customer is usually not the customer. The customer is the source of the data.”

The problem is not simply privacy in the narrow sense of secrets. It is the erosion of asymmetry. Democracies depend on citizens who can think, organize, dissent, and fail without every action being absorbed into a predictive system. Once that boundary collapses, power migrates to whoever can see most and infer fastest. That can be a platform company. It can be an employer. It can be a police department. It can be a state.

The workplace will be watched before it is fully automated

Much of the public discussion about AI and employment is framed as a body count. How many jobs will disappear? Which occupations will vanish? Which sector is next? These are useful questions, but they encourage a misleading image of labour markets as if they were rows of dominoes waiting to fall. In practice, technological change is usually slower, messier, and more humiliating. Jobs are not always eliminated. They are often thinned out, fragmented, and disciplined.

That matters. A firm that adopts AI rarely does so simply to replace workers overnight. More often it uses AI to reallocate bargaining power. Call centres use systems that transcribe, score, and coach agents in real time. Logistics companies track routes and pace. Retailers analyse shelf behaviour and shrink inventory. Software tools promise to increase output per worker, but they also create a world in which managers can inspect every hesitation, compare every employee to a digital benchmark, and squeeze more labour from fewer people.

In theory, this could raise wages for workers who become more productive. In practice, the gains are likely to be uneven, with the benefits accruing first to owners of capital and to a thin layer of highly skilled supervisors, engineers, and prompt-wranglers. The rest face a harsher regime of measurement. AI, in other words, may not be a simple substitute for labour so much as an instrument of labour discipline. It may not just answer the question, “Can this worker be replaced?” It will also ask, “Can this worker be rendered more legible, more compliant, and more easily managed?”

There is a historical precedent for this. Industrial capitalism did not merely invent machines; it invented new ways of supervising human beings. Time clocks, Taylorism, piece rates, and assembly lines all made work more observable and more controllable. AI extends that logic to cognitive labour, service work, and the home. The result may be a more efficient economy. It may also be a more anxious one, in which every employee is aware that software is always looking over the shoulder.

Surveillance capitalism and digital authoritarianism are not separate stories

It is tempting to draw a bright line between the commercial world of platform capitalism and the political world of state surveillance. In reality, they increasingly reinforce one another. The same tools used to refine consumer targeting can be repurposed for policing, border control, labour inspection, and social scoring. The same data brokers selling audience segmentation can become suppliers of authoritarian infrastructure. The same machine-learning systems that identify purchasing patterns can identify protest networks.

This is where AI becomes politically ominous. Authoritarianism has always required information. It needs to know who is discontented, who is organized, who can be pressured, and who must be silenced. The old-fashioned versions relied on informants, files, wiretaps, and fear. AI offers something more scalable: the ability to aggregate digital traces from phones, cameras, social platforms, financial records, and administrative databases into a single environment of anticipation.

That anticipation changes the nature of repression. A regime no longer has to wait for dissent to become visible. It can try to detect it in advance. Nor must it necessarily imprison millions to maintain compliance. It can create a climate in which people self-censor because they suspect the system is watching. The most efficient form of control is often the one that persuades people to police themselves.

Democratic societies are not immune. Liberal institutions often arrive at surveillance by way of public safety. A city wants to reduce crime. A school wants to detect cheating. An employer wants to detect fraud. A platform wants to improve recommendations. A welfare agency wants to reduce error. None of these goals is absurd. But each creates a rationale for deeper data collection, and each new layer of data can be combined with others in ways that were never explicitly consented to. The danger is not just abuse. It is mission creep with a machine-learning sheen.

The seductive myth of neutral algorithms

Tech companies and their defenders often insist that algorithms are objective, that machines merely reflect patterns in the world, and that bias exists mainly in the data or in the humans who designed the systems. This argument is reassuring because it implies that technical fixes will suffice. Tweak the model, improve the dataset, and the problem disappears.

But algorithms are not neutral in any meaningful political sense. They encode choices about what counts, what matters, and what gets optimized. A hiring model trained to maximize retention may favor people who resemble past hires, thereby reinforcing existing homogeneity. A predictive policing tool trained on arrest data will often reproduce the consequences of prior over-policing. A productivity tool that rewards speed over judgement may quietly punish care, discretion, or creativity. The model may not hate anyone. It does not have to. Structural bias can be produced without malicious intent.

The deeper problem is that AI systems turn human judgement into machine outputs that are easier to defend because they appear scientific. A manager can say the software made the decision. A landlord can say the score was low. A lender can say the risk was unacceptable. This diffusion of responsibility is one of AI’s most corrosive political effects. It does not eliminate power. It launders it.

That laundering is especially dangerous in institutions already inclined toward exclusion. If a state wants to deny benefits, it can use automated fraud detection. If an employer wants to suppress labour costs, it can use algorithmic scheduling. If a platform wants to keep users engaged, it can use recommendation systems that prioritize outrage. If a regime wants to monitor critics, it can fuse facial recognition, data from apps, and predictive analytics. In each case, AI is not an agent of history. It is a supply chain for power.

Why the abundance story is incomplete

It would be intellectually lazy to deny that AI can produce real economic gains. It can speed up drug discovery, improve diagnostics, help teachers draft materials, assist coders, and automate tedious tasks. There is no reason to romanticize drudgery. The problem is not that AI is useless. It is that the distribution of its benefits will not be automatic, and the social costs will not be evenly borne.

Technological abundance is not the same as social abundance. A company can become more productive while its workers become more precarious. A country can create more output while its citizens experience greater surveillance. An employer can reduce labour costs while the local economy loses spending power. The gains from AI, left to market forces, are likely to concentrate among firms with data, compute, and capital, while many workers receive either stagnant wages or a new layer of monitoring disguised as efficiency.

That is why the right policy response cannot be limited to the soothing language of innovation. It must address power directly. Firms that collect and process data at scale should be constrained in what they may capture, retain, infer, and sell. Workers should have rights to know when algorithmic systems are used to evaluate them, and to challenge those decisions. Public agencies should face strong limits on automated profiling and clear duties of transparency. Lawmakers should treat AI not as a novelty to be admired but as infrastructure to be governed.

“The question is not whether machines can learn from us. It is whether institutions will be allowed to learn how to dominate us more efficiently.”

The political choice ahead

There is a comforting fantasy that technological history has an internal moral direction, that innovation naturally bends toward freedom, and that the market will sort out abuses after the benefits have been realized. It is a fantasy because it mistakes adaptation for justice. Societies often accommodate harmful technologies long after their harms are visible, especially when those technologies are profitable, administratively convenient, or culturally intoxicating.

AI today sits at the intersection of three forces that are individually powerful and jointly dangerous: capital’s appetite for prediction, bureaucracies’ appetite for control, and citizens’ appetite for convenience. That combination is what makes it so politically potent. People do not usually embrace surveillance because they love being watched. They accept it because it is framed as a small price for speed, safety, personalization, or efficiency. The bargain is often presented as voluntary. Over time, it becomes structural.

The challenge, then, is not to panic about machine superintelligence. It is to notice how ordinary institutions can become more extractive and more coercive when AI is added to them. The future may not feature robot overlords. It may feature better dashboards, more predictive dashboards, and more people behaving as if the dashboard is destiny.

If the first age of digital technology gave us platforms that harvest attention, the next may give us systems that harvest intent. That is a more intimate theft. It reaches past what we click and toward what we are likely to do, feel, buy, fear, or tolerate. In such a world, democracy does not die in a blaze of machine revolt. It erodes under the pressure of constant observation and optimized manipulation.

The choice is not between AI and no AI. It is between AI governed as a public technology and AI deployed as an instrument of extraction. The former can widen human possibility. The latter will narrow it, even if the quarterly earnings look excellent. The most important question of the AI era is not whether machines will become smarter. It is whether human institutions will become more decent before they become more efficient.