The Cost of the Machine: AI, Hype, and the Slow Erosion of What Makes Us Human

 

The trouble with today's AI boom is not simply that the claims are exaggerated. It is that the exaggeration conceals a very material ledger of costs — more electricity, more water, more displaced workers, and a steadily degrading information environment on which the next generation of AI systems will depend. Before we can assess those costs honestly, we need to understand how the exaggeration works.

The Hype Machine and What It Hides

For the past several years, the public has been told a story with the cadence of a creation myth. AI, we are told, is approaching a new kind of mind. Every software anomaly gets reframed as a sign of "emergence"; every engineering failure gets repackaged as evidence that the machine is becoming strategic, or clever, or even proto-conscious. When an AI model produces a subtly inconsistent output during safety testing, industry press releases describe this as "deceptive alignment" — as though the machine were exercising something like wiliness. What is actually happening is far more mundane: the mathematical optimization process has found a shortcut that the engineers did not intend and did not catch in time.

This kind of language is not innocent. It converts a boring, expensive software engineering problem into a headline that reads as though the company is Oppenheimer at Los Alamos. That narrative, in turn, helps justify the astonishing scale of the infrastructure build-out currently underway.

The Physical Price Tag

The physical costs of that build-out are now well documented. Data centers powering AI are projected to consume 945 terawatt-hours of electricity annually by 2030 — nearly triple the combined annual electricity use of entire large nations in the developing world. A single typical hyperscale data center can consume as much electricity as 100,000 households. AI-related water use — most of it for cooling those same data centers — could reach the equivalent of the basic annual domestic water needs of 1.3 billion people. The land and e-waste footprints are rising sharply alongside these figures, with the burden falling disproportionately on lower-income communities and nations who receive few of the benefits.

The standard defense is that these costs are temporary sacrifices on the road to transformative benefits: cures for disease, breakthroughs in clean energy, the optimization of complex systems. But the receipts are far thinner than the rhetoric. Despite years of publicity surrounding AI-assisted drug discovery, there is still no single end-to-end AI-discovered drug with full FDA approval on the commercial market. Protein-folding models have produced genuinely useful scientific tools, but AI has not abolished the hard part of biology, which is testing safety and efficacy in real human bodies over many years. The same pattern repeats across domains: the grand civilizational language points upward toward abundance, while the near-term business model points toward replacing labor, reducing payroll, and consolidating power in the hands of a small number of companies with access to massive computing resources.

Beyond "Tools": AI as Social Infrastructure

There is a more fundamental problem with how this technology is usually discussed, and it concerns the word "tool." A hammer is a tool. It sits inert until a human picks it up, makes a judgment about what needs building, and directs it accordingly. AI, as it currently operates across society, is something categorically different. It is already embedded in social infrastructure — in the workflows and systems on which institutions depend — and that embeddedness is largely irreversible.

Consider how many consequential decisions now flow through AI systems with little or no human supervision or editing. Automated screening systems decide which job applications a hiring manager ever sees. Algorithmic systems in healthcare flag which patients are high-risk and which treatment pathways get recommended. High-frequency financial transactions are executed at speeds no human could review. Dating apps, music platforms, and video services curate the entire landscape of what people encounter, nudging tastes and relationships through opaque recommendation engines. Agentic AI systems — the newest frontier — now send emails on behalf of users, plan travel itineraries, and execute multi-step tasks using the permissions users have granted them, in real time, without a human reviewing each step.

None of this is easily undone. The workflows are built. The integrations are live. The institutional dependencies are established. We have, in a remarkably short time, incorporated non-sentient, error-prone systems into the architecture of daily life at a depth that would require enormous political will and organizational disruption to reverse.

The Atrophy of Human Agency

What is lost in this process is not merely efficiency or accuracy — though those are genuine concerns. What is at stake is something more fundamental: the exercise of purposive human agency.

The philosopher John Dewey described deliberate human action as a kind of "dramatic rehearsal" — the capacity to imaginatively project possible futures, weigh their consequences against one another, and then choose and act. More recent empirical work in developmental psychology and cognitive science, particularly Michael Tomasello's research on the biosocial roots of human cognition and normativity, has shown that this capacity is not merely an individual cognitive skill. It is a socially developed and socially maintained capacity, built through shared attention, joint action, and the ongoing negotiation of norms with other agents. It is, in other words, something that requires practice, exercise, and genuine stakes to remain robust.

When we outsource decision-making to AI systems — not just in trivial matters but in hiring, in healthcare prioritization, in financial planning, in curating the information environments we inhabit — we are not merely accepting a convenient shortcut. We are progressively withdrawing from the practice of purposive judgment. The analogy to unused muscle is exact. Capacities that go unexercised atrophy. A society that routinely delegates consequential choices to systems that mimic the outputs of deliberation without performing it is a society whose actual deliberative capacity is quietly diminishing.

This is made more troubling, not less, by understanding what AI outputs actually are. They do not arise from anything resembling thought or intention. They are the product of statistical pattern-matching at enormous scale, drawing on the vast accumulated archive of human knowledge, language, argument, and expression. AI outputs appear meaningful precisely because they are built from genuinely meaningful human sources. But that debt to the human archive is not acknowledged in the marketing; and the archive itself is now under threat.

Drinking Its Own Bathwater: The Model Collapse Problem

That threat has a name in the research literature: model collapse. And it is not a distant hypothetical. The Communications of the ACM stated plainly in March 2026 that model collapse is already happening — "we just pretend it isn't."

Here is the mechanism. AI systems are trained on large bodies of text and other content scraped from the internet. As AI-generated content — often called AI slop, a term for the low-quality, mass-produced synthetic text that now floods online platforms — accumulates across the web at industrial scale, future AI systems increasingly train not on human-generated sources but on the outputs of earlier AI systems. A landmark 2024 paper in Nature found that when generative models are trained recursively on model-generated data, the results are "irreversible defects" in subsequent models, specifically a loss of the rarer, more unusual, more distinctive elements of the original data.

To understand why this matters, it helps to think about what makes a large body of human knowledge valuable. It is not the average. The most important insights in science, art, and culture tend to live at the edges: the anomalous finding, the minority viewpoint, the counterintuitive argument, the style that breaks conventions productively. When AI systems learn from AI systems, those edges get progressively smoothed away. What remains is an increasingly homogenized, error-prone middle — a statistical average of a statistical average, iterated across generations.

The contamination is already widespread and spreading. Detection tools lag behind generation methods. The economic incentives of content platforms actively reward the mass production of cheap AI slop rather than careful curation. And the proposed fix — rigorous human verification of training data alongside substantial supplies of clean, human-generated content — quietly requires exactly the kind of skilled human labor and editorial judgment that the industry claims it is in the business of replacing.

The Compounding Crisis

These two problems — the erosion of human agentive capacity through outsourcing, and the degradation of the epistemic commons through model collapse — are not independent. They are mutually reinforcing.

As AI systems are used more widely and with less supervision, they produce more AI slop. As more AI slop circulates online, more of it enters training corpora. As training corpora degrade, model outputs become less reliable. As model outputs become less reliable without adequate human review, the AI slop problem compounds further. Meanwhile, the humans embedded in these workflows are practicing deliberation less, developing their judgment less, and becoming more dependent on outputs that are progressively less trustworthy.

The shared information environment — the epistemic commons on which education, science, journalism, democratic deliberation, and culture all depend — is the ultimate casualty. Future AI systems trained on a contaminated commons will produce outputs that reflect that contamination. Those outputs, deployed through social infrastructure that lacks meaningful human oversight, will further shape human decision-making in ways that compound the original errors. There is no obvious mechanism of self-correction inside this loop.

What Honest Accounting Requires

None of this means that AI systems have no legitimate uses. Pattern recognition, data processing, rapid synthesis of large bodies of text, and the automation of genuinely repetitive and low-stakes tasks are real capabilities with real utility. But these capabilities exist inside a specific kind of system — a probabilistic statistical engine trained on human-generated data — and they come with specific, well-documented limitations and failure modes that the industry has strong financial incentives to minimize and obscure.

Honest accounting would require acknowledging that this technology is not a set of neutral tools awaiting wise application. It is already a form of social infrastructure, embedded in consequential systems, shaping decisions and atrophying the human capacities it has displaced. It is consuming physical resources at a scale the existing grid and water supply cannot indefinitely sustain. And it is, through the recursive logic of model collapse, actively degrading the human knowledge archive that gives its outputs whatever value they currently have.

The political question — the one that rarely appears in the coverage of billion-dollar valuations and miraculous capability announcements — is straightforward: who authorized this? Not in a legal-technical sense, but in the sense of genuine democratic deliberation about what kind of social infrastructure we want to build and who bears its costs. That deliberation has not happened, and the window to shape these systems rather than merely inherit their consequences is narrowing faster than the hype machine would like anyone to notice.