The Cost That Stayed Hidden

I. The question of the week

There is a particular move the technology conversation makes when a critique becomes too loud to ignore: it names the problem, quantifies it, and in the same breath converts it into a management challenge. The hidden costs of artificial intelligence — the water evaporated from cooling towers, the carbon vented by data centers, the rare earths torn from the ground, and the human beings paid pennies to label the images that teach a model to see — have travelled precisely this route over the past eighteen months. In late 2024 they were an exposé. By early 2025 they were a “sustainability challenge.” By the autumn they were a line item in a return-on-investment calculation, filed under “Responsible AI.”

This week’s question is what that journey did to the costs themselves. Naming a harm is supposed to be the first step toward reckoning with it. But naming can also be a way of neutralizing — of absorbing an objection into the vocabulary of the very institutions the objection was aimed at. The environmental toll of AI has been measured with increasing precision; the tonnes of carbon now have a figure attached. The human toll — the data labelers in Nairobi and Manila, the miners in the Congolese cobalt belt, the crew on the container ships — has proven far harder to metabolize, because there is no clean unit in which to express it and no obvious efficiency gain to promise.

The arc that follows traces two things that ran in parallel and rarely touched: what the discourse said about AI’s buried costs, quarter by quarter, and what those costs actually did. The gap between them is the story. It is not that no one was watching. It is that watching, by itself, changed remarkably little.

II. What we’ve been saying

The rhetoric arrived already split. In December 2024 two pieces published within days of each other set the two poles that would organize everything after. One, from an energy-industry trade outlet, opened with a number designed to shock — training a single AI model can generate pollution equal to ten thousand car trips between Los Angeles and New York — and then pivoted, within the same article, toward the language of drivable solutions and better corporate stewardship. The other, from the International Labour Organization, refused the pivot: its account of the invisible workforce fueling the “automated” economy insisted that AI’s success “hinges on an invisible workforce performing low-paid, precarious tasks under challenging conditions,” and offered no reassuring turn at the end. From the outset, then, the environmental cost was framed as a problem with an engineering answer, and the labor cost as a problem with no comfortable exit.

The first quarter of 2025 is where the optimistic register surged — the probe counts show eleven positively framed pieces against seven critical ones, the widest pro-tilt of the entire arc. The tilt is worth reading closely, because it was not denial. It was assimilation. A representative essay described the hidden cost of AI in energy, water, and sustainability terms, conceding the footprint in full before framing it as a challenge the sector was already rising to meet. Thomson Reuters ran the same structure under the banner of industrial AI and sustainability opportunities, where energy consumption appeared not as an indictment but as one line in a matrix of adoption barriers alongside ROI-projection difficulty. Harvard Business Review, addressing AI’s growing waste problem, embedded the crisis in a title that promised the remedy — circular-economy principles — before the reader had finished absorbing the crisis. The grammar was consistent: name the cost, then reach for the fix. The cost became the setup; the solution became the sentence.

Not every voice in that quarter took the bait. A report on the supply-chain cost of Big Tech’s AI investments held its gaze on data centers, material sourcing, and the mounting social costs of the pipeline without offering the tidy exit. And an essay written for human-rights defenders posed the AI dilemma for peacebuilders in terms that resisted the celebratory frame it described — the frame in which AI had been “hailed as the key to unlocking unprecedented” gains across every sector. But these were the minority. The dominant Q1 sound was of a critique being welcomed indoors and given a seat at the strategy table, where it could be managed.

By the second quarter the register tilted the other way — five critical pieces to two optimistic — and the reason is instructive. The sustainability frame had begun to strain against its own contradictions, and a sharper counter-voice emerged to name the strain. The most direct came from an academic writing for a general audience, arguing that the public needs to challenge the “good AI” myth pushed by tech companies — that the barrage of positive messaging about what AI can do was itself a product to be sold. Against this, the industry’s own vocabulary hardened into a formula: the ROI of Responsible AI, which promised a future where AI “not only transforms industries but does so with an unshakeable commitment to human values and the well-being of our planet,” and which read less like an argument than like a brochure. Meanwhile the university librarians — at Amherst and Tulane — were quietly assembling ethics guides that listed the social costs plainly, without the promise, for students who would have to reckon with them. The trade press, for its part, framed the whole thing as a wager: a $500-billion gamble on whether ethics and energy could keep up with the capital already committed.

By the third quarter the two registers had settled into an even split — eight and eight — and the vocabulary had matured into something like a stable genre. Harvard Magazine asked whether Green AI was hype or hope, the question mark doing the work of balance. A widely shared essay argued that regulating AI use could stop its runaway energy expansion, moving the conversation from corporate virtue toward public rule. And Bank of America’s institute weighed in on addressing the “AI” in sustainability, noting a “growing disconnect between AI’s potential to advance sustainability and its implementation” — an admission, from inside finance, that the promise and the practice had come apart. What is striking across all four quarters is what the rhetoric almost never did: it named the environmental cost obsessively and returned to the human cost only rarely, and when it did, without a remedy to offer.

III. What’s been happening

While the vocabulary matured, the measurements got worse. The single most durable finding of the period is that the environmental footprint kept expanding no matter which rhetorical register was ascendant. Research tied the AI boom to roughly eighty million tonnes of carbon emissions in 2025 alone — not a projection for some distant year but an accounting of the one just past. The United Nations University’s water institute compiled the fullest inventory, documenting the carbon, water, and land footprints of AI’s energy use as a single interlocking system, and by June 2026 the UN was reporting that AI’s environmental costs threaten water, land, and climate at a pace it called capable of outrunning the mitigation strategies being proposed to contain it. The numbers did not respond to the discourse. They followed the buildout.

The buildout has a physical body that the “cloud” metaphor is designed to hide, and the period’s reporting kept insisting on that body. An analysis of AI hardware’s environmental impact traced the lifecycle backward from the specialized chip to the rare-earth extraction that produced it and forward to the electronic waste it becomes — the part of the story that data-center efficiency figures, however honestly reported, systematically exclude. This is the terrain Kate Crawford mapped in The Atlas of AI (2021), which follows the pipeline “from the mining sector, where resources are extracted and transported to create the core infrastructure of AI systems, to the software side, where distributed workforces are paid pennies per microtask.” The costs the discourse itemized in 2025 were the same costs Crawford had already located at every stage of the chain; the reporting caught up to the map.

The chain includes the ocean, which almost no coverage in this arc mentioned. Crawford’s accounting of the shipping that moves the hardware notes that the cargo fleet is “estimated to emit as much pollution as fifty million cars, and sixty thousand deaths every year are attributed indirectly to cargo-ship-industry pollution,” and that even industry-friendly sources concede thousands of containers are lost annually, some carrying toxic substances that “leak into the oceans.” That this leg of the footprint went essentially unremarked while carbon-per-query became a familiar statistic tells you something about which costs the conversation was equipped to see. The measurable ones got measured. The rest stayed offshore, literally.

The human cost has its own longitudinal reality, and it did not improve. The ILO’s account of the invisible workforce was not describing a residual problem on its way to being solved; it was describing the load-bearing labor of the entire enterprise. The people who annotate training data, moderate the content, and clean the datasets — the petites mains of the industry, working from Nairobi, Manila, Lagos — remained, throughout the period covered here, low-paid, precarious, and structurally invisible, for the simple reason that their invisibility is the product. A system marketed as “automated” cannot advertise the humans it runs on. The MIT Press primer AI Ethics (2020) puts the arrangement in exactly these terms: “human labor is also hidden behind the scenes: miners, workers on ships, click workers who label data sets, all in the service of capital accumulation by very few people.” The concealment is not incidental. It is the business model rendered as an aesthetic.

What changed over the four quarters was not the conditions but the framing infrastructure built around them. Institutions began commissioning the surveys and reports that let the costs be spoken about at scale. An original global survey prepared for the 2025 Human Development Report found, when it asked whether societies were ready to meet the expectations of AI for development, that roughly a fifth of respondents were already using AI in ways that outran any framework for governing its costs — a finding that measured enthusiasm and preparedness moving in opposite directions. The apparatus of measurement grew impressively. The apparatus of remedy did not keep pace with it.

There is a deeper claim underneath the environmental accounting that the arc’s reporting gestured toward without quite stating, and AI Ethics names it directly: the danger that AI “becomes an alienation machine: an instrument to leave the Earth and deny our vulnerable, bodily, earthly, and dependent existential” condition. The clean-tech image of AI — weightless, placeless, running in a “cloud” — is not merely inaccurate marketing. It is a way of thinking that treats the Earth, and the bodies laboring on it, as a support system to be transcended rather than a set of limits to be honored. Every water-footprint report published in this period was, whether it said so or not, an argument against that transcendence fantasy. The water is real. The cooling towers are somewhere. The labelers have names.

IV. Where they meet, where they miss

They meet on the environmental ledger, and they meet there because the ledger could be built. Carbon has a unit. Water has a unit. Land has a unit. Once the UNU inventory and the eighty-million-tonne figure existed, the discourse had something to hold, and holding it produced the whole optimistic register of Q1 through Q3 — the Green AI question marks, the circular-economy remedies, the regulatory proposals. This is genuine progress, and it should be credited as such. A cost that can be measured can, in principle, be capped. The rhetoric and the reality converged on the environment because measurement gave them a common language.

They miss on the human ledger, and they miss for the mirror-image reason: there is no unit. The pennies-per-microtask economy that Crawford and the ILO both describe cannot be aggregated into a single number that fits on a sustainability dashboard, and so it slid to the margins of a conversation organized around dashboards. Notice the asymmetry in this very arc: nearly every quarter produced fresh environmental figures, while the labor story was carried almost entirely by two pieces — the ILO’s at the start and the supply-chain accounting that followed. The costs that resist quantification are the costs that get to stay hidden, and the industry has every incentive to keep them that way, because a data labeler in Nairobi is not a rounding error in a carbon model — she is a person whose wage is someone else’s margin.

This is where the “Responsible AI” vocabulary earns its skepticism. The ROI-of-responsibility framing that hardened in 2025 performs a specific trick: it promises that doing right by the planet and doing right by the balance sheet are the same act. For carbon, this can occasionally be true — energy efficiency saves money. For labor, it is almost never true, because the entire economic logic of the microtask pipeline is that the labor is cheap. The essay warning against the “good AI” myth was pointing at exactly this seam. When Bank of America’s institute concedes a growing disconnect between AI’s sustainability potential and its implementation, the disconnect it names is real, but it is described as a gap to be closed rather than as a structure that profits from staying open.

Our own earlier work on the social-aspects beat noticed the pattern in an adjacent form. As we argued in a critical analysis published on 4 May 2025, the discourse around AI and equity repeatedly emphasizes “the transformative potential” of the technology while the material conditions underneath it go under-examined — the same rhetorical center of gravity, the promise pulling attention away from the pipeline. The environmental turn is that tendency’s most sophisticated version: it does examine the material conditions, rigorously, but only the fraction of them that can be counted.

V. The longer view

Eighteen months of conversation taught the industry to see its own carbon and water, and that is not nothing — a cost named in tonnes is a cost that can, eventually, be legislated. But the same eighteen months taught it to keep not seeing the people, because the people do not reduce to tonnes, and a vocabulary built for the ledger has no column for a wage that is somebody’s whole day. The environmental cost was domesticated by measurement; the human cost was protected by its absence from the metrics. Both are hidden, in Crawford’s sense, at every stage of the pipeline — but only one of them was ever going to be found by a footprint calculator, because the other one is the footprint calculator’s blind spot, engineered rather than accidental. The clean machine was never clean, and it was never a machine alone; it was cooling towers in a drought and a woman labeling images for pennies while a headline called her labor automation. The measure of whether this conversation has actually traveled anywhere is not whether the next report has better numbers. It is whether the numbers ever learn to count the person who is not a number — and on that question, so far, the arc bends toward the ledger and away from her.

The Cost That Stayed Hidden

I. The question of the week

II. What we’ve been saying

III. What’s been happening

IV. Where they meet, where they miss

V. The longer view

References