AI and Its Environmental and Social Impact

Part I: AI Is Not "Virtual"

Most discussions of AI treat it as pure information — weightless, placeless, detached from the physical world. We speak of "the cloud" as if computing happened somewhere clean and frictionless, beyond geography and consequence. That framing is not just imprecise; it is ethically dangerous. It invites us to evaluate AI purely on what it produces while ignoring everything required to make that output possible.

In reality, artificial intelligence is a material and social system. It requires energy-intensive computation and cooling in massive data centers, water to keep them from overheating, and land and grid capacity competed for with other uses. It is built from minerals — cobalt, lithium, copper, rare earths — extracted through supply chains whose costs fall on communities far from where the AI is used. It relies on human labor at every stage: miners, data labelers, content moderators, technicians. And hardware replaced on short upgrade cycles becomes electronic waste, a burden that also falls unevenly.

A few concrete numbers make this vivid — though each has to be handled carefully, because environmental statistics about AI circulate much faster than they are checked. Training GPT-3, OpenAI's 2020 language model, is estimated to have consumed about 1,287 megawatt-hours of electricity — roughly the annual use of 120 American homes — and emitted around 550 tonnes of CO₂. The estimate comes from outside researchers , not OpenAI, and covers a single run of a model several generations old — it says nothing about the far larger runs since.

Microsoft reported consuming roughly 6.4 million cubic meters of water company-wide in 2022, a 34% increase over the previous year, most of it cooling its cloud data centers; in July 2022 alone its cluster outside West Des Moines, Iowa drew about 11.5 million gallons — around 6% of that water district's total for the month. And the Democratic Republic of Congo, one of the world's poorest countries, supplies roughly three-quarters of the world's mined cobalt — the metal in the lithium-ion batteries behind both our devices and data centers' backup power — often from mines with severe safety and environmental problems. Chatbot users are connected to these realities by supply chains designed not to be visible.

Why Collective Harms Are Hard to See

AI's social and environmental impacts are distributed, delayed, and offloaded. A single chatbot query feels trivial to the person who types it; multiplied across billions of daily interactions, it becomes a significant infrastructure load. The benefits are immediate and visible — a faster answer, a completed task. The costs are indirect: a marginal increase in electricity demand, a fractional draw on cooling water, a small contribution to hardware wear.

This structure — immediate, visible benefits; delayed, dispersed, often invisible costs — recurs across environmental and social harms, and it resists individual decision-making, because no individual's choice looks significant in isolation. Hence the need for institutional ethics — collective standards enforced through regulation, procurement policy, reporting requirements, and democratic oversight. Consumers choosing "greener" AI products is not a substitute for systemic accountability.

In plain English: AI's costs are not merely large but misdirected. Economists call this an externality — a cost of production borne by someone outside the transaction. If a data center's neighbors absorb the noise, the water draw, and the higher electricity prices, they pay part of the price of a product they may never use. The third premise does the real work, and it is a claim about incentives, not bad faith: even a scrupulous company has no market signal telling it how much harm it causes in a Congolese mining town, because that harm never appears on its balance sheet. Externalities are, by definition, problems individual choices cannot solve, because the chooser is not the payer.

Part II: Environment and Climate Change

AI's relationship to the environment is genuinely two-sided, and that complexity is itself ethically important. AI systems contribute to environmental problems through their energy and material demands; they may also help address them through better prediction and resource management. Understanding both sides is more useful than either dismissing AI's environmental costs or accepting "AI for climate" uncritically.

AI's Environmental Footprint

AI's environmental impact has several dimensions that no single metric like carbon emissions captures, and each raises distinct ethical and policy questions.

The International Energy Agency (IEA) is the standard reference here. Its April 2025 report Energy and AI put global data-center electricity use at roughly 415 TWh in 2024 — about 1.5% of world consumption. Ten months later, Electricity 2026 raised the 2025 figure to about 485 TWh: data-center demand grew roughly 17% in a year, against about 3% growth in global electricity demand overall. The IEA projects consumption will roughly double to about 950 TWh by 2030, around 3% of global electricity, with the AI-specific share growing fastest — up roughly 50% in 2025 alone, projected to triple by 2030. In the United States, data centers account for close to half of projected growth in electricity demand through 2030. For emissions, the IEA earlier attributed around 330 Mt CO₂e in 2020 to data centers and transmission networks together — about 0.9% of energy-related greenhouse-gas emissions, a baseline predating most of the build-out.

Two features teach something about reading a forecast. First, between the two IEA reports the headline 2030 projection barely moved — from roughly 945 TWh to roughly 950 TWh — while the near-term baseline rose substantially. The destination stayed put; the starting point caught up to it. A reader who saw only the 2026 headline might think the outlook had worsened dramatically, when the agency's view of 2030 was almost unchanged. Second, the 2026 report emphasizes bottlenecks: more than 2,500 GW of projects sit stalled in grid connection queues worldwide, and the IEA treats supply-chain and grid constraints as a real brake on the more aggressive scenarios. A projection is not a prediction. It says what would happen if a trend met no obstacles, and here the obstacles are visible.

These numbers are neither trivial nor catastrophic in isolation. What makes them politically significant is context: they compound pressures on already-stressed energy and water systems that are simultaneously being asked to support a renewable transition.

The Rebound Problem

A common response to concerns about AI's energy footprint is that models keep getting more efficient — each query costs less energy than a year ago. True, but it does not follow that total energy use is declining. The rebound effect, also called the Jevons paradox, is a well-documented pattern in which efficiency improvements raise total resource use because cheaper systems get deployed more widely.

Applied to AI: as models get cheaper per query, firms integrate them into more products at larger scales, inference volume rises dramatically, and aggregate electricity and hardware demand still grows even as unit costs fall. This is why "smaller and cheaper models" cannot be straightforwardly celebrated as an environmental solution. History offers analogies. Fuel-efficient cars after the 1970s oil crisis did not cut gasoline consumption proportionately — Americans drove more miles because driving was cheaper. LED lighting increased total light consumption, because lights went into more places and stayed on longer. Efficient jet engines were followed by more flights and greater fuel burn. Each time the efficiency gain was real; the savings were offset by expanded use.

The argument is easy to overstate. It does not say efficiency is bad, only that efficiency and reduction are different things and the first does not automatically deliver the second. Premise 3 contains a real empirical condition — "if total use expands faster than efficiency improves" — and it can fail. Refrigerator standards did cut household refrigeration energy absolutely, because there is a ceiling on how many refrigerators a household wants. The worry about AI is that no such ceiling is in view once each task costs a fraction of a cent. The upshot is a test for any sustainability claim: if emissions per query fell 40%, ask what happened to the number of queries.

Training and inference also differ. Early public discussion focused on the spectacular energy cost of training frontier models, but evidence suggests inference — running the model to answer actual queries — is the larger long-run burden, because it scales with adoption. AI's environmental story is not about a few dramatic training events; it is about what happens when AI becomes ordinary infrastructure embedded in search, office software, medicine, and government services.

AI for Climate: A Comparative Judgment

AI is also deployed for climate applications, and this is not merely marketing. There are genuine use cases where it can help reduce emissions or adapt to a changing climate.

This is why the chapter refuses both available slogans. "AI is an environmental disaster" and "AI will save the climate" are claims about a category containing wildfire detection systems, protein-folding models, warehouse routing software, and advertising-copy generators — things with nothing in common but the technique used to build them. Premise 4 is a familiar principle of practical reasoning, one you would apply without hesitation to a medical procedure's risks, and it forces the comparison to be made case by case.

Two cautions follow. The standard is demanding: someone has to measure both the costs and the averted harms, and firms disclose far more about the second than the first. And it is silent about distribution — a system could pass the comparative test in aggregate while the harms fall on one population and the benefits on another, exactly what the Externalized Cost Argument warned about. Comparative justification is necessary, not sufficient.

Part III: Medicine and Research

Medicine is often presented as AI's strongest case — the domain where the benefits are most tangible and hardest to dismiss as convenience or commercial gain. A model that helps a radiologist catch a tumor she would have missed saves lives in a direct and comprehensible way. That makes medical AI worth taking seriously on its merits, and worth examining critically, because the stakes of getting it wrong are correspondingly high.

Scale and Concentration of Medical AI

AI deployment in medicine has grown dramatically. Stanford's AI Index reports that the U.S. Food and Drug Administration authorized 223 AI-enabled medical devices in 2023 alone, up from just six in 2015. The agency's public list passed 1,500 entries in early 2026 and continues to rise. It is revised several times a year, which is why this chapter gives no exact figure: any precise count is stale before the semester ends.

Two features of that list matter more than its length. The first is concentration. Roughly three-quarters of the authorized devices are radiology devices, because medical images are abundant, standardized, and well suited to the classification and anomaly-detection tasks current systems excel at. Cardiovascular and neurology follow at a distance; most of medicine barely appears. Adoption is real, and narrow. A student who reads "1,500 AI devices approved" and pictures AI spread evenly across medicine has misread it.

The second feature is more instructive: the list is not what it appears to be. The FDA itself cautions that it is "not a comprehensive resource" — it is assembled by keyword-searching the public summaries accompanying device authorizations, so a device whose paperwork used different vocabulary may simply be absent, and inclusion depends on what a manufacturer wrote rather than on what the software does. What looks like a census is really a search result. That does not make it useless — it is the best public inventory available, and its broad trends are almost certainly right — but the number carries an error bar it never displays. Asking how a statistic was constructed, not merely how large it is, is among this chapter's most transferable skills; we will need it again in Part V.

A Scientific Success: AlphaFold

The clearest example of AI delivering transformative scientific value is AlphaFold, DeepMind's protein-structure prediction system. Predicting how a protein folds from its amino-acid sequence — the "protein folding problem" — was among the hardest unsolved problems in structural biology, because a protein's three-dimensional structure determines its function and the search space is astronomically large.

AlphaFold broke through at the 2020 CASP competition, and the public database built from it now holds roughly 200 million predicted structures, freely available. In 2024 the work received the Nobel Prize in Chemistry: half to Demis Hassabis and John Jumper for structure prediction, half to David Baker for computational protein design. It is a proof of concept — AI can genuinely accelerate research in ways that create broad social value, especially when paired with open access and public goals.

Scientific promise, however, does not erase concerns about access, intellectual property, and data control. When AI accelerates drug discovery, who can afford the resulting drugs — and who owns the patents — remains as urgent as ever. Faster discovery is not better distribution.

Risks of Medical AI

Medical AI's risks are not limited to inaccuracy. The most serious arise when systems perform well on average but unevenly across populations, or when their place in clinical workflows changes how clinicians think and decide.

The World Health Organization has argued that AI in health should be governed by human rights, public accountability, and the needs of health workers and patients — not only by technical performance metrics. The framing refuses to let "accuracy on a benchmark" suffice for deployment. Consider a system with 95% accuracy overall but 82% for patients with darker skin. Deploying it without acknowledging that disparity is an ethical failure, not merely a technical limitation.

Part IV: Good Governance

Democratic governance has always faced a tension between efficiency and legitimacy. Efficient government decides quickly, cheaply, with uniform rule application. Legitimate government makes decisions citizens can understand, contest, and hold accountable — decisions that respect rights and remain correctable. AI intensifies the tension by promising efficiency at precisely the level — large-scale, consequential, population-affecting decisions — where legitimacy requirements are most demanding.

What Good AI Governance Requires

The OECD's AI Principles — first adopted in 2019, updated in May 2024, and now adhered to by 47 parties, including every OECD member state and the European Union — identify inclusive growth, human rights and democratic values, transparency, robustness, and accountability as core requirements for trustworthy AI. These are not technical standards but normative commitments about what relationship public institutions should have with the citizens they govern.

The OECD notes that governments are cautious adopters of AI relative to the private sector — precisely because public decisions affect rights, not merely convenience, and accountability requirements are stricter. The caution is appropriate: the legitimacy logic of democratic governance imposes constraints that the efficiency logic of markets does not.

The Governance Triangle

Public-sector AI almost always claims to improve one of three goods: efficiency (faster, cheaper decisions at scale), fairness (equal treatment, non-discrimination), or accountability (explanation, appeal, assigned responsibility). The political question is what gets sacrificed when one corner is maximized without attention to the others.

The general pattern: governments introduce automated screening, ranking, or fraud detection when they face scarcity, political pressure, or public distrust, and the burdens fall most heavily on people already dependent on public institutions. Automation appears to solve a technical problem (efficient allocation) while the political problem (inadequate resources, structural disadvantage) remains intact. AI becomes a technology for making exclusion look objective.

The pattern is visible across countries. In the United Kingdom, the Home Office ran a "visa streaming" tool from 2015 that sorted applications into red, amber, and green risk tiers partly on the basis of nationality, with red-tier applications drawing far more scrutiny and far more refusals. In August 2020 the department suspended it rather than defend it in a judicial review brought by the Joint Council for the Welfare of Immigrants and the campaign group Foxglove, who argued it was discriminatory under the Equality Act 2010 and irrational at common law. In the United States, Allegheny County's child-welfare models drew so heavily on data generated by poverty itself — welfare applications, housing assistance, public mental-health visits — that families could be flagged partly for being poor. The algorithm did not create the inequality; it automated and laundered it.

Be precise about what automation is and is not responsible for. Britain's better-known Windrush scandal, in which long-settled British residents of Caribbean origin were wrongly classified as illegal immigrants and some detained or deported, was not an algorithmic failure. It grew out of the "hostile environment" policy, which shifted the burden of proving lawful status onto individuals, combined with the Home Office's own destruction of the landing-card records that would have supplied that proof. The lesson cuts both ways: automated systems can encode injustice, and ordinary bureaucratic policy can do the same without any software. Blaming the algorithm is sometimes accurate and sometimes a way of avoiding the political question.

Part V: Poverty and Global Inequality

The most consequential questions about AI's social impact may be global rather than national. The World Bank's 2024 Poverty, Prosperity and Planet report estimated that about 700 million people — 8.5% of the world — were living in extreme poverty on less than $2.15 per day, and about 3.5 billion, 44%, on less than $6.85. Sub-Saharan Africa held 67% of the world's extreme poor in 2024, with 16% of global population.

Those figures deserve exactly the scrutiny we gave the FDA's device list. In June 2025 the World Bank moved its international poverty line from $2.15 in 2017 prices to $3.00 in 2021 prices, folding in new household surveys from more than 160 countries. The same underlying world was suddenly recorded as holding roughly 125 million more extremely poor people; the Bank's estimate for 2025 is about 808 million, close to one person in ten . Nobody became poorer the day the line moved; the count changed because the definition did. Not a reason to dismiss poverty statistics — the revision made them more accurate — but a reason to ask of any headline figure who drew the line, in what currency, and for what year.

Against this background, claims that AI will transform human welfare need careful evaluation: not just whether particular applications provide value, but whether AI development as a whole reproduces or disrupts the structural inequalities that generate poverty.

The Case for AI in Development

Advocates for "AI for development" argue that the technology can genuinely reduce poverty by extending scarce expertise to underserved populations. A translation tool can help a non-English speaker navigate a benefits office. A mobile agricultural assistant giving localized weather alerts and crop advice may reach farmers who lack extension services. Low-cost diagnostics could identify disease earlier where physicians are scarce. These claims are not empty.

The Structural Critique

But the optimistic story depends on power relations that AI, in its current form, tends to reproduce rather than disrupt. The structural questions are stark: Who owns the models, the cloud infrastructure, and the chips? Whose languages appear in training data, and whose are absent or distorted? Which countries bear the environmental costs of extraction and disposal? Who captures the profits when AI is deployed in lower-income countries?

Development is not just higher output. It is an expansion of what people are actually able to do and be.

A paraphrase of Amartya Sen's central claim in Development as Freedom (1999)

That sentence is a compression rather than a quotation, but it captures what makes Sen's framework useful. Output is a proxy; capability is what the proxy stands for. Separate the two and you can ask a question no productivity statistic answers — whether the people an AI system touches end up with more genuine options than before.

Algorithmic Bias and Racial Justice

Within wealthy countries, AI's impact on inequality has a racial dimension that needs its own analysis. Two scholars have been central to that critique.

Extractive AI and the Question of Dependency

The pattern across environmental, labor, and development concerns is sometimes called extractive AI or AI colonialism: data, labor, energy, and materials drawn from communities with less power, while benefits and control centralize in a few firms and wealthy countries.

Key Points

This chapter has argued that AI must be evaluated not only by what it does but by how it redistributes benefits, risks, resources, and decision-making power. Its impacts on the environment, medicine, governance, and global inequality are not secondary concerns to address after the technology is built; they are constitutive of what it is for. Sustainability, human benefit, legitimacy, and justice are four tests that must be applied together.

References