-I- Background and Facts
Overview
Artificial Intelligence (AI) has moved from the pages
of science fiction into the fabric of everyday life. No longer just a
futuristic idea, it is now a powerful force reshaping how we work, communicate,
shop, and govern. Much like the steam engine, computers, or the internet, AI
represents a technological revolution—driving unprecedented efficiency and
innovation while also challenging long-standing economic systems and social
agreements.
Put
simply, AI is changing the world around us in visible and invisible ways. From
workplace automation to online conversations and from personalized shopping to
government decision-making, it is influencing nearly every aspect of life. The
rise of AI represents a fundamental paradigm shift from previous waves of
automation. While earlier technologies primarily disrupted "routine"
tasks, AI is uniquely capable of automating "nonroutine" tasks,
exposing a far wider portion of the workforce to potential disruption than
anything that came before.
Currently, AI can be defined as a machine-based system
that can perceive its environment and generate predictions, recommendations, or
decisions to achieve specific objectives. It underscores the remarkable
capabilities of modern AI systems such as GPT, which can now perform tasks once
considered uniquely human. The OECD has defined AI as a machine-based system
that can perceive its environment and make predictions, recommendations, or
decisions to achieve given objectives
AI is not a brand-new concept—it has existed since the
1960s, passing through several stages: symbolic AI in the mid-20th century,
machine learning in the 1990s, and deep learning in the 2010s. Early Artificial
Neural Networks (ANNs) were designed for narrow tasks with restricted learning
capacity, performing well in specific areas like image recognition but
struggling to grasp the complexity of human language. Recurrent Neural Networks
(RNNs), though capable of sequential processing, were slow and ineffective at
managing long-range relationships in text. The current surge in AI, often
described as the "AI Spring," stems from a single transformative
innovation introduced in 2017—the Transformer architecture, developed by
Google. At its core is the "attention mechanism," which enables a
model to assign importance to different parts of a sentence simultaneously,
much like how a person focuses on key words when reading. This parallel
processing made it highly efficient to train models on massive datasets, representing
a qualitative leap and the beginning of a new technological era.
This
breakthrough gave rise to Large Language Models (LLMs) such as OpenAI's GPT
series, Google's PaLM and Gemini, and Meta's LLaMA. Unlike earlier narrow-task
models, these "foundation models" are trained on vast datasets and
learn general patterns of language, code, and knowledge, making them adaptable
to a wide variety of applications—from writing and programming to customer
support and art generation, supported by powerful diffusion models in computer
vision. These advances have collectively propelled progress across natural
language processing, computer vision, generative AI, and complex
decision-making systems.
In
essence, AI has evolved from being a specialised automation tool into a
general-purpose technology with applications across nearly every domain.
However, as AI has evolved into its present form, it
has also sparked several debates. One concerns its usefulness — how beneficial
it truly is, whether it will help or harm society, take away jobs, endanger
humanity in the long run, or genuinely possess the “creativity” to rival the
human mind. The other debate revolves around its economic impact — how it
creates surplus, who controls that surplus and the means of production, and
whether it will transform existing production relations. To fully grasp AI’s
impact, we must look not only at the technology itself but also at its effects
on workers, the economy, and the political choices that shape its growth. Let
us briefly examine these aspects.
AI
presents major opportunities, including substantial productivity gains, the
creation of new goods and services, and more effective policymaking—for
instance, improved bail assessments or more accurately targeted welfare
distribution. At the same time, it brings serious challenges such as risks to
privacy, potential anti-competitive practices, worker displacement, and the
deepening of social issues like discrimination through biased algorithms or the
rapid spread of misinformation. AI has created both automation (replacement of the
workforce) and augmentation (assisting the worker with enhanced skills).
First, let us consider
the general benefits and perils of AI.
Advantages of AI
Every major
new technology brings advantages, and AI is no exception. Its downsides tend to
stem from how it's implemented rather than from the technology itself — a point
worth returning to when discussing how these benefits play out under the
current economic system.
1. Automation
and Efficiency: AI takes over repetitive tasks — data entry, report generation,
logistics, manufacturing — freeing humans for creative and strategic work.
Running continuously without fatigue, it keeps output consistent around the
clock. Its speed in processing massive data volumes also enables faster,
better-informed decisions than human teams could match.
2. Data
Analysis and Pattern Recognition: AI excels at extracting insights from enormous,
complex datasets. By spotting subtle patterns, it improves prediction accuracy
in finance, marketing, and healthcare, while reducing errors caused by human
fatigue — making it essential for data-driven organizations.
3. Risk
Mitigation and Safety: AI operates in environments too dangerous for humans
— bomb disposal, deep-sea exploration, mining, high-radiation zones — cutting
risk to human life. In healthcare, it supports earlier, more accurate
diagnosis, acting as a buffer between people and harm.
4.
Personalization and Digital Assistance: AI tailors experiences to individual needs — curated
streaming content, personalized learning, targeted healthcare advice,
responsive customer support. Virtual assistants like Siri and Alexa extend this
with round-the-clock help for millions of users at once.
5. Economic
and Long-Term Operational Benefits: Despite high upfront costs, AI delivers strong
long-term returns by streamlining operations and cutting costly errors. It also
drives innovation, accelerating research in medicine, materials science, and
engineering by exploring vast solution spaces far faster than traditional
methods.
6. Objectivity
and Unbiased Decision-Making: Trained on well-curated data, AI can base decisions
on logic rather than prejudice — valuable in recruitment, credit assessment,
and legal analysis. This objectivity depends entirely on data quality, but the
potential for fairer, more consistent decisions remains compelling.
7. Support for
Human Creativity and Innovation: Rather than replacing creativity, AI collaborates
with it — suggesting novel patterns in art, music, and design that humans might
not consider. In research and software engineering, it compresses years of work
into months, with its biggest contributions still ahead.
Now for some
commonly perceived disadvantages.
1. Job
Displacement and Economic Inequality: Automation threatens workers in manufacturing, data
entry, and customer service, especially where retraining and safety nets are
weak. Benefits concentrate among firms and nations with capital and
infrastructure, widening inequality. Algorithmic management also drives
deskilling and job precarity, leaving workers with less power over opaque
systems.
2. Bias,
Discrimination, and the "Black Box" Problem: AI is only as
fair as its training data, and biased historical data leads to discriminatory
outcomes in hiring, credit scoring, facial recognition, and sentencing. Complex
models are often hard to interpret, even for developers — this opacity
undermines accountability and makes errors difficult to catch or contest.
3. Privacy,
Surveillance, and Security Risks: AI's demand for data fuels mass surveillance, letting
governments and corporations track individuals often without real consent. It
also enables new security threats — cyberattacks, deepfakes, disinformation —
and remains vulnerable to adversarial manipulation designed to trigger
dangerous outputs.
4. High Costs
and Implementation Challenges: Deploying AI at scale requires major investment in
hardware, talent, and infrastructure, often out of reach for smaller
organizations. Ongoing maintenance and retraining add further cost, while
over-reliance risks leaving organizations exposed if systems fail — especially
once the human expertise they replaced has atrophied.
5. Absence of
Human Judgment, Empathy, and Common Sense: AI lacks genuine empathy, moral reasoning, and
commonsense understanding — qualities essential to caregiving, therapy, and
crisis management. It mimics creativity by recombining existing patterns rather
than truly originating ideas, making it a poor substitute where human
connection or ethical judgment is required.
6. Skill
Erosion and Intellectual Dependence: As AI takes over summarizing, drafting, and
diagnosing, people risk outsourcing core thinking skills — eroding critical
thinking and expertise, especially among younger generations. Organizations may
also eliminate roles prematurely, only to later miss the institutional
knowledge those people provided.
7.
Misinformation, Hallucination, and Epistemic Risk: AI can
generate fluent, confident, but false content — "hallucination" —
leading to fabricated legal citations, wrong medical claims, or false history
stated with authority. At scale, it lowers the cost of producing convincing
deepfakes and propaganda, threatening the shared truth that democratic
discourse relies on.
8. Existential
and Ethical Risks of Autonomous Systems: As AI gains autonomy in finance, military use, and
infrastructure, serious questions arise over accountability — who's responsible
when a self-driving car kills someone, or a weapons system misfires? Without
adequate oversight, delegating life-and-death decisions to machines challenges
justice itself, with the longer-term risk of misaligned AI goals now taken
seriously by leading researchers.
AI is no longer knocking at the door—it’s already
inside, woven into nearly every sphere of our lives. Yet access to it is
uneven: those who know how to wield it—and can afford to—are suddenly capable
of things that were unthinkable just a few years ago. That alone should give us
pause. Will it widen the gaps in social structure?
Because alongside this surge of capability comes a
growing unease. Is AI beginning to edge out human labour? Are jobs quietly
disappearing behind the curtain of “automation”? Should its expansion be
curbed—or is that even possible anymore? What about privacy, when these systems
can surveil, predict, and even manipulate? And perhaps most fundamentally: who
really owns the value AI generates—the creators, the corporations, or the
countless invisible workers whose data fuels it? And then there’s the deeper,
more unsettling question: are we building tools to serve humanity, or systems
that might one day render it obsolete? These are not abstract concerns. They
demand a serious reckoning. Let’s confront them head-on. These
aren't hypothetical debates for philosophers and academics. They are urgent,
real, and demand answers — now.
Let's start with the question that's keeping workers
up at night: What is AI doing to employment — and is it creating anything worth
having in return?
One camp argues AI isn't just disrupting jobs — it's
devouring them. And their solution? Restrain it. Restrict it. Or abandon it
altogether. But before we rush to that conclusion, let's ask a harder question:
Is AI really the disease, or merely the latest symptom of a much older
sickness?
Because this crisis didn't begin with algorithms, it
was born the moment economies were built, designed not around human need, but
around the relentless generation of surplus. Capitalism, by its very nature,
doesn't ask "what does society need?" — it asks "what generates
more surplus?" AI is its newest, most efficient instrument.
The Luddites[i]
didn't smash machines out of ignorance — they smashed them out of desperation,
watching their dignity and livelihoods ground to dust by the gears of so-called
"progress." The Sabotage[ii]
movement didn't arise from laziness — it arose from rage, from workers who had
no other weapon left. They were dismissed as obstacles to advancement. Sound
familiar? The real question was never "should we fear the machine?"
The real question — then and now — is: Who does the machine serve?
At present, AI is rapidly displacing roles such as
junior programmers, content writers, translators, graphic designers, and
accountants—but the real issue is the extent to which these roles can truly be
replaced. As one content writer put it, their job has increasingly become
reviewing and reworking machine-generated text. That is hardly the role people
envisioned for AI. Ideally, the relationship should be reversed: humans create,
and AI assists—by checking grammar, ensuring consistency, and handling repetitive
support tasks. Instead of taking over creative work, AI ought to relieve us of
mundane chores (laundry, dishes, etc.)—whether in professional workflows or
everyday life.
We will return to this later; for now, let us examine
some concrete facts.
How AI is
Affecting Jobs: Losses, Transformations, Creations
A. Scale of
Change
Around one in four jobs faces substantial
transformation from generative AI — through task substitution or partial
automation — though outright elimination is less likely than restructuring (ILO
Working Paper 140). McKinsey projects ~11.8 million U.S. workers in shrinking
occupations will need to transition to new roles by 2030.[iii]
Even workers who keep their jobs face growing pressure to reskill. Those who
fail to adapt risk significant income losses, driven mainly by prolonged
unemployment — older and mid-educated workers are especially vulnerable
(Netherlands evidence).
B. Jobs Most
Affected
- High exposure (replacement/heavy transformation): data entry, bookkeeping,
transcription, first-line customer service, routine legal review, basic
coding, templated reporting — areas where generative AI automates rapidly.[iv]
- Medium exposure (augmentation): accountants, paralegals, marketers, software
developers, analysts, radiology screening — AI speeds up and assists
rather than fully replaces.[v]
- Lower exposure (resilient): roles requiring social intelligence, complex manual
dexterity, creative leadership, empathy-driven care, and skilled trades.[vi]
C. Jobs Created
- Direct: AI researchers, ML engineers, prompt engineers, data annotators,
MLOps/cloud roles, AI safety and policy specialists.
- Indirect: AI trainers, integration specialists, data-centre operations,
domain experts who leverage AI, AI consultancy and auditing roles — all
reporting strong hiring demand (PwC, McKinsey, NASSCOM).[vii]
II-
Economic Implications
Intensifying Exploitation through Surplus Labour: Generation of Surplus: Capitalist growth
depends on cutting "socially necessary labour time" — the time needed
to produce a commodity. As this time shrinks, "relative surplus labour
time" grows — the portion of a worker's day spent generating profit beyond
their own subsistence. AI functions as a productivity multiplier: a task once
requiring ten workers may now require one worker directing an AI system, with
output the same or higher but labour cost sharply reduced. This compression of
necessary labour time does not vanish — it resurfaces elsewhere. Historically,
and unless deliberately redirected, it flows into the profit margins of firms
owning the technology and the capital markets financing it, rather than into
workers' wages. Productivity gains accrue mainly to owners of capital.
Displaced workers face unemployment or precarious gig work, swelling the
reserve army of the unemployed, while those who remain become appendages or
overseers of machines, as capital extracts more value from their time.
Data as a Second Source of Surplus: A second, less visible source of surplus lies in data
itself. Foundation models are trained on vast quantities of text, images, and
code produced by millions of people — writers, programmers, artists, ordinary
internet users — almost none of whom are compensated for their contribution.
This unpaid, diffuse labour becomes raw material that a small number of firms
convert into proprietary assets. Some argue this constitutes a new form of
value extraction: surplus is not generated by the AI model at the moment of
use, but by the accumulated, uncompensated human labour and expression embedded
in its training data — value that is then owned and monetized by whoever built
the model.
Who Controls the Means
of Production_ Concentration of
AI's Means of Production: AI appears
open to anyone, but it is not. Training a state-of-the-art foundation model
demands massive computing infrastructure, specialized hardware, and capital
investment. Only a small number of corporations and governments can afford this
scale. The result is structural concentration. Compute clusters, proprietary
datasets, and trained model weights — the "means of production" for
AI — sit in very few hands. Yet the technology's effects ripple across the
entire economy. This mirrors a familiar pattern in the history of
general-purpose technologies. But AI's concentration is unusually extreme,
driven by enormous capital requirements and highly specialized expertise. This
concentration shapes the data itself. Those who control training data also
control its biases. They decide what gets generated, how it's shaped, and whose
interests it serves.
Does AI
Generate Value, or Merely Transfer It?
Does AI create value on its own, or simply amplify human labour's productivity?
One view treats AI as a sophisticated machine, not a new kind of worker. AI
lacks the agency or stake in outcomes that living labour has. Its apparent
"value" is really the congealed labour of researchers, engineers, and
data contributors — transferred gradually into the goods and services it helps
produce. AI does not create value in a new sense. It reshuffles who is needed
to produce existing value, and at what price.
This aligns
with Marx's theory of value, where value originates from socially necessary
human labour. AI, like any machinery, increases productivity and transforms the
labour process, but does not create new value independently. It transfers value
embodied in its own production costs — the labour of design, training,
infrastructure, and maintenance — into the commodities it helps produce. In
this sense, AI functions as constant capital, not a source of surplus value.
That
production cost is itself human labour: software engineers, data annotators,
chip manufacturers, maintenance workers, and the vast workforce sustaining data
centres. AI also depends on accumulated knowledge and data generated through
social labour. The wealth attributed to AI comes not from AI as an autonomous
agent, but from human labour embodied in its production and continuously
required for its operation. What looks like AI generating value independently
is what Marx called the fetishism of capital — the productive power of human
labour appearing as a property of the machine itself.
Productivity
gains from AI do not automatically flow to the workers whose tasks are
compressed or eliminated. That depends on bargaining power, ownership
structure, and policy choices — not on the technology's capabilities alone. It
is not AI itself that generates wealth. It is the labour hidden inside it.
These
structural dynamics translate into concrete outcomes for workers.
Rather than
liberating workers, AI is being used by capital as a tool of control: from
wearable trackers like Amazon's wristband that monitor and correct warehouse
workers' movements in real time, to the broader threat of automation creating a
"reserve army of labour" — with up to 92 million jobs potentially
displaced by 2030 and over 184,000 tech jobs already cut in 2025 alone — which
disciplines the remaining workforce into accepting lower pay and worse
conditions out of fear of replacement. At the same time, AI deskills work by
breaking complex tasks into simple ones that machines or untrained workers can
perform, stripping skilled labour of its former value and turning both skilled
and unskilled workers into interchangeable, disposable parts of the production
process.
The
political economy of AI points toward a deepening of capitalist contradictions.
Immense wealth is generated, but concentrated in the hands of those who own
AI's means of production. Labour's bargaining power is, in turn, weakened. This
drives greater inequality — and risks a crisis of underconsumption and
overproduction.
III Environmental and other
Issues
1. Dematerialisation is a myth
There has been sustained propaganda around a myth: that AI
will dematerialise official work altogether — 'the cloud', weightless data,
frictionless intelligence. Paperwork will shrink to near-zero, computing time
will fall, everything will become more efficient, and energy consumption will
decrease as a result. This is simply false. First, AI carries a built-in
tendency to consume ever more energy just to exist and grow. Second, capitalist
production, trapped by Jevons' Paradox[viii], will never allow such
savings to materialise. Let's look at this briefly. A little technical
grounding is needed to see it clearly, so we'll keep it simple.
In reality, every AI query is an industrial process: it
consumes electricity, water and rare minerals, and releases heat and waste. The
underlying deep-learning operation — multiplying inputs by a weight matrix, y =
f(Wx + b) — looks mathematically trivial. It isn't. The function is nonlinear,
the variables are not single numbers but arrays and matrices, and this
operation must be repeated across billions of parameters and trillions of data
points. It also has to run in reverse. Take a stored image of a pen as an
example: AI runs millions of forward passes to learn it, then makes a
prediction. If that prediction is only 70% accurate, the model must run
backwards to trace the error — more data can be fed in too — and this forward-and-backwards
cycle repeats until recognition is near-perfect. This back-and-forth is
precisely what forces the physical scale: massive parallel computation demands
massive, parallel extraction of energy and matter.
This is best understood through Ilya Prigogine's concept of
'dissipative structures' — systems (a flame, a hurricane, a corporation) that
sustain their internal order only by importing low-entropy (more ordered)
resources such as energy, freshwater and minerals, while exporting high-entropy
(disordered) waste as heat, carbon and degraded matter. AI is one such
structure. Worse still, because model performance scales as a power function of
parameters and compute ('super linear scaling'), each further increment of
'intelligence' costs more resources than the intelligence gained. There is no dematerialised
endpoint in sight — only a deepening metabolic rift between computing and the
biosphere it feeds on.
Every system that keeps itself organised — a flame, a
hurricane, a living cell, or a data centre — has to keep drawing in energy and
matter from outside and pushing waste back out; nothing stays
"ordered" on its own. This is basic thermodynamics: the second law
says any process that does useful work also degrades some energy into unusable
form, usually heat, and that degradation — called entropy — cannot be undone.
Once energy is dissipated as heat or matter is broken down into waste, it can't
be gathered back into its original, useful state; the disorder it created is
permanent. So a system can look stable and orderly on the inside only because
it is constantly exporting disorder into its surroundings — the price of that
internal order is a growing, irreversible mess outside it. But physics alone
doesn't explain why we're generating so much of it: that comes down to a
handful of tech monopolies locked in a competitive arms race, not any physical
necessity. So while thermodynamics measures the damage, it's capitalism's drive
for accumulation that causes it — which is also why the industry's fixes don't
work: carbon offsets are just accounting tricks with no real effect, and more
efficient hardware only makes computing cheaper and demand grow faster (the
Jevons Paradox), rather than actually cutting total consumption. There's no
technical shortcut here — only choosing to compute enough rather than compute
more can address the problem at its root.
2. How the energy crisis is being created
Electricity has become the binding constraint on AI
expansion, with the industry absorbing power at a pace that is outrunning
grids. A single AI query already uses roughly 10 times the electricity of an
ordinary web search, and training a frontier model like GPT-4 consumed tens of
gigawatt-hours — enough to power a small city for weeks. The IEA projects that
global data-centre electricity demand could exceed 1,000 TWh by 2026, a figure
comparable to Japan's entire national electricity consumption. The strain is
already visible on the ground: Northern Virginia, the world's largest
data-centre hub, is seeing growth outstrip transmission capacity; Ireland's
data centres could reach 32% of national electricity demand by 2031, crowding
out renewables meant for households; and Singapore was forced into a multi-year
moratorium on new facilities.
Because AI needs constant, 24/7 'baseload' power that
intermittent renewables cannot reliably supply, technology firms are signing
long-term contracts for gas plants and reviving nuclear sites (the infamous
Three Mile Island among them) — extending fossil-fuel infrastructure under a
green rebrand. This is compounded by the Jevons Paradox: each new generation of
chips (A100 → H100 → Blackwell) is more efficient per watt, but instead of
reducing total energy use, cheaper computation simply stimulates more of it.
OpenAI's own data show compute demand for frontier training runs doubling
roughly every 3.4 months — a
growth rate no efficiency gain can offset. In India, an estimated 200 planned
data centres would draw around 4 GW (1–2% of
peak demand), deepening reliance on coal and straining urban grids in Delhi
NCR, Mumbai and Bengaluru.
3. How the water crisis is created[ix]
Water is AI's 'hidden metabolism'. Data
centres require vast volumes of freshwater — both to cool servers directly and
to cool the thermal power plants that generate their electricity. Engineers are
caught in a stark trade-off: 'dry cooling' conserves water but draws more
power, while 'evaporative cooling' saves power but consumes millions of litres
of freshwater, most of which is lost to the local watershed rather than
returned — this is water consumption, not mere withdrawal.[x]
|
Process |
Water
footprint |
|
Training GPT-3 |
~700,000 litres |
|
20–50 ChatGPT questions |
~0.5 litres |
|
Microsoft, annual increase (2022) |
34% |
|
Google, annual increase (2022) |
20% |
|
Greater Noida facility, 160 MW |
~4.2 billion litres/year (town-scale) |
India
makes the stakes plain: it houses roughly 18% of the world's population but
holds only about 4% of its freshwater. Yet data centres are being concentrated
precisely in its already water-stressed cities — Hyderabad, Chennai, Bengaluru,
Mumbai, NCR. One projection has India's data-centre water use climbing from
roughly 150 billion litres in 2025 to around 358 billion litres by 2030. Since
this is a slow, cumulative draining of aquifers rather than a single dramatic
disaster, it amounts to 'slow violence' — a gradual, largely invisible
diversion of water away from agriculture and drinking supply and toward
corporate cooling towers.[xi]
4.
Why GPU and memory-chip costs are rising
A CPU is designed
for sequential control and logic — a handful of powerful cores working through
varied tasks one after another. Deep learning, by contrast, is dominated by
matrix multiplication (y = f(Wx + b)) repeated across millions of independent
neurons and data points — a workload that is 'embarrassingly parallel'. GPUs,
made up of thousands of simpler cores, can run these identical operations
simultaneously, delivering 10–100× speedups over CPUs. This is why training a
large model takes days on a GPU cluster instead of months on CPUs, and why
modern AI would be practically impossible without them.
During
training, a GPU must hold model weights, input batches, intermediate
activations (needed for backpropagation), gradients and optimiser state — all
simultaneously, and all accessible at very high bandwidth. System RAM is large
but relatively slow; VRAM (memory built directly onto the GPU) is smaller but
far faster, which is exactly what parallel matrix math demands. A model with
100 billion-plus parameters needs roughly 1–2 terabytes of VRAM once training
overheads are factored in — well beyond what any single chip can hold — forcing
firms to link thousands of GPUs together (multi-GPU/model-parallel clusters)
just to fit one model.
This
has led to a steep rise in costs. This technical demand — ever-larger,
ever-faster, high-bandwidth memory built into GPU dies — is what's driving GPU
and memory-chip prices upward. Demand for high-end AI accelerators (NVIDIA's
A100/H100/Blackwell class, for instance, each drawing 300–700 watts and needing
40–80 GB of VRAM) far exceeds supply, and the specialised high-bandwidth memory
(HBM) stacked onto these chips comes from only a handful of manufacturers,
creating a bottleneck across the entire memory market — pushing up prices even
for chips with no connection to AI. Training a single frontier model can now cost
$15–50 million, with the most advanced runs topping $100 million, GPU compute
being the dominant expense.
Not
just the costs, manufacturing these chips is detrimental to environment.
Building this hardware is itself ecologically extractive, and the burden falls
disproportionately on the Global South: The Democratic Republic of Congo
supplies roughly 70% of the world's cobalt through hazardous labour and land
degradation; lithium extraction in the Bolivia–Chile–Argentina 'Lithium
Triangle' drains saline aquifers vital to high-altitude ecosystems; and
rare-earth processing concentrated in China and Myanmar generates radioactive
and toxic waste. Compounding this, AI's competitive pressure enforces a 2–3
year hardware replacement cycle (planned obsolescence), pushing global e-waste
from 62 million metric tons in 2022 toward a projected 82 million metric tons
by 2030 — much of it exported to West Africa and South Asia for hazardous,
informal processing.
5.
Case study: the Vizag (Visakhapatnam) data centre, India[xii]
The
Adani–Google AI data-centre project in Visakhapatnam, Andhra Pradesh, shows how
state power is deployed to fast-track ecologically damaging projects on behalf
of monopoly capital. Clearance itself was rushed through in just nine days —
filed on April 9, 2026 and cleared by April 18 — timed deliberately to precede
an April 28 foundation ceremony. The project also exploited a regulatory
loophole: by keeping the built-up area at 1.49 lakh sq. metres, just under the
1.5 lakh threshold, and with the help of consultancy Pridhvi Envirotech, it
avoided a full Environmental Impact Assessment and the mandatory public
hearings that would otherwise apply. Ecologically, the site overlaps 90% with
the Pedda Chukka Konda Reserve Forest and sits barely 1.53 km from the
Kambalakonda Wildlife Sanctuary — proximity that should legally trigger
Category A status and federal oversight, yet this too was bypassed. The context
is further militarised: the state government also pledged to fast-track
clearances for the nearby 'Project Varsha' nuclear-submarine port, showing how
state and defence interests align to shield the project from scrutiny. And the
resource draw is substantial — 501 kilolitres of water per day and 1,626 MW of
power, backed by roughly 350 diesel generators — a massive regional claim on
resources set against comparatively limited local employment.
The
crisis here is not simply that a data centre consumes resources — it is that
the normal channels of environmental accountability (impact assessment, public
hearings, protected-area buffers) were deliberately routed around, in a city
and state already under water stress, to suit corporate and geopolitical
timelines. It stands as a template for how monopoly capital converts regulatory
weakness into a de facto subsidy.
6.
How this is helping monopolies grow
After
understanding how AI uses resources, it is easy to understand how it is helping
monopolies to grow. The sheer scale of capital, energy and hardware needed to
compete in AI is itself a barrier to entry — one that disproportionately
favours the handful of firms already able to absorb it. The industry is
dominated by a small set of hyperscalers — Google, Microsoft, Amazon, Meta and
their Chinese counterparts — for whom computational capacity is not merely a
production input but a strategic weapon: the firm commanding more compute
trains bigger models, attracts more users and accumulates more data, further
entrenching its market position. This is a self-reinforcing 'computational arms
race' — no single firm can unilaterally pull back without ceding ground to
rivals, so the industry as a whole keeps expanding its resource footprint well
beyond what any honest assessment of social benefit would justify.
The
same dynamic underlies the pattern seen in Vizag: regulatory shortcuts, cheap
land and lax enforcement act as an indirect subsidy that only firms with
billions in capital — Adani, Google — can capture at this scale, while
surrounding communities absorb the water stress, ecological loss and diesel
pollution. Meanwhile, the extractive end of the supply chain — cobalt, lithium,
rare earths, e-waste disposal — is offloaded onto the Global South, so the
ecological cost of monopoly growth in the North is settled elsewhere. Because
efficiency gains do nothing to slow this expansion (the Jevons Paradox again),
and because the real advantage lies in capital scale rather than social
usefulness, market competition left to itself tends to concentrate AI's gains
rather than distribute them — deepening monopoly power even as it deepens the
ecological and social costs outlined above.
Book Destruction for AI Training: A Legal Loophole
Anthropic, the company behind Claude, ran an internal
effort known as "Project Panama" to scan virtually all the world's
books. It spent millions acquiring physical books, then stripped their
bindings, scanned the pages, and discarded the originals — all to build
training data for its AI models. Many rare books, books with no other physical
copy, have been destroyed. Now no physical copy is left, and digitized on is in
possession of Anthropic.
A separate, more serious issue involved over seven
million books pirated from shadow libraries like Books3 and LibGen. A court
ruled that training AI on lawfully purchased books counts as fair use, but
retaining a permanent library built from pirated copies does not. Anthropic
settled the piracy claim for $1.5 billion.
The deeper loophole lies in copyright law itself:
buyers of a physical book have the right to dispose of it. Courts extended this
to allow converting purchased books into digital copies and using them to train
AI — all as fair use. This means a company can legally buy a book, destroy it,
and profit from an AI system built on its contents, while the author has no
claim to compensation.
Copyright law asks only whether reproduction is lawful
— not whether a book is culturally irreplaceable. Anthropic says it avoided
rare or antiquarian titles, but the sheer scale of destruction, driven purely
by profit, raises a conservation question the law was never built to answer:
should companies be free to treat physical cultural artefacts as disposable raw
material in pursuit of AI dominance?
Workers as Data: The New Face
of AI Labour
A new and disturbing
dimension of artificial intelligence is emerging in the workplace: workers
themselves are becoming sources of data for AI systems. Companies are
increasingly using wearable devices, scanners, smartphones and cameras to
record workers' movements, location, task completion and work patterns. In
systems of algorithmic management, these data can be used to determine
schedules, measure productivity, monitor idle time and even influence
disciplinary decisions. The worker is no longer merely supervised by a human
manager; their physical activity is converted into continuous streams of
machine-readable data.
In India, an even more
striking development has been reported. Workers in factories and the services
sector are being asked—or in some cases reportedly required—to wear
head-mounted cameras or smartphones while performing their ordinary jobs. These
devices record “egocentric data”, showing precisely what a worker sees and does
while operating machinery, cooking, handling materials or performing household
tasks. Such recordings are valuable to companies developing embodied AI and
humanoid robots, because they can be used to teach machines how human beings
perform complex physical tasks. There are also paid data-collection jobs in
which people voluntarily mount smartphones or cameras on themselves and are
paid by the hour or by the quantity of accepted recordings.
This creates a remarkable
contradiction at the heart of the emerging AI economy. Workers may be paid to
produce the data needed to teach machines how to perform their own jobs—and
potentially to replace them. A worker's labour is therefore being commodified
twice: first as the work they perform, and second as the behavioural data
generated while performing it. The immediate payment may be modest, while the
resulting dataset can become a valuable commercial asset for AI and robotics
companies. The central questions are consequently not only about employment and
automation, but also about consent, privacy, ownership of workplace data,
workers' rights over the value generated from their activities, and whether
workers should have any say when their accumulated skills and movements are
converted into the knowledge base for machines designed to replicate them.
IV Conclusion
AI has undoubtedly opened
enormous possibilities. It can raise productivity, assist human creativity,
improve healthcare and research, perform dangerous work, and relieve people of
repetitive and exhausting tasks. Yet the evidence examined here shows another
side of the AI revolution: jobs are being displaced or deskilled, workers are
subjected to algorithmic surveillance, their movements and skills are converted
into proprietary data, books and other cultural resources can be appropriated
as raw material for private AI systems, while the enormous requirements of
computing are intensifying the consumption of energy, water and minerals. At
the same time, the immense capital required for AI is concentrating its means
of production in the hands of a few giant corporations.
From a Marxist perspective,
however, the fundamental problem is not AI—or technology itself—but the class
that owns, controls and deploys it. Technology is a product of human labour and
accumulated social knowledge. Under a different social relationship, the
enormous productivity that AI makes possible could mean shorter working hours,
elimination of drudgery, better healthcare and education, safer workplaces, and
greater material abundance for all. But under capitalism, the decisive question
is not what society needs or what human beings could gain from technological
progress; it is what technology can do for the accumulation of capital. AI
becomes another extraordinarily powerful instrument for reducing necessary
labour time, intensifying exploitation, weakening workers' bargaining power and
appropriating surplus.
This is why the same
technology that could liberate human beings from monotonous labour can instead
be used to monitor them, discipline them, replace them and turn their own
skills and activities into data for machines that may eventually replace them.
The same AI that could help conserve knowledge can be used to turn books into
disposable raw material for corporate profit. The same computational power that
could serve socially useful purposes can drive enormous demands for
electricity, water, minerals and land, while the environmental and social costs
are pushed onto workers and communities. The question is therefore not simply
what AI can do, but what benefits it produces for society and what costs
society is required to pay for those benefits.
Nor can technological
development be treated as sustainable merely because individual machines or
data centres become more energy-efficient. AI requires a vast material
infrastructure—electricity, water, minerals, semiconductor production, data
centres and continuous hardware replacement. The pursuit of ever-greater
computing capacity cannot be allowed to proceed without regard to ecological
limits. A technology that increases corporate profits while exhausting water
resources, intensifying energy demand, degrading ecosystems and shifting
environmental costs onto already vulnerable communities cannot be called
socially progressive. The environmental cost of technological development must
be measured alongside its economic and social benefits, and those costs cannot
simply be externalised onto society and future generations.
Capital has historically
appropriated every major advance in the productive forces and subordinated it
to the generation of surplus value. AI represents a particularly powerful
extension of this tendency because, as a general-purpose technology, it can penetrate
almost every sphere of economic and social life. If its development remains
subordinated to the interests of monopoly capital and the pursuit of
superprofits, its consequences can be correspondingly pervasive—and potentially
disastrous. The alternative is not technological backwardness or opposition to
AI, but social ownership and democratic control over its development and
deployment, ensuring that its benefits are socially distributed, its
environmental footprint remains within sustainable limits, and the costs of
technological progress are not imposed upon workers, communities and nature for
the enrichment of a small class.
Ultimately, the question is
not whether humanity should use AI, but for whom, for what purpose, under whose
control, and at whose cost. Technology must serve society—not society serve
technology, and certainly not society be sacrificed to the accumulation of
capital. Only when the productive forces created by humanity are brought under
the conscious and democratic control of humanity can technological progress
become a means of genuine social emancipation rather than another instrument of
exploitation and accumulation.
[i] Luddite Movement
The Luddite movement emerged in early
19th-century England (1811–1816) as a militant response by skilled textile
workers to the rapid mechanisation of production under industrial capitalism.
Faced with wage cuts, job displacement, and the degradation of artisanal
control over work, workers organized clandestinely and targeted
machines—particularly power looms—that symbolized their dispossession. Operating
under the mythical leadership of Ned Ludd, the Luddites were not simply
“anti-technology”; rather, they opposed the specific use of machinery by
employers to intensify exploitation and undermine labour rights. The movement
combined economic demands with direct action, including machine-breaking,
collective organization, and nocturnal raids, making it one of the earliest
forms of organized industrial resistance by the working class.
[ii] The sabotage struggle of workers emerged as a radical tactic in the
early industrial era, particularly championed by figures like Émile Pouget and
William Trautmann, who viewed it as a non-violent yet potent form of class
warfare against exploitative bosses. Unlike outright strikes, which could leave
workers vulnerable to retaliation, sabotage involved subtle disruptions—slowing
production lines, "accidentally" damaging machinery, mislabeling
goods, or feigning incompetence—to impose direct economic costs on capitalists
without drawing immediate blame. Rooted in the Industrial Workers of the World
(IWW) motto "Sabotage is the withdrawal of efficiency," it symbolized
workers' refusal to be mere cogs in the profit machine, echoing the French term
"sabotage" from wooden shoes ("sabots") allegedly thrown
into gears. This strategy peaked during labour unrest in the 1910s U.S. and
Europe but waned under legal crackdowns and anti-syndicalist laws. However, its
spirit persists in modern gig economy tactics like mass app deletions or
algorithmic gaming.
[iii] ILO Study May 2025 “Joint
study One in four jobs at risk of being transformed by GenAI, new ILO–NASK
Global Index shows”
[iv] McKinsey Global Institute, The Future of Work in America,
updated 2023
[v] McKinsey Global Institute Report
November 2017 “Jobs lost, jobs
gained: What the future of work will mean for jobs, skills, and wages”
[vi] Ibid January 2025 “Superagency in the workplace:
Empowering people to unlock AI’s full potential”
[vii] PwC.com, June 2025 “The
Fearless Future: 2025 Global AI Jobs Barometer”
[viii] The Jevons Paradox is
an economic concept first identified by the English economist William Stanley
Jevons in 1865 during his analysis of British coal consumption. Jevons
discovered that improvements in the efficiency of steam engines—meaning it took
less coal to perform a given unit of work—actually accelerated total coal
consumption rather than reducing it. This paradox occurs because as efficiency
gains lower the cost of using a resource, they stimulate greater demand and
expand energy-intensive economic activity. Today, it is recognized not just as
a historical quirk, but as a structural feature of capital accumulation
[ix] AI programs consume large volumes of
scarce water, UCR study by David Danelski
[x] Resources, Conservation and Recycling “The water use of data center workloads: A
review and assessment of key determinants” Vol 219, June 2025
[xi] International Journal of Computing & Artificial Intelligence, 2026,
Vol. 7, Issue 5, Part A Silent
thirst: Water footprints of AI data centres in India and routes to resilience
[xii] Reuters August 6, 2026 Google’s $15 billion India data centre project battles water, wildlife
concerns
