AI Is Not Magic. It Just Has Really Good PR.

Sunday mornings on Doordarshan meant one thing for an entire generation of Indian children: Shaktimaan, descending on his golden disc to remind us that power without discipline can turn into harm. Almost three decades on, a very different kind of power has landed in our hands, and most of us have quietly forgotten the lesson.

Artificial intelligence has now placed an extraordinary amount of that power in our hands. We can ask a machine to write, draw, compose, explain, translate, edit, and create things that would once have taken hours, specialised skills, or an entire team of people. That same power arrives soaked in warnings about its misuse: a fabricated face in a video can cause real harm, a cloned voice can impersonate someone we trust, and synthetic images can spread misinformation faster than anyone can correct it.

We also use that same technology simply because we are bored.

We ask it to place us on a houseboat in Kashmir we have never visited, in a pre-wedding shoot at a palace we could never afford, in a stadium selfie with a cricketer we have never met, or in a different body or outfit. A fake holiday picture is not morally equivalent to an AI-generated scam, and a teenager’s imaginary celebrity selfie does not cause the harm that a deepfake built to deceive millions does. The comparison is not one of equivalence.

The connection is subtler. We have been given an extraordinarily powerful tool, trained to worry about how other people might misuse it, and left almost entirely uninterested in examining how casually we use it ourselves. Creation has become so effortless that we have stopped noticing what it costs.

Consider what happened to photography. A photograph used to be evidence of a moment: a particular place, a particular person, a particular afternoon. It could be staged or retaken, but a real moment sat underneath it. Generative AI has radically changed that relationship. The background can be invented, the lighting rewritten, people added or erased, and if the result is not quite right, we simply generate another.

The image itself takes seconds, but the system producing it does not.

Generative AI runs on data centres filled with specialised processors, cooling systems, and physical hardware. We call it “the cloud,” a word that makes it sound weightless, when in reality it requires electricity, water, minerals, land and manufactured hardware. The International Energy Agency estimates that data centres consumed around 415 terawatt-hours of electricity worldwide in 2024, roughly 1.5% of global electricity use. It projects that this could more than double to around 945 terawatt-hours by 2030, with AI the most important driver of the increase.¹

Those numbers also need context. A simple text query and a high-resolution video generation do not have the same computational footprint. Newer uses such as complex reasoning and agentic systems can demand substantially more computation than a basic request. The environmental cost of AI, therefore, cannot sensibly be reduced to a single number attached to every prompt.

India is part of this expansion too. S&P Global estimates that data centres in India consumed around 13 terawatt-hours of electricity in 2024 and projects that figure could reach nearly 57 terawatt-hours by 2030.² As AI adoption grows, the infrastructure supporting our increasingly digital lives will become an increasingly physical part of the country’s energy demand.

AI is becoming more efficient, which sounds as though it should solve the problem. Efficiency alone, however, does not necessarily solve the underlying issue.

When something becomes cheaper to produce, people often find more reasons to produce it. Economists have studied this for more than a century through what is known as the rebound effect, and in its stronger form, the Jevons Paradox: efficiency can reduce the resources required for each individual task while encouraging enough additional activity that total consumption still rises.

AI offers an unusually clear modern example. A 2026 study of Chinese listed firms found that AI adoption reduced carbon emissions intensity while increasing firms’ total carbon emissions. The authors describe this as an “AI Jevons Paradox.” Their analysis suggests that productivity improvements and expansion associated with AI adoption can contribute to the rise in total emissions even while emissions become lower relative to output.³

That distinction matters because if AI becomes ten times more efficient but we use twenty times more of it, celebrating the efficiency number alone tells only half the story.

None of this means the answer is to tell people to stop using AI. That would be absurd, and arguably counterproductive. AI can help a student understand a difficult concept, a designer prototype an idea, or a researcher analyse a dataset that would otherwise take a team of people months.

Its environmental record is not one-directional either. The IEA estimates that widespread adoption of existing AI applications in areas such as electricity-grid optimisation and building efficiency could reduce emissions by an amount equivalent to several times the emissions produced by data centres themselves by 2035.¹

AI is not simply replacing tasks that already existed. It is also creating entirely new demand for computation: more generated images, more synthetic videos, more automated content, more iterations, and more systems operating in the background because computation has become cheap enough to spend almost casually.

That changes what sustainability should mean. The goal cannot simply be making every AI task more efficient. It has to include getting better at deciding which tasks deserve that efficiency in the first place.

Responsibility for that cannot rest entirely on the person holding the phone. If AI companies optimise only for speed, engagement and revenue, efficiency gains can simply become a reason to sell more computation rather than a reason to spend less of it.

Resource efficiency needs to become part of the equation. When a company decides which model should handle a task, it could weigh not only accuracy, latency and price, but also how much computation the result requires. A simple request should not automatically be routed through the most powerful system available if a smaller model can produce an equally useful result.

At commercial scale, companies could also report meaningful measures of computational efficiency alongside conventional performance metrics. Instead of treating resource consumption as an invisible engineering detail, they could make it something that competes directly with speed, cost and capability.

This does not mean making AI worse; it means becoming more selective about where its power gets spent.

There is a real irony here. We spent decades trying to make computers faster, cheaper and more capable so that they could do more for us. We succeeded so completely that we now have to answer a harder question: what happens once “more” costs us almost nothing to ask for?

AI is going to become more powerful, more efficient and more deeply embedded in ordinary life. Reversing that is neither realistic nor desirable. The responsibility is making sure efficiency never becomes, by itself, an invitation to consume more.

We began on a Sunday morning with a masked hero on a golden disc, and a lesson children’s television gave us for free: power without discipline can turn into harm.

The responsibility of this generation is not to use less technology because it is powerful. It is to get better at deciding what is worth asking it to do.

The future of AI should not be about doing less with it. It should be about deciding, deliberately, what deserves it.

References
International Energy Agency. Energy and AI. 2025.
S&P Global. Will Data Center Growth in India Propel the Country to Global Hub Status? 2025.
Yu, Dongyang & Xu, Bingjie. “Artificial intelligence and carbon emissions: Evidence of an AI Jevons Paradox.” Economic Analysis and Policy, Vol. 90, 2026, pp. 946–966.

WriteToWin Monsoon 2026