AI works. Now we need to manage its impact
At SEOZoom, we use artificial intelligence every day. We studied it when it was still a topic for industry insiders; we integrated it into our platform as soon as it stopped being an experiment; and today, part of what we offer—writing tools, visibility analysis in generative search engines, and prompt monitoring—exists because AI works. We’re saying this up front because any discussion of the costs of artificial intelligence written by those who use it—and, in part, sell it—risks sounding like a hypocritical lecture and a mere attempt to absolve ourselves. We can’t lecture anyone.
However, we can say something more useful, precisely because we use it every day: we know how to distinguish where AI becomes work and where it becomes waste. If it were a useless technology, exposing it would be easy, and the debate would resolve itself. Instead, it works, it integrates into processes, and it improves operational steps. And its effectiveness is measured in terawatt-hours, in trillions of liters of water, in square kilometers of land occupied by its physical infrastructure.
That’s precisely why it’s become dangerous. When a tool is useful, we stop questioning it. We normalize it. And everything it consumes—both materially and cognitively—disappears behind the convenience of using it.
We need AI. That’s what makes it a problem.
Convenience is the mechanism that silences questions
Think about what it was like to search for something just three years ago. You’d type a query into Google, scroll through the snippets on the first page trying to figure out which one actually met your need, work your way through ads, clickbait headlines, and sites that promised one thing but delivered another. You’d click, be disappointed, go back, and rephrase your query. It was a process full of friction, and it was annoying.
Today, you type a question and get a ready-made, relevant answer in natural language. No one, given the choice, would willingly go back to the old way.
That convenience—more than just laziness—is relief. And it forces those of us working in this industry to make an uncomfortable admission: the obstacle course that AI is freeing us from was, in part, built by the web itself. Pages padded out to cover one more query, answers dragged out paragraph by paragraph to keep the user engaged, comparison sites disguised as guides, headlines that were more compelling than the content they promised. A part of the web that today fears being bypassed by AI helped make that leap desirable. We say this from the inside, because that’s our line of work, and we’ve seen those trends up close.
That friction, however, also served another purpose. Every snippet to be evaluated was an exercise in judgment, every source to be weighed a small test of reliability, every reformulated query a refinement of how you framed the problem. The effort of searching was the moment when your expertise came into play. The concise answer gives you the result and takes away the journey, and with the journey go the two most important questions—how much do I trust what I’m reading, and what impact did obtaining it have? Convenience acts as an anesthetic, right on the nerve endings that should signal the weight of the information to you.
Then there’s a level where choice is taken away from you entirely. Generative responses have become the default across half the web—from the AI Overview at the top of search results to assistants built into operating systems and professional apps, all the way to thesilent calls running in the background within everyday software.
It consumes computational resources and performs inferences even when you haven’t asked for them. A single conscious action matters, but it must be understood within an architecture that pushes in the opposite direction—one that has transformed our interaction with these systems from a decision into an environmental condition.
The cost of AI has shifted into our daily lives
The fluidity you experience in front of the screen has a heavy body on the other side, made of silicon, copper, drinking water for cooling, and power plants pushing data centers to their limits. For years, the debate about this body has focused on training—the training of large models—portrayed as a massive and distant event, the exclusive responsibility of the few companies with the means to afford it.
It was a convenient way of looking at things, because it kept the problemout of our reach. The most recent data debunks this view. According to a report published in June 2026 by the United Nations University, once a model is in production, inference—the responses to users’ daily requests—accounts for 80–90% of all the energy that system will consume. ChatGPT alone processes about 2.5 billion prompts per day. The impact of AI has shifted, moving from the lab to the desk of anyone who uses it.
The overall numbers are even more alarming. In 2025, data centers consumed approximately 448 TWh of electricity globally, and the projection for 2030 reaches 945 TWh—roughly the annual consumption of all of Japan and nearly triple the combined consumption of Pakistan, Bangladesh, and Nigeria, countries with a combined population of over 650 million.
In 2024, these same facilities consumed 4.5 trillion liters of water—enough to meet the needs of more than 600 million people in sub-Saharan Africa—and generated 189 million metric tons of CO₂, equivalent to Argentina’s annual emissions.
Estimates for 2030 point to 9.3 trillion liters of water per year—the basic domestic needs of 1.3 billion people—and a land footprint exceeding 14,500 square kilometers, roughly double the size of the Jakarta metropolitan area. When it comes to electronic waste, discarded AI hardware could reach 2.5 million metric tons per year by the end of the decade.
Among these data points are requests that drive medical diagnoses and scientific research, as well as those asking the model to rewrite a three-line email. The system handles them all with the same diligence. For now, the only one who can distinguish between them is the person typing them.
Five drops of water multiplied by billions
Google has attempted to quantify the cost of a single request by publishing the first comprehensive measurement of its system in production. The median Gemini prompt consumes 0.24 watt-hours of energy, emits 0.03 grams of CO2 equivalent, and uses 0.26 milliliters of water—about five drops; Sam Altman has reported similar figures for ChatGPT.
Five drops of water amount to less than what evaporates from the glass on your desk.
But try doing the math. One million median prompts amount to 240 kWh and 260 liters of water; 2.5 billion prompts per day—for just one service among many—account for exactly the national-scale consumption that the UN report documents.
And there’s a second variable that matters just as much as quantity: the type of request. A typical conversation consumes about 200 times the energy of basic text classification; image generation reaches 1,450 times, and for the latest reasoning models, independent estimates rise even more sharply. The cost depends on how you use AI even more than on how much you use it.
The numbers are incomplete, and that incompleteness is part of the problem
The figures just cited should, however, be treated with caution, because the industry measures itself using methodologies that contradict one another. Google claims a comprehensive approach—which includes active chips, idle machines, infrastructure, and cooling—and disputes previous academic estimates, which reached as high as 7 watt-hours and 50 milliliters of water per request. Shaolei Ren, the author of those estimates, responds that the comparison is misleading, because his calculation also includes the indirect water consumption of the power plants that generate electricity—an item that Google’s measurement leaves out. Missing from the picture is the data that would settle the matter once and for all: the total volume of queries, which no platform publishes.
This lack of transparency has a consequence that affects you directly. Without the full set of numbers, you decide how to use AI without being able to truly gauge what it consumes; you can estimate in orders of magnitude, but the precise figures remain in the hands of those with a vested interest in presenting them in the most favorable light. If you work with data, this asymmetry is familiar to you.
Efficiency improves, but the total cost still rises
The platforms respond to criticism with an argument that appears solid: progress in efficiency. And it is real progress—even spectacular. Over the course of twelve months, Gemini’s median prompt reduced its energy consumption by 33 times and its carbon footprint by 44 times. During the same period, however, Google’s total emissions grew by 51% compared to 2019, driven precisely by the expansion of the computing capacity needed to train and run generative systems. These two figures coexist without contradicting each other, and illustrate a mechanism the economy has known for a century and a half: the rebound effect, whereby when a resource costs less, more of it is consumed.
The cheaper a single prompt becomes, the more uses it accommodates—more built-in features by default, more automated requests, more generations generated simply because “it costs nothing.” Efficiency per unit is the best ally of growth in absolute consumption. Expecting technical progress to solve the problem on its own means ignoring the fact that it is precisely technical progress that makes it cost-effective.
The web is paying a second price, and it’s made of words
The material footprint is the visible part of the cost, and there is a second one—less measurable in liters and watt-hours but just as concrete for those who work with content. It concerns what all this generative capacity is actually producing.
Every piece of generated text consumes energy twice: first when it is created, and second when it must be processed, indexed, and scanned by crawlers and models tasked with deciding whether to use it. If that text is useful to someone, the cost is justified.
If it’s yet another anonymous variation of existing content, it’s pure waste—energy spent adding noise to an archive that’s already overflowing with it—and more energy spent by everyone else to dispose of it.
Half of all new online content is generated by a machine
The scale of this phenomenon was measured in a study published by Graphite, which analyzed 55,400 English-language articles published between 2020 and the first quarter of 2026, classifying them using three different detection systems. Before ChatGPT, machine-generated content accounted for only a marginal fraction of total output; in the first quarter of 2026, it made up 49.9% of everything published—a share that has remained stable at around half for over a year now. The new web—the content created every day—is half automated writing.
The figure that changes the way we interpret this, however, is the second one. 86% of the pages that rank on Google are still written by humans, and the same is true for 82% of the sources cited by ChatGPT and Perplexity. Taken together, these two numbers describe texts that consume energy to exist and are discarded by all the channels that are supposed to distribute them. Selection systems—search engines and generative engines—continue to prioritize experience, data, and recognizability.
Cognitive debt is the cost you pay last
The third level concerns those who use AI, and it has begun to emerge in experimental studies. The most discussed study comes from the MIT Media Lab, which monitored 54 participants using electroencephalograms as they wrote essays under three different conditions—with an LLM, with a search engine, and without any tools.
Those who wrote with AI showed the weakest neural connectivity of the three groups, reported the lowest sense of ownership over the text they produced, and—in a revealing detail—struggled to correctly cite their own work just a few minutes after submission. The researchers called this phenomenon cognitive debt, by analogy with technical debt in software—an immediate efficiency gain that accumulates structural costs over time. The study is a preprint that has not yet undergone peer review, and its limitations must be kept in mind, but the direction it points to is consistent with what psychology has documented for years regarding cognitive offloading.
The professional version of this same mechanism has already produced textbook cases. In the fall of 2025, Deloitte Australia reimbursed part of the 440,000 Australian dollars it had received from the government for a 237-page report that contained references to academic articles that never existed and a fabricated citation from a federal court ruling; the corrected version had to disclose, after the fact, the use of generative AI in its drafting.
Model hallucination is a well-known and documented limitation; the decisive error lies with the human who incorporates that output into an official process without verifying it. The fluidity of the responses works against you, because a well-written text seems true even when it’s completely fabricated.
The line between what you control and what is beyond your control
Even if every user on the planet were to use AI flawlessly starting tomorrow, data centers would continue to grow, because the current race is infrastructural and financial rather than behavioral, fueled by investments measured in the hundreds of billions and by competition among platforms that is driven by installed capacity—far exceeding the demand expressed by users. Conscious use by individuals reduces marginal waste—and that’s where it ends. Anyone who tells you that the solution lies in your habits is selling you a get-out-of-jail-free card, and we, for one, refuse to sell it.
The fact remains that part of the balance really does depend on you, and that’s the part where a serious discussion can make a difference. The waste of unnecessary requests, the production of text that no system will ever select, delegation without verification, the atrophy of skills you stop practicing. These are small items compared to the trillions of liters, but they’re the only ones entirely under your control, and when added up across millions of professionals, they cease to be small. Managing that part is less than what’s needed and more than what we’re currently doing.
The Ten Rules About AI We’ve Learned by Using It
The rules we follow at SEOZoom come from daily use, from years of integrating AI into our processes and tools, and some we’ve learned through trial and error—we, too, have generated our share of futile attempts before figuring out where the tool delivers and where it falls short.
We’re sharing them for what they are: work practices—a sort of query ecology—that improve the quality of what you produce and reduce your share of waste, within the boundaries just outlined.
- Reserve for it the tasks that are worth delegating. If explaining the work to the machine takes more effort than doing it yourself, do it yourself. Delegation makes sense when it frees up real time; the rest is waste disguised as innovation.
- Automate the long and repetitive tasks; keep the thinking to yourself. Automation streamlines high-volume, low-judgment activities; ideation, perspective, and thesis remain your tasks.
- Choose a tool that’s proportionate to the task. A quick check is done with a search, a summary with a lightweight model, and image generation should be reserved for when it’s truly needed, since it consumes hundreds of times more resources than a text request. The cost depends on how you use AI rather than how much.
- Prepare your prompt before submitting it. Every regeneration incurs a cost—both in energy and time. A prompt crafted with clear context and criteria achieves in one go what ten haphazard attempts fail to do.
- Have a human review everything, statement by statement. Every piece of output intended for the public must be reviewed, and every piece of data must be verified, because fluid responses can sound true even when they’re fabricated. The responsibility for what you publish remains yours even when the drafting is delegated.
- Optimize what you build as well. If you integrate generative functions into software or a workflow, every redundant call is a waste multiplied by the number of your users. Call efficiency is a quality metric, and default automations should only be enabled where they’re needed.
- Publish only what you would have signed off on anyway. Text devoid of experience, data, or perspective adds noise, and consuming resources to produce noise is the very definition of waste.
- Keep honing the skills you delegate. Use AI as an expert in the task—as someone capable of evaluating its output. The judgment you stop exercising atrophies, and at that point, you also lose the ability to spot the machine’s errors.
- Disclose its use when it matters. With readers, customers, and collaborators, transparency about AI intervention costs little; disclosing it later, once an error is discovered, costs much more.
- Measure value; forget about volume. If AI makes you produce more without making you produce better, it’s multiplying waste with your name on it. The only output that justifies its use is the kind that serves someone’s purpose.
Ten rules hold together the three levels of discourse—the material, the editorial, and the cognitive—because, when put to the test,they are one and the same level. A pointless request wastes water and electricity, produces text that no one will read, and dulls your judgment—all in a single action. Proportionate and verified use takes the opposite approach: it conserves resources, produces content that selection systems reward, and keeps sharp the skills that allow you to use the machine rather than be at its mercy.
These tools will remain in our processes and yours, and their resource consumption will continue to grow at a rate that best practices can only slow down. What we can decide, every day, is which side of the divide to stand on—work or waste, tool or habit.
The problem isn’t using artificial intelligence; it’s using it without even realizing it.
