Not using AI is a luxury

In June last year, more than seventy authors, Colleen Hoover, R.F. Kuang and Lauren Groff among them, signed an open letter demanding that their publishers "never release books that were created by machines." Every one of them has a publisher.

The instructions keep coming, from every direction. During the January 2025 Los Angeles fires, a post blaming ChatGPT for helping warm the planet collected more than 150,000 likes. The Washington Post put "a bottle of water per email" in a headline. Merriam-Webster made "slop" its word of the year. The Associated Press published advice on cutting back: don't ask it for recipes, directions or store hours. New York's public schools banned it outright, for a while.

All of it shares one assumption, and nobody says it out loud: every argument against using AI comes from someone who already has the thing it replaces, whether that is the editor, the tutor, the doctor, or the money, time and people who answer the phone. Not using AI is a luxury, and the bill for refusing it is being handed to the people holding none of those things.

The case is really four claims

Strip the discourse to its parts and the case against using AI comes down to four claims: it is destroying the planet, it is rotting your mind, it is cheating, and it is slop no serious person should touch. Each one carries real truth, and this essay concedes that truth in full as it goes. But each one also gets cheaper to hold the more alternatives you own, and lands heavier the fewer you have. So take the claims in turn, with the strongest numbers available, the critics' own numbers wherever they exist. If the case for abstaining survives the arithmetic, take it.

"It's destroying the planet"

Start with what one prompt costs. Three estimates that share no methodology converge: Google measured its median Gemini text prompt at 0.24 watt-hours in August 2025, Sam Altman claimed 0.34 for an average ChatGPT query, and Epoch AI independently modelled roughly 0.3. Treat all three with suspicion, two are self-reports, and they still land in the same place: about a third of a watt-hour, less energy than Google's own 2009 figure for a single web search, and about 11 to 16 seconds of video streaming on the International Energy Agency's (IEA) numbers.

Water is the sharper accusation and the messier fight, because it is an accounting fight. Google counts what its data centres drink on site: 0.26 millilitres per prompt, five drops. Mistral, the only AI company to publish a full lifecycle analysis reviewed by third parties, counts the water behind the electricity too and lands at 45 millilitres. Shaolei Ren, the researcher behind the original AI water study, puts the honest total near 50. So take 50: ten prompts a day, every day for a year, comes to about 180 litres. Then watch what circulation did to the research. Ren's paper says GPT-3 needed a 500 millilitre bottle for roughly 10 to 50 medium-length responses. By September 2024 that had become the Post's bottle per email. Ren himself now puts a GPT-4 prompt nearer 15 millilitres. The bottle survived every retelling. The denominator did not.

Two honest caveats before the bigger point. The per-prompt numbers hide a heavy tail: load a model with 100,000 tokens of context, most of a book, and one query can near 40 watt-hours. And every figure in this fight goes stale fast: Google says its median prompt's energy fell 33-fold in one year, which retires the reassuring numbers as fast as the scary ones.

The bigger point is about where the real problem lives. The total is real: the IEA counted about 415 terawatt-hours for the world's data centres in 2024 and expects that to roughly double by 2030, to more than Japan consumes. The local harms are real too, and they have street addresses: thermal imaging found as many as 35 gas turbines running at xAI's Memphis data centre where 15 were permitted, beside a majority-Black neighbourhood; Google drew 29 percent of one Oregon city's water in a year and fought for 13 months to keep that secret; Google planned a Uruguayan centre that would drink 7.6 million litres a day during the country's worst drought in 74 years.

Now notice what those three stories share. In each one a company created the harm, and where it got fixed, pressure on the company or a regulator fixed it: protests pushed the Uruguay design to air cooling, journalists forced the Oregon numbers out, a county commission capped the Memphis turbines at 15. Nobody's deleted chatbot appears anywhere in that chain. That is the whole argument of this section: the planet case is real, and it points at boardrooms and permitting offices. Aimed at you, it asks for a third of a watt-hour and fixes none of the above.

"It rots your brain, and it's cheating"

The brain-rot front rests almost entirely on one study. In June 2025, MIT Media Lab researchers posted a preprint: 54 people wrote essays while wearing electrode caps, some with ChatGPT, some with Google, some with neither, and the ChatGPT group showed the weakest brain connectivity on the task. The authors called it "cognitive debt." The study is real and worth taking seriously. It is also 54 people writing practice essays, only 18 of whom returned for the final session, and a year later it remains an unpublished preprint. Ask the authors themselves whether it shows AI makes us dumber and their own project FAQ answers: "No!", with a plea to media not to use words like "brain rot", "harm" or "damage". The headlines used them anyway, within days. The bottle got a sequel.

What the measured record actually shows is that structure decides everything. At Harvard, a purpose-built AI tutor more than doubled the learning gains of an active-learning physics class in a randomised trial of 194 students, in less time. Across 37 studies, a 2025 meta-analysis found ChatGPT moderately improved academic achievement (a separate, more flattering meta-analysis was retracted in April 2026, so treat the modest number as the honest one). In Nigeria, six weeks of teacher-guided, after-school GPT-4 tutoring produced gains the World Bank authors equate to one and a half to two years of business-as-usual schooling, with the caveats carried: a short pilot, gains skewed to girls and stronger students, no long-term follow-up. And where structure was absent, the story flipped: a GPT-4 business mentor given to 640 Kenyan entrepreneurs produced no average benefit at all; strong operators gained, strugglers lost. Unguided access rewards the already equipped.

Concede what deskilling evidence exists, because some does. After AI-assisted colonoscopy arrived at four Polish centres, experienced doctors' detection rates in their non-assisted procedures fell from 28.4 to 22.4 percent. That is measured, patient-relevant, and an argument for designing guardrails, not for banning the tool that also catches what humans miss.

As for cheating: New York's schools banned ChatGPT in January 2023 and reversed within four months, the chancellor writing that "the knee-jerk fear and risk overlooked the potential of generative AI to support students and teachers." Cambridge's current policy makes unacknowledged AI content misconduct while explicitly permitting AI "to support their personal study": read closely, the cheating rule was always about acknowledgment, never about use. Ohio State now requires AI fluency of every undergraduate from the class of 2029 onward. Schools have run this exact loop before. The College Board allowed calculators on AP exams in 1983, banned them in 1984 on the grounds that it wasn't fair to students who didn't have one, then made them mandatory a decade later. Note the shape of that ban: faced with unequal access, the gatekeepers took the tool from everyone rather than getting it to everyone. A ban always lands unevenly. The student with a tutor at home loses nothing. The student in a classroom with one trained teacher for 58 pupils, the sub-Saharan African average against 14 students per teacher across the OECD (the club of mostly rich countries), loses the only tutor she was ever going to get.

"It's slop. Hire a professional."

The craft objection deserves its strongest voices. Nick Cave, sent a ChatGPT song written "in the style of Nick Cave," called it "a grotesque mockery of what it is to be human." Ted Chiang argued in The New Yorker that art is the accumulation of thousands of choices, and prompting a machine surrenders them. The writers' letter runs on a grievance no fair reader can dismiss: these models were trained on their books without consent or payment. On art, they are substantially right, and nothing in this essay argues a chatbot should write your novel.

But watch the category slide. When Simon Willison mainstreamed the word "slop" in 2024, his definition was AI content "mindlessly generated and thrust upon someone who didn't ask for it," and he was explicit that reviewed, accountable AI-assisted work is exempt: "Not all AI-generated content is slop." By the time "slop" was being crowned word of the year, the label had stretched to cover nearly anything a machine touched. Even the purity badge is softer than its users think: the "Not By AI" badge certifies, by self-declaration, that a mere 90 percent of the content is human, with AI grammar checks and "inspiration" allowed inside the 90.

The aesthetic objection is strongest where the stakes are aesthetic. Where the stakes are access, it collapses into its bluntest form, "hire a professional," and that is where the arithmetic turns brutal. The OECD averages 3.43 physicians per 1,000 people; sub-Saharan Africa has 0.25. In Nairobi, clinicians working with an AI reference tool across 39,849 patient visits made 16 percent fewer diagnostic errors than colleagues without it (a study co-authored by OpenAI and not randomised; hold that against it, and notice it does not explain the errors away). In Ghana, an AI maths tutor on school phones cost about $5 per student per year; ordinary human tutoring runs about $2.60 an hour even in Manila, one of the cheapest markets with published rates, so a single month of weekly sessions costs more than the AI's two years. The small business owners I build marketing systems for in South Africa have no agency, no analyst, no copywriter. One owner, one phone, one chat window. Call that slop and what you are really telling her is to hire a professional instead. There is no professional in this story. There never was. Her choice is the chatbot or nothing.

The poor don't get to refuse it, either

Be honest about the other gate. Refusing AI assumes you could have used it: 2.2 billion people remain offline, 96 percent of them in low- and middle-income countries; an entry-level basket of mobile data costs 4.2 percent of average national income per person in Africa against 0.3 percent in Europe; 810 million women in those countries do not use mobile internet at all. Owning a laptop and a connection already puts you ahead of a quarter of humanity. And refusal is gated from above as well: Kali Holloway argued in The Nation this April that "the human touch has become a luxury good," with brands now selling human craftsmanship as the premium tier while the poor have the most consequential parts of their lives, benefits, care, even parole, decided by AI systems by default, without ever being asked. Both halves are true at once. The full luxury is having the choice: the option to use AI, and the option to still reach a human. The people with both options are lecturing the people with neither.

We have shamed the cheap version before

Every technology that made a scarce good cheap was met by people holding the expensive version, explaining that the cheap version corrupts. When writing threatened trained memory, Plato had a king warn that this invention "will produce forgetfulness in the minds of those who learn to use it," offering "the appearance of wisdom, not true wisdom." We know the objection because Plato wrote it down. When novels made stories cheap, the essayist Vicesimus Knox wrote in 1778 that "the great multiplication of Novels probably contributes to its degeneracy," and a 1797 magazine piece was titled, in full, "Novel Reading, a Cause of Female Depravity." When the penny press made news cheap, politicians and clergy declared a "Moral War" on the New York Herald. The psychologist Amy Orben calls this the "Sisyphean cycle of technology panics": novels, radios, smartphones, each panic burning down and restarting on the next machine. And the sociology of abstention says the quiet part: media refusal works through "conspicuous displays of non-consumption" that read to everyone else as "a performance of elitism."

The geography of AI opinion fits the pattern uncomfortably well. In Ipsos's 30-country 2025 survey, the people most excited about AI are in Indonesia (80 percent), Thailand (79) and Malaysia (77); Great Britain (37) and the United States (38) sit at the bottom. One honest complication, because the pattern is not clean: within countries, Pew finds less-educated people are more worried about AI, not less. The asymmetry lives between countries, and in who gets to publish the objections: the op-eds demanding restraint are written where the compute is, in the 32 countries that host AI data centres, about the users in the 150 countries that host none. Even flight shame, the nearest modern cousin of this discourse, at least aimed upward, at the 1 percent of the world's population likely responsible for more than half of passenger aviation emissions. AI shame aims at the newest users of the cheapest expertise available, who mostly live in the countries most excited to finally have it.

Hold the objections. Drop the shame.

Every one of the four claims contains something worth acting on, and none of it is aimed at the user. If you hold the environmental objection, demand lifecycle disclosure at the grade Mistral already met and permitting with teeth: Bloomberg found about two-thirds of American data centres built since 2022 sit in water-stressed places, which is a zoning fact no consumer created, and the serious campaigns already target policy (Food & Water Watch's demand is a Congressional moratorium, not your abstinence). If you hold the learning objection, demand structured use and teaching, because that is what the evidence rewards, and know that a ban hands the advantage to the child who has a tutor anyway. If you hold the craft objection, demand consent and payment from the companies that trained on writers' work, which is what the letter actually asks its publishers to enforce. All three point at institutions with names and addresses.

What remains, after the institutions get their share, is a student in Benin City whose six weeks of after-school tutoring, a third of a watt-hour at a time, bought her most of two years of school she was not otherwise going to get. The authors of the June letter can hold their principle; it costs them nothing, because they already have the thing the machine replaces. For her the same principle has a price, and it is the highest price in the whole discourse. Not using AI is a luxury. Stop billing it to the people who can least afford it.

A note on numbers

Per-prompt environmental figures are company self-reports (Google, Mistral, OpenAI) or independent modelling (Epoch AI); none is independently audited, and the water figures differ mainly by accounting boundary, so where the argument leans on water it uses the higher, critics' number. The MIT study is described with its authors' own stated caveats and remained an unpublished preprint as of mid-2026. Learning-effect sizes are reported from randomised trials and the surviving (non-retracted) meta-analysis, translated into plain terms. Google reports its median prompt's energy fell 33-fold in a year, so expect every static figure here, alarming or reassuring, to age. Written July 2026; sources are linked or named where each claim appears.