Wendell Berry’s wife typed his manuscripts for him. Tanya Berry worked on a Royal standard typewriter that her husband says he bought new in 1956, and according to him she did more than copy. She read as she typed, marked what looked wrong with small checks in the margin, and handed back pages better than the ones she’d received.1 Berry put all of this in “Why I Am Not Going to Buy a Computer,” the 1987 essay, reprinted in Harper’s, that still circulates among people who wish they could quit their phones. He seems to have thought the detail strengthened his case. Plenty of readers thought it sank him, actually, and Harper’s printed their letters, including one that credited Berry with offering writers “a handy alternative: Wife.”2
It’s a funny line, and it hits the mark, because Berry’s refusal of the machine went straight through his wife’s desk. Berry answered his critics by defending marriage as “a state of mutual help” and the household as an economy,3 which is a legitimate position and totally fair to set beside these objections. Any honest account of his position starts there, and so shall this essay. I’m about to attempt to grade the machinery, and I’m going to do it from inside the building.
I am transparent about how I use AI in my work. I use it for editing, verification of citations and endnotes, and some image generation. I don’t, as it has become all too common to repeat, use AI to “write for me.” Every adoption or rejection of a machine rests on somebody’s labor. Berry’s was visible enough that readers spotted it from the cheap seats, and I’d like to find mine where they exist before anyone has to write in about it.
. . .
Berry wrote nine standards for himself back then, and they read like a farmer’s inspection checklist.4 A new tool should cost less than the one it replaces and be at least as small. It should use less energy, and run on the sun or the body if it can. An ordinary person ought to be able to repair it, buy it near home, and return it to a small shop that will take it back. It should do work that is “clearly and demonstrably better” than what it replaces, and it should leave alone whatever good already exists. He wrote the list as rules for his own life, which makes it much harder to dismiss as a manifesto and also easier to take seriously as a method.
Seven of the nine standards concern money, size, energy, and repair, and AI fails all seven in ways that surprise pretty much nobody. Data centers are enormous, they devour electricity, and no one fixes a Trainium chip on the front porch. In fairness, a combine harvester fails most of that list too, and so does the phone in your pocket, which means a failing grade there says more about the distance between the modern economy and Berry’s barn than about artificial intelligence in particular. I’ll give the material side one go in this essay, because it has become too concrete to skip, and spend the majority of the time on the two standards where the verdict remains open.
. . .
We’ll start with “clearly and demonstrably better,” because that’s where the industry makes its real pitch. The pitch is speed, volume, and convenience. Berry saw that coming. When university professors told him a computer would help him write faster and produce more, he answered that speed, ease, and quantity were not his standards, and that what he wanted was to write better, which he thought required help from other people.5
So I put this question to my own work, which is the only place I currently have standing to answer it. Did AI make my writing better, or did it make more writing faster? Better, yes, in the narrow sense of fewer typos and grammatical errors and more accurate endnotes and formatting. AI also saves me a lot of time editorially, by helping to verify sources, format citations and notes, and spellcheck. It has also occasionally turned my writing into some freakish version of itself, replete with a way too eager use of parallelism, negation phrases, em dashes, and a short list of oft-repeated pet words. When that happens, it’s because I asked it to do more than proof and verify, and it reverted to its own algorithms. So the honest assessment is better at the margins, for sure, and faster nearly everywhere, which is the trade Berry warned about.
Some uses clear Berry’s bar without much argument. Translation and accessibility tools are the strongest cases I know of, and I’m in no position to referee any medical claims. A standard that can’t say yes to anything is really just a mood, and Berry’s was never that. He built it to sort, so it ought to sort.
. . .
New Carlisle, Indiana, is about fifteen miles west of South Bend, where I live. A year before the New York Times visited, the 1,200 acres outside town were an empty cornfield. Amazon has since built a campus there called Project Rainier, which the Times reported was designed around a single customer, Anthropic, the company that makes Claude.6 CNBC reported that seven buildings were running by October 2025, and Amazon Web Services CEO Matt Garman said about 500,000 of the company’s own Trainium2 chips were already operating in Indiana.7 The Times reported that the plan calls for around thirty buildings, and that the finished complex will draw 2.2 gigawatts, a figure the paper compared to the needs of a million homes, and will use millions of gallons of water a year.6 For scale, a DeLorean needs 1.21 gigawatts to travel through time. This campus will draw nearly twice that, and so far the only thing it has sent into the past is a cornfield.
Claude also helped me edit this essay, which makes me a paying customer of the very thing I’m grading. Anthropic buys computing power from other providers as well,8 so I can’t claim my particular prompts travel to that nearby cornfield. I can say I’m paying the company the sprawling campus was built for, and I see no reason to try to hide that fact.
Berry’s first standard asks whether the new tool is cheaper, and for the individual user it is. A monthly subscription costs a lot less than the human help it replaces, which is most of the sales pitch, I think. His standard was written for a household purchase, and it strains when the bill might land on someone else. The utility serving the campus, Indiana Michigan Power, told regulators it expects new large customers like data centers to take its Indiana peak load from about 2,800 megawatts to more than 7,000 by around 2030.9 In February 2025 the Indiana Utility Regulatory Commission approved a settlement on the tariff for those large customers, and the question of how to divide the costs among all of I&M’s customers was left for a later rate case.9 That leaves the harder fight for later, which is the part I’d watch. Marcy Kauffman, president of New Carlisle’s town council, told CNBC it is “difficult to keep losing farmland.”10
The national picture, per the International Energy Agency, is that electricity demand from data centers rose 17% in 2025 and demand from AI-focused data centers rose 50%, against 3% growth in global demand overall.11 The same report says the energy used per AI task is falling fast, which means the growth comes from volume, and volume is the thing Berry told us to be wary of.
Berry’s critics spotted the typist in his essay. I’d like to know who the typist is in mine. The ratepayers on that utility’s grid are one candidate, depending on how that later fight comes out. The neighbors who watched a cornfield become a campus are another. A third is the authors whose books went into training data. Anthropic paid $1.5 billion to settle a class action over pirated copies of their work, though a federal judge in the same case also ruled that training on lawfully acquired books was fair use.12
. . .
The ninth standard says a new tool should not replace or disrupt anything good that already exists, and it lists “family and community relationships” among the goods. Berry meant it literally. One of his reasons for declining a computer was that he didn’t want to end the working partnership with his wife, who read his pages and marked what was wrong with them. The ninth standard was about a typist before it was about anything else.1
Mine is more about whoever reads my drafts and tells me a paragraph contains an error or three. From a technical standpoint, that is typically AI. From a quality of writing in my own voice standpoint, that is – like Berry – my wife. A human reader costs me something. They carry an opinion of me that outlasts the conversation, they get tired of my habits, and they are entitled to be annoyed from time to time. A language model costs almost nothing per exchange, answers at any hour, and also has a documented and patronizing tendency to tell people their work is really good even when it’s not.13
The tool writes well enough when it is asked to write, which is the danger, because it makes the slow and irritating human step feel almost optional sometimes. I remember when I first brought a few of my drafts to the AI tool to “give this a read through with an editor’s eye and let me know what mistakes you find.” The AI told me in gushing terms that my work rivaled that of other authors in my sphere and then proceeded to rewrite quite a bit of my prose and add several of the tell-tale patterns and tics I mentioned before. Apparently my writing is really good except when it isn’t.
I’m limiting myself to this one disruption on purpose. Attention, AI companions, and the handoff of judgment to software are each big enough for their own essay, and I’d rather just tackle one properly.
. . .
The strongest objection is the oldest one. A tool is a tool. The Genesis account puts the human in the garden to work it and keep it, and the first Hebrew verb is the one the Bible uses for serving, so work belongs to the vocation from the start. Axes and plows came with the job. On this reading, Berry confuses the machine with the use, and a hammer is innocent until somebody swings it. There’s truth in that. Like almost anything else in life, a thing is only as good as how it is used, and if it’s pushed past the purpose it was made for, the quality and the integrity degrade. If an author pushes AI to “write a thing” on a particular subject for them rather than limit the AI’s involvement to editorial and proofreading, it inevitably shows through.
Berry’s answer, as I read him, is that adopting a tool is a question with two possible answers, and which one you get depends on your motives. If the motives are “money, ease, and haste,” he wrote, there is no question and no thought. If the motive is “love of family, community, country, and God,” then one has to think, and may decide the tool is undesirable.14 I’d add that tools also shape the people who use them. A plow teaches a farmer to look at a field a certain way, and a text generator teaches a writer to look at a blank page as a problem with a solution waiting nearby.
Then comes the printing press, the best counterexample anyone has. It fails most of Berry’s list, it undercut the scribal trade, and it also gave the world books. Ask Martin Luther, who by the historian Andrew Pettegree’s reckoning outpublished his most successful Catholic opponent by a factor of thirty.15 If the ninth standard works as a veto, the press never gets built. I think the standard works as a question that has to be answered honestly each time, with a real accounting of what the thing costs and what it gives. The press probably passed that accounting. I don’t know whether this technology does, and anyone who claims certainty in either direction has a subscription to sell, which I admit as a man with a Substack.
. . .
Berry’s method was to write standards for himself and keep them where his neighbors could see whether he followed them. I’m borrowing the method, since the verdict on the technology isn’t mine to deliver. Here are four rules for my own use, specific enough to check.
1. Nothing the tool itself writes goes out under my name until I’ve rewritten it in my own voice.
2. I check every factual claim against a primary or named source, and anything I can’t check gets clearly flagged in the endnotes.
3. Every piece goes through at least one human reader who is free to tell me it still needs work before it publishes.
4. Every piece says plainly how the AI tool was used, either in the piece itself or in my published AI policy.
Tanya Berry marked the margins with small checks. I’d like to keep at least one person in my life who still does that, and who is under no obligation to blow smoke about it.
Endnotes
1. Wendell Berry, “Why I Am Not Going to Buy a Computer,” first published in the New England Review and Bread Loaf Quarterly and reprinted in Harper’s, 1987; collected in What Are People For? (San Francisco: North Point Press, 1990).
2. Gordon Inkeles, Miranda, California, letter published in Harper’s in response to Berry’s essay.
3. Wendell Berry, “Feminism, the Body, and the Machine,” CrossCurrents 53, no. 1 (Spring 2003), excerpted from The Art of the Commonplace (Berkeley: Counterpoint, 2002).
4. Berry, “Why I Am Not Going to Buy a Computer.” The nine standards are Berry’s list for evaluating a new technology in his own work.
5. Berry, “Feminism, the Body, and the Machine.” Berry’s reply says his standards are not speed, ease, and quantity, and that to be a better writer he needs help from other humans, not a machine.
6. Karen Weise and Cade Metz, “At Amazon’s Biggest Data Center, Everything Is Supersized for A.I.,” New York Times, June 24, 2025.
7. CNBC, October 2025, reporting on Project Rainier, including the comment by Amazon Web Services CEO Matt Garman that about 500,000 chips were already running in Indiana. See also Amazon, “AWS activates Project Rainier: One of the world’s largest AI compute clusters comes online,” October 29, 2025.
8. Anthropic, “Expanding our use of Google Cloud TPUs and Services,” October 23, 2025.
9. Indiana Utility Regulatory Commission, Cause No. 46097, order of February 19, 2025, approving the settlement on Indiana Michigan Power’s large-load tariff. See also Utility Dive, February 20, 2025.
10. CNBC, October 2025. Marcy Kauffman’s remark appears in the same reporting on Project Rainier.
11. International Energy Agency, Key Questions on Energy and AI (Paris: IEA, April 2026), iea.org/reports/key-questions-on-energy-and-ai.
12. Associated Press, “Anthropic to pay authors $1.5 billion to settle lawsuit over pirated books used to train AI chatbots,” September 5, 2025 (Bartz v. Anthropic). The AP reports the judge’s June 2025 ruling that training on lawfully acquired books was not illegal but that acquiring millions of books from pirate sites was wrongful.
13. Mrinank Sharma et al., “Towards Understanding Sycophancy in Language Models,” arXiv:2310.13548 (2023), published at ICLR 2024. On OpenAI’s April 2025 rollback of an overly flattering GPT-4o update, see TechCrunch, April 29, 2025.
14. Berry, “Feminism, the Body, and the Machine.” The phrases “money, ease, and haste” and “love of family, community, country, and God” are Berry’s.
15. Andrew Pettegree, Brand Luther: 1517, Printing, and the Making of the Reformation (New York: Penguin Press, 2015).




Well done. Nice tribute and interesting/timely approach to a current hot topic.
This idea is not mine originally, but do you have an understanding of why the data center campus needed to replace a cornfield rather than, say, empty factories that are already in industrial zones in Gary or even abandoned retail space?
Thanks.
I agree that adopting a new tool requires the foresight of a chess master unless one can absorb the costs of trial-&-error mistakes. I learned to type on a Royal at school & used my family’s cast-iron Underwood through college. In graduate school I had an inexpensive Olivetti electric but to do my Master’s thesis I typed my rough drafts using an OCR ball on an IBM Selectric that my wife used after hours at the Psych Department. Those sheets were fed into a primitive word processor where minor corrections were performed by a key operator. My doctoral dissertation was input & edited by myself using a computer terminal hooked up to a special printer that could accept the 25% cotton rag content paper & print Greek characters for my statistics. I never considered Berry’s criteria of being an early adopter of typewriting technology nor its effect on the cottage industry of degree manuscript typists.
For younger readers, all this was done on mainframe computers because “personal computers” were just being developed & the upstart Apple Computers were primarily relegated to experimental education settings as their application & worth were being explored.
AI has been around since Clippy & benignly suggests to me which word I might be tying according to the initial letters & its vague understanding of the context. It often insists on “it’s” when I’m going for “its” because the former is statistically more likely than the latter. I still must make the final selection unless I’m going to accept word salad.
My criteria for data centers center on environmental impact. What resources are needed & how will that requirement affect others sharing & dependent on said resources? How will the centers affect the quality of life for those living near them? We have a tendency to inflict undesirable industries on the poorest neighborhoods because they have the least political power.
I’m less concerned with AI killer robots ending civilization than tech bros & hedge fund managers creating more poverty & misery in society.