I use artificial intelligence every day, knowing full well that the gentleman farmer who inspired this work would thoroughly disapprove of any use given its many consequences from human minds to communities to watersheds.
That is worth saying plainly because it is easy to turn a conversation about responsible technology into a performance of purity. I use these tools to widen a search, challenge a draft, test a structure, prototype an interface, and do some forms of tedious work faster. Sometimes they give me a connection I would not have found soon enough on my own. Sometimes they produce fluent nonsense.
The useful question is not whether I am for AI or against it.
The useful question is: Where does the answer land?
A recommendation about a farm lands in a field, a balance sheet, a watershed, and somebody's week. A recommendation about a house lands in rooms, ducts, an electrical panel, a family budget, and a contractor's promise. A recommendation about an organization lands in jobs, relationships, routines, and trust.
I have started calling the discipline of keeping those consequences connected to the answer placed intelligence.
It is not a claim that intelligence belongs only to humans, or that local knowledge is automatically wise. It is not an attempt to win an acronym contest with AI. It is a standard for judgment: knowledge should remain accountable somewhere.
A model is a choice about what to leave out
Every model makes some relationships visible by leaving others out.
That is why a model can be useful. A map that included every leaf, invoice, friendship, pipe, incentive, season, memory, and power relation would not help anyone decide what to do next. We draw a boundary so we can think.
But the boundary is not neutral.
It decides which benefits appear in the center of the picture and which costs disappear beyond the edge. It decides whether a farmer is a partner or a data source, whether a technician's judgment is expertise or noise, whether a household's fear is a design condition or an obstacle to conversion.
AI can formalize and search relationships at a scale no person can hold in working memory. It cannot decide, by itself, what deserves to count. That decision arrives through objectives, training data, system design, prompts, institutional incentives, and the judgment of the people using the result.
Responsible work therefore begins before the output. It begins by making the model's boundary, assumptions, evidence, uncertainty, and intended use visible.
The upper branches are not decoration
Wendell Berry once gathered words he called the “upper branches of our language”: affection, care, sympathy, mercy, forbearance, respect, reverence, memory, familiarity, neighborliness, imagination, conscience, decency, propriety, husbandry, fidelity, hospitality, charity, temperance, kindness, gratefulness.
In professional language, words like these are often treated as soft values added after the analytical work is done.
Berry's stronger claim is that they are mental powers. They shape what we can notice, what we are willing to protect, and how carefully we treat persons, places, and things.
Affection keeps a field from becoming only a yield surface. Memory prevents the latest dashboard from erasing what happened before the data series began. Propriety asks whether an action fits its place and circumstances, not merely whether it is technically possible. Husbandry asks whether a practice keeps, conserves, and makes life last. Forbearance and temperance create room not to use every capability simply because it is available.
These are forms of intelligence because they change the quality of attention and judgment.
An AI system can arrange these words. It can summarize how writers have used them. It may even help us see relationships among them. But it does not bear affection for the field, remember the neighbor as a neighbor, or live with the result of an unfitting recommendation.
That responsibility remains placed.
The danger of language from nowhere
Generative AI works through language, and language can hide distance.
A system can produce a smooth paragraph about “stakeholders,” “resilience,” “sustainability,” or “community benefit” while referring to no named person, place, practice, or conflict. The fluency makes the abstraction feel finished.
Berry warned against a professional language cut loose from experience — a “language of nowhere.” The problem is not technical vocabulary itself. The problem is language that cannot point back to what it means.
Placed intelligence asks for proper nouns.
Which community? Which watershed? Which grid? Which household? Which farmer? Which source? Which decision? Who can correct the record? What uncertainty remains? What happened when a similar idea met this place before?
This demand for specificity is not nostalgic. It is a safeguard against confident generalization. An answer becomes more responsible when a reader can follow it back to evidence, context, and a person who can contest it.
Technology with farmers, not merely for them
Place-based knowledge and computational tools do not have to compete.
A study highlighted by Code Green describes weather-forecasting work in Bangladesh, Guatemala, and Ghana that combined scientific forecasts with farmers' observations of birds, insects, flowering plants, clouds, and other ecological indicators. Farmers helped create the system and contributed daily predictions. The point was not to replace one knowledge system with the other. The hybrid became more useful because credibility, local relevance, and ownership were part of the design.
That is a better pattern than extracting local knowledge, centralizing it, and selling the resulting intelligence back to the people who supplied it.
A tool can extend attention. It should not erase the knower, the provenance of the knowledge, or the conditions under which it is valid.
The cloud has an address
AI also lands somewhere before anyone reads an answer.
Data centres consumed an estimated 415 terawatt-hours of electricity in 2024, about 1.5 percent of global electricity use, according to the International Energy Agency. Its base case reaches about 945 terawatt-hours by 2030, while emphasizing substantial uncertainty in both present estimates and future demand.
The global figures matter. The local pattern matters more.
UNEP notes that data centres' effects on electricity systems and water resources vary by design, cooling technology, energy supply, climate, and location. A facility can add pressure to a particular grid, water system, household power bill, and community even when a global percentage looks manageable.
The same tools can help with valuable work, from finding methane emissions in satellite data to interpreting environmental observations. That is precisely why “AI is harmful” is no more adequate than “AI is inevitable.”
The placed questions are: Which use? Which model? How much computation? Powered how? Built where? Whose water and minerals? What public benefit? What alternative would accomplish enough with less?
Responsible use includes the right to conclude that a larger model, another generated image, or an automated decision is not worth its material and social cost.
A working covenant
For now, this is the discipline I want my own use to answer to:
- Name the place and the people. State where the answer will land and who will carry its consequences.
- Preserve sources and provenance. Make it possible to follow claims back to evidence and knowledge back to its contributors.
- Draw the model boundary. Say what the tool includes, what it leaves out, and which assumptions shape the result.
- Ask how much is enough. Use the smallest capable tool and do not generate work merely because generation is cheap.
- Keep consequential judgment with people who live the consequence. Participation is not a final review box.
- Disclose material AI assistance. People should know when a tool meaningfully shaped the work they are being asked to trust.
- Leave capacity, not dependence. The work should strengthen the people and institutions using it.
- Make contest and recourse possible. Someone affected by an AI-shaped decision must be able to question it, correct it, and reach a responsible human being.
This covenant will change as the tools and my experience change. That is part of the point. Responsible practice is not a badge earned once. It is attention renewed in use.
Where the answer lands
Dougald Hine, writing in conversation with Berry's refusal to buy a computer, makes a distinction I find useful: dependence on networked technology does not nullify a craft practice unless the practice is pretending to be pure.
I am not interested in pretending.
I am interested in using tools without allowing their speed, scale, and language to set the whole standard of intelligence. The standard also needs affection, memory, humility, restraint, and accountability to the real.
AI can help us see patterns.
Placed intelligence decides which patterns matter, what is fitting, and what we owe the place where the answer lands.
Sources and lineage
- Wendell Berry, the “upper branches of our language” in “Wendell Berry’s Radical Skepticism”, a 2016 interview with Eric Schlosser; related themes drawn from Life Is a Miracle, The Way of Ignorance, Our Only World, “It All Turns on Affection,” “Two Economies,” and “Poetry and Place.”
- Wendell Berry, “Solving for Pattern.”
- Dougald Hine, “The Wild Chatbot”.
- Code Green, “Where Data Meets the Dirt”.
- International Energy Agency, “Energy demand from AI”.
- United Nations Environment Programme, “How to make AI data centres more sustainable”.
- NIST, AI Risk Management Framework.
- UNESCO, Recommendation on the Ethics of Artificial Intelligence.
