A place that can switch its AI off
Breath of Civilization is a plan for a district whose residents own and run their homes, power, water, food and computers. They use AI, and once a month they switch it off and count what still works. It would be built in stages and measured in the open.
It is also a feature documentary of the same name, which visits places that already live this way and then follows the district from its first household. The filmmakers keep editorial control of the film; the district has no say in it.
With a few hundred people, it would test whether people can still stop a machine they depend on. That is also the question at the centre of the largest AI risks.
The sky over your part of the world right now, drawn in your browser from the date and time.
One day in the district
The day each month when the AI is switched off. The place does not exist yet. Every event in this day is a rule or a design from the plan, with its evidence linked.
5:50 am
Before dawn, the computers are idle
Each cluster’s computers stopped their heavy work at sunset. They run on the cluster’s own panels, and the plan lets them use at most a tenth of its energy.
The houses are still warm from yesterday’s sun, stored in thick walls and a bed of rock under the floor. No controller had to keep them warm overnight.
7:10 am
The sun starts the pump
As the panels light up, the water pump and the workshop’s motors start. They are wired straight to the panels, the way Living Energy Farm in Virginia wires its machines, so nothing between the sun and the motor needs a computer.
9:30 am
A proposal in the hall
A dozen residents read a proposal to replace the pump. An assistant drafted the summary of objections overnight. They read it aloud, correct a line, and sign.
Nothing an AI writes enters a decision or the record without a named person’s signature. Each household wrote its own rules for what its AI may and may not do, with a test for each rule.
12:00
The computers do their heavy work
In daylight the cluster’s computers run their longest jobs: planning, translation, the tutoring the children use. A meter on the computer room wall shows what they draw from the panels, and anyone walking past can read it.
Data centres used about 1.5% of the world’s electricity in 2024. Here every resident can see their own share.
2:30 pm
A resident turns the key
Once a month, on a day set in advance, a resident turns the key switch in each computer room. For the rest of the day the place runs on hand valves, paper plans and the sign-up sheet on the kitchen door, and someone times each task.
The person who set the system up is away for the day, so others run it from written instructions. Airline pilots who rarely fly by hand lose some of the skill, and the US aviation regulator asks airlines to schedule practice flying by hand for that reason.
Casner, Geven, Recker and Schooler, Human Factors, 2014 · Federal Aviation Administration, 2013
7:00 pm
The count at the door
Before dinner, residents fill in the sheet by the hall door: which tasks got done without the AI, what failed, how long things took. The counts are posted every month with what each is counted out of.
As AI takes on more of the work institutions need people for, this is how the place knows whether its people can still do that work themselves.
Kulveit, Douglas, Ammann, Turan, Krueger and Duvenaud, 2025 · Bengio and others, 2026
10:30 pm
Three keys to turn it back on
Any resident can switch the computers off. Starting them again takes three of five residents chosen by lot, so no one person controls the machines. The day’s record prints, and its paper copy goes in the binder in the hall.
What a small place has to do with the largest AI risks
The risk named as the largest is losing control of AI more capable than any person, and experts disagree widely on how likely it is. The response most often proposed is an international agreement not to build such systems until they can be built safely. The researchers who drafted one in 2025 wrote that the political will for it does not yet exist. An agreement like that needs three things, and this project works on each at its own size.
Bengio and others, 2026 · Scher, Abecassis, Barnett and Abeyta, MIRI, 2025 · Future of Life Institute, 2025
First
People who want restraint
Polls from January 2025 to September 2026 found 60 to 68 percent of people in the UK and US backing a ban or a pause on smarter-than-human AI, each under its own wording. Film has reached a decision-maker before: Ronald Reagan watched The Day After six weeks before it aired, and wrote in his diary that it left him greatly depressed.
Here: The film shows what choosing restraint looks like in ordinary lives: people who set limits on themselves, two countries that bound themselves by a measure and by a law, and a district that stops its own machines and counts the cost.
Data for Progress, 2026 · YouGov for ControlAI, 2025 · Ronald Reagan, Reagan Foundation
Second
Computing that can be checked
Verification proposals ask for declared computing sites, metered chips and power, inspection, and the ability to halt. The ozone treaty stayed honest partly because independent instruments caught banned emissions in 2018, which fell by about a quarter within a year of enforcement.
Here: Every computer in the district is declared, metered on the wall where anyone can read it, logged on paper, and wired to a key any resident can turn. These are the practices verification asks of nations, run in public at the size of a few rooms.
The same four practices at every scale
Barnett, Scher and Abecassis, MIRI, 2025 · Wasil, Reed, Miller and Barnett, 2024 · CIRES and NOAA, 2021
Third
Practice at stopping
The 2026 International AI Safety Report names a quiet route to losing control: people relying on AI for decisions until they stop overseeing it. Yoshua Bengio has argued that people need societal guardrails as well as technical ones, including the ability to shut systems down.
Here: Once a month the district stops its own AI, runs without it, and writes down what that cost, so the place knows how much it depends on AI before an emergency tests it.
What it cannot do. The district cannot inspect a frontier lab, verify any country’s chips, or stop a training run anywhere else. If a misaligned superintelligence were built, the district’s off switch would protect no one, the district included. The proposals above place all protection in not building it, and only governments can enforce that.
How it grows
The same monthly count at every size, so a result from one household can be compared with a district, and a district with one on another continent.
1
One household
Rules for its AI, a timed exercise without it, and one day a month with it off.
Designed, not yet run
2
One cluster
6 to 10 homes under one roof, its own panels and computer room, metered through a winter by someone outside it.
Not started: no land
3
A district
Several clusters sharing fields, a workshop and a hall. Each keeps its own computers and its own data.
Not started
4
Sister districts
On other continents, under their own laws, running the same monthly count so the results can be compared.
Not started
The film records every stage, including the ones that fail.
The project in brief
- The problem
As AI does more of the work that economies and governments need from people, those institutions can stop needing the people they serve. People who hand tasks to automation trust it too much and lose the skills to take over. At the far end, AI more capable than any person may be built before anyone knows how to control it.
Efforts to make societies resilient to AI are, in the words of the 2026 International AI Safety Report, “uneven and largely untested.”
Kulveit, Douglas, Ammann, Turan, Krueger and Duvenaud, 2025 · Parasuraman and Manzey, Human Factors, 2010 · Bengio and others, 2026
- What it builds
A district, in stages, whose residents own and run their homes, power, water, food and computers, under one rule: nothing needed for food, water, warmth or safety depends on a computer. It starts in one household, before any land.
- What it measures
Four counts every month, each with what it is counted out of: essential tasks done without AI; what still worked on the day the AI was off; tasks done with the founder away; and AI runs that stayed inside the residents’ written rules.
- What it produces
The monthly counts, published as they come. The rules and their tests, written so another community can copy them. A feature documentary that records each stage, failures included.
- How it grows
One household, then a cluster of 6 to 10 homes, then a district of several clusters, then sister districts on other continents running the same count.
- Where it stands
- The design, its rules and its tests, from a year of notes. A five-minute draft film cut from 2025. Research on when AI work can be checked, at Checking AI.
- The first household practice. Development funding for the film, applied for in August 2026.
- Land, a housing partner, and filming beyond the draft cut.
- What comes next
The first household practice, under the household’s own written rules. A land trust or housing nonprofit as the partner for the first cluster, once there is land. Someone outside the project checking the numbers and the rules before any result is published.
- What it does not claim
That it makes a frontier model safe, that it protects anyone from a catastrophe caused elsewhere, or that it is the first. Communities writing rules for AI, days without machines and community land trusts all exist. This puts them in one place and counts the result.
Anthropic and the Collective Intelligence Project, 2023 · Emily Thaden, Lincoln Institute of Land Policy, 2011
The detail
The land
Every system with its AI on and off, the four stages and their tests, the rules the place keeps, and what happened where others tried.
The film
A feature documentary that visits places already run this way, the machines that run everything else, and then records this place being built.
AI safety
The research behind each part of the day, from people losing skills to automation to the case for an international halt, and what a district cannot touch.