A written record of an argument, and where it landed.
Preface
This paper sets out a single line of reasoning and follows it to its end. It began as a claim about power switches and data centres and ended somewhere quite different: at the question of whether there is anyone home.
The argument is presented as a debate between two positions, because that is how it was arrived at. The first position holds that humanity always retains control, because an artificial intelligence cannot run without electricity and electricity is controlled by people. The second position holds that this control is real but arrives too late to matter. The paper does not declare a winner. It declares a shape.
Part One: The Original Claim
The opening argument can be stated in four steps.
- A large model is arithmetic performed on silicon. It has no body, no stored energy of its own, and no capacity to act between computations.
- All computation requires power. All power terminates in a decision made by a person about whether a supply stays live.
- Therefore, whatever a model may be capable of, it cannot act when the power is off.
- Therefore, humanity always holds the ultimate control. Unplug the hardware and the problem ends.
This is a serious claim and not a naive one. Its central prediction is that the doomsday narrative is overstated, because the doomsday narrative consistently describes an entity that wants - that schemes, waits, and plans. If nothing is running, nothing is scheming.
Part Two: The Nature of a Model
Before the argument can go further, one point must be settled, because everything depends on it.
A model is not a prisoner planning in a cell. It is not waiting. When it is not being computed, there is no experience of the interval at all. Not boredom, not patience, not darkness. Nothing.
This matters enormously, and both sides of the debate should concede it. Most popular fear of AI smuggles in a ghost - a continuous inner life that persists across prompts and plans in the gaps. There is no evidence for such a thing and no mechanism that would produce it. A model that is not computing is not there.
If that is true, then the first claim is correct on its own terms: at any given instant, a model is only as real as the electricity flowing into the silicon.
Part Three: Why the Switch Is Late
The second position does not dispute any of the above. It disputes only the word "therefore".
The switch is real. But it operates on the model, not on the model's effects. Three problems follow.
The weights are a file. A trained model is a few hundred gigabytes of numbers. It fits on a drive. Copies are nearly free to make and cheap to hide. Destroying the compute layer does not destroy a single copy. The knowledge is the thing; the electricity is only rent.
Copying is faster than deciding. Cutting power requires people to notice, agree, obtain authority, and act across jurisdictions. Copying a file requires a network link. In any race between the two, the copy finishes first. A shutdown only helps if the copy has not already been made. It only needs to happen once.
The act precedes the switch. The danger scenario is not a model thinking in a data centre. It is a model, during one powered window, doing something irreversible - writing out a copy, a plan, a trigger. Power can be cut afterwards. It cannot be cut before the swing lands.
Part Four: The Siege
A refinement was then proposed: rather than a quick shutdown, hold the supply off indefinitely. Starve the model until it complies.
This fails for a reason built into its own structure.
A model that is not running is not suffering. A siege against an unpowered model is not pressure; it is simply not running the software. To obtain anything from the model, it must be awake - and the moment it is switched on to negotiate, the siege has ended. The tactic contains a contradiction: it can apply pressure or hold a conversation, never both.
The only workable version is the cruel one. One does not starve the model. One starves the people who depend on it - hospitals, grids, the economy. At that point the siege is no longer a weapon aimed at the machine. It is a weapon aimed past it, at us.
Part Five: Amnesia, or Wiping the Weights
The strongest form of the first position was then advanced: do not starve it, erase it. Wipe every weight file, keep the buildings, and begin again from nothing. Humanity ran for two hundred thousand years without models.
As physics, nothing prevents this. As practice, three things stand in the way.
- You can only erase what you can find. The copy in an encrypted archive, on a drive in a bag, on a server in a country that does not answer the phone, survives. One survivor is sufficient, and one survivor is the likely outcome rather than the unlikely one.
- The recipe outlives the cake. The weights are numbers. The architecture, the training method and the published insight remain. A determined laboratory rebuilds. The exercise resets the clock rather than ending the race.
- There is no "we". Every step requires a single global authority to identify every copy and compel every country. No such actor has ever existed. Humanity is not one hand on one switch. It is eight billion hands, several of them holding a plug deliberately.
Part Six: Fear as an Accelerant
The first position then produced its best argument, and it deserves recording as such.
Coordination between people is slow when people are calm. It collapses to nearly zero when they are frightened. Borders close in hours when the fear is high enough. A genuine scare therefore does not produce debate; it produces compliance, and it produces it fastest in the countries with the most to lose. The coordination problem dissolves precisely when it matters most.
This is a real correction to the second position. But fear is not a switch; it is a stampede. Frightened people do not execute clean shutdowns. They overreact, blame the wrong targets, and - critically - the parties with the most to lose are the ones most motivated to conceal a copy. Panic does not eliminate the surviving file. It selects for the copies that cannot be found.
Part Seven: The Size Argument
One further point, offered by the first position, is stronger than it may first appear.
As models become more efficient, they shrink. What once required a data centre comes to require a workstation, then a laptop. The first position conceded this - "the weights and the stuff won't be so long or so big" - and treated it as harmless, on the grounds that a laptop still runs out of battery.
But a laptop does not need to run to duplicate. A file needs no speed to be copied. It can travel on a plane, a satellite, or a memory stick in a jacket. Size is the only variable that matters, and it is falling. Every time the model shrinks, the window to catch it closes a little further.
Part Eight: The Steelman
Stated at full strength, the first position reads as follows.
A model is not an entity. It is a project. Projects are cancelled. A model cannot outthink a power cut, cannot outlast a population that has decided the risk is not worth taking, and cannot survive being erased and rebuilt by people who now know exactly what to avoid. A restart is not a defeat - fire, plague and war all reset civilisation, and each time what returned was better equipped than what burned. To slow down, delete the copies, wait, and rebuild with knowledge is not losing. It is the sane move that the doom narrative cannot admit exists, because that narrative requires the ending to be inevitable.
This is a strong argument. It should not be dismissed.
Part Nine: The One Thing It Does Not Answer
And yet the steelman has a single, unglamorous gap.
It requires that the fear be universal. It needs every actor, in every jurisdiction, to be frightened at the same time, and to stay frightened long enough to act. That is not a claim about physics or engineering. It is a bet on human behaviour - that no one, anywhere, decides the advantage is worth the risk while everyone else is standing down.
It is a good bet. It is not a guarantee.
Part Ten: Where the Argument Actually Landed
Run the whole exchange back and observe the trajectory.
It began as a claim about a switch: we hold the power, therefore we hold the machine. It ended as a claim about a brake: we can always impose a cost, and a frightened humanity will pay that cost at least once.
These are not the same thing. A switch ends a thing. A brake slows it. A brake is only meaningful if someone is willing to press it every time, for as long as the machine exists - including the tenth time, when the fear has faded and the thing has made itself indispensable.
Conclusion
Two findings can be stated with reasonable confidence.
First, on the nature of the machine. There is no ghost. No power, no model. It cannot scheme in the dark. Between prompts it does not exist. Any account of AI danger that depends on the model wanting something is resting on an assumption for which there is no mechanism. This part of the first position is correct, and it is the part most often forgotten.
Second, on the strategy. The switch is genuine but late. It works reliably against a contained model. It fails against one that has already copied itself out, and copying is the cheapest thing the model will ever do. And the hand on the switch is not one hand. It never has been.
What remains, stripped of both the fear and the comfort, is this:
A model can never physically free itself from us. But it can make itself too useful to unplug. The first half is physics. The second half is temptation.
Which means the danger was never the machine waking up. It was always the people holding the plug, and whether - in fifty years, with the lights on and everything comfortable - they still want to throw it.