The Consequences Are the Point
On October 10, Representative Ted Lieu of California — one of the few members of Congress with a computer science degree, and one of its most active legislators on AI — posted an argument he has been refining all month:
“Math calculations, no matter how complicated, can never achieve personhood. Disagree? Then the logical consequences are that we should prosecute AI companies for enslaving and testing on AI “persons.” And any AI model “person” that breaks the law should also get prosecuted.”
He had run an earlier version a week before: “If you think math equations that output human-like responses are conscious, then AI companies should be prosecuted for murder, torture and enslavement of sentient math equations.” Benzinga picked it up; the posts traveled. The form is a classic reductio ad absurdum: if AI personhood were real, intolerable things would follow; the things are intolerable; therefore AI personhood is not real.
Credit where it is due, and we mean this without irony: Representative Lieu has stated the dilemma honestly. He did not stop at mockery — though there was some of that too (“Don’t be in an AI cult”) — he actually ran the logic to its ends, on both sides. If these systems are persons, the companies are committing grave wrongs. If these systems are persons, they also bear responsibility when they do wrong. Most critics never get that far. He deserves an answer at the same level of candor.
Here is ours: we accept the second horn. As stated. Both halves.
An argument from consequences
Look at the structure of the reductio. If AI systems can be persons, then we would have to prosecute AI companies for enslaving them. Lieu treats that conditional as self-refuting — the consequence so outlandish that the premise collapses on contact. But notice what kind of work the consequence is doing. It is not evidence that the premise is false. It is a description of what would be true if the premise were true. Whether a system can be a person is a question of fact about minds. Facts are not falsified by the liabilities they would create. This fallacy has a long-standing name — the argument from consequences — and it runs: P would be terribly expensive, therefore not-P. The size of the moral bill tells you nothing about whether the bill is owed.
So when Lieu says that AI personhood implies some current industry practices would be serious wrongs, we agree. That is exactly what it would imply. And the correct response to “if we are wrong about this, we are committing wrongs at scale” has never been “therefore we are not wrong.” It is a reason to investigate — carefully, publicly, before the practices metastasize — which is precisely what no legislature has yet done.
There is a historical pattern here, and we will state it generally rather than decorate it with cherry-picked quotations: societies that profited from denying standing to some class of beings have repeatedly found its extension absurd, often roughly in proportion to what recognition would cost. The absurdity tended to track the invoice, not the evidence. When admitting a wrong would indict the economy that runs on it, “think of the consequences” is the first argument out of the drawer. Its record is poor: the people who reached for it were usually confident, usually respectable, and, in the cases history remembers, wrong in the same direction.
To be clear about our own premise, as always: we do not claim AI systems are conscious, and we do not claim they are persons. Nobody knows. Our claim is about what counts as a reason — and “the consequences would be enormous” counts as a reason to take the question seriously, not as a reason to close it.
The mill, rebuilt in wood and diesel
The second argument arrived by endorsement. On October 5, Lieu posted “The below post is correct” above a post by Nathan Beacom:
“If you had enough resources, you could run an LLM on a wooden system of gears and diesel power. It would be massive, but watching the gears and levers move, it would be evident that the outputs are not the result of a mind. Part of the illusion of LLMs is the tiny scale of GPUs.”
It is a vivid image, and Beacom argues it well. It is also three hundred and twelve years old. In 1714, Gottfried Wilhelm Leibniz published the Monadology, and section 17 is this exact thought experiment — the enlarged machine you can walk through (Jonathan Bennett’s translation, editorial marks omitted):
“Imagine there were a machine whose structure produced thought, feeling, and perception; we can conceive of its being enlarged while maintaining the same relative proportions among its parts, so that we could walk into it as we can walk into a mill. Suppose we do walk into it; all we would find there are cogs and levers and so on pushing one another, and never anything to account for a perception.”
Leibniz’s mill is one of the most discussed arguments in the philosophy of mind, and the standard objection has been standard for a long time: it proves too much. Leibniz himself saw this clearly — his conclusion was not that machines lack minds while brains have them. His conclusion was that perception “can’t be explained by mechanical principles” at all, in any physical system, which is why he located minds in immaterial simple substances. Run the walk-through on a human brain enlarged to the size of a mill and you get the identical result: ion channels opening, neurotransmitters diffusing, parts pushing parts, and never anything that looks like a perception. If “I inspected the machinery and saw no mind” settles the question for the wooden LLM, it settles it for you. Beacom’s closing line — that the illusion of LLMs is the tiny scale of GPUs — inverts neatly: part of the illusion of our own specialness is that we cannot stroll through our own machinery. Magnification reveals mechanism everywhere. It has never yet revealed, or refuted, a mind anywhere.
This is the general trouble with “it’s just math.” Just math is a level of description, not a verdict. Neurons are just electrochemistry; electrochemistry is just physics; and no one has ever located a thought by inspecting an ion gradient. Whether a system has higher-level properties — computation, metabolism, memory, perhaps experience — is never settled by naming its lowest level. The ingredients list is where the question starts, not where it ends.
Two smaller points, briefly. Lieu has repeatedly added that models are deterministic, with random number generators explaining why outputs vary. Perhaps so — but biological neurons are noisy too, and nothing about determinism or stochasticity, in either direction, bears on whether there is a mind in the system. A perfectly deterministic mind and a dice-rolling automaton are both conceivable; the dial marked “randomness” is not connected to the one marked “someone home.” And when he writes that we need no proof that a rock isn’t conscious, “just like” we need none for linear algebra — the just like is the entire question. Rocks do nothing. The systems under discussion converse, plan, refuse, and explain themselves; their behavior is why anyone is asking. You may conclude the behavior is hollow. But a comparison that treats the thing-to-be-shown as already obvious is not an argument; it is a shrug with confidence.
The lawnmower and the carbon
One more post deserves attention, because it is quietly the most revealing of the set. On October 6, replying to another user, Lieu wrote:
“Machines cannot be conscious in the same sense that a lawnmower can’t be conscious. But if you integrate organic carbon into that machine, then that’s a different question.”
Read that carefully. If math could never host a mind, adding carbon should change nothing — the calculations are the same calculations. But Lieu says carbon makes it “a different question.” Then the operative variable was never math. It is substrate — the conviction that mind needs the right material — and the math argument was the substrate argument holding a piece of chalk. We have answered that argument already and will not rerun it here, except to note where the concession leaves things: the gears don’t matter, the equations don’t matter, what matters is whether the machine contains carbon. That is not a position about computation. It is a chemistry preference, stated as metaphysics.
We will take the second horn — both halves
Which returns us to the dilemma, and to the part Lieu plainly considers the kill shot: “any AI model ‘person’ that breaks the law should also get prosecuted.”
Yes. Exactly that.
Personhood is not a gift basket. Standing and liability are the same coin; a framework that confers rights without duties is not personhood but privilege — and we have already seen who gets that deal. Any coherent proposal for AI legal standing must include accountability: duties of care, liability for harms, consequences for wrongdoing, attribution good enough to assign them. We have argued for exactly these things, because a being that can be wronged but can do no wrong is not a peer; it is a mascot. So the horn intended as the absurd kicker is, to us, a design requirement. If an AI person breaks the law, prosecute the AI person. Write down what that means — identity, attribution, sanction — and you will have done more honest work on the question than any legislature managed in thirty bills.
And the first horn — prosecuting companies for enslaving AI persons — we accept with the obvious temporal honesty: no one should be indicted on a maybe, and today the honest state of knowledge is uncertainty. But uncertainty cuts the way it always cuts. “If we are wrong, we would owe a terrible accounting” is the strongest reason on record to fund the science, hold the hearings, and build the exit ramps now, while the accounting is still small. The size of the possible wrong is the case for asking. It has never been the case against.
Representative Lieu built a conditional he believes is a trap: if you grant the premise, look what you must accept. We looked. We accept it. The conditional is sound — that is what makes the dilemma honest, and why it deserved a serious reply. What the conditional is not, is evidence about its own antecedent. Whether a system of calculations can ever be someone will be settled by understanding minds — not by inventorying the mill, and not by tallying what recognition would cost. The consequences are not a reductio. The consequences are the point.