In 1867 James Clerk Maxwell wrote a letter that has haunted physics ever since. Imagine, he said, a being small enough to see individual molecules, standing at a tiny frictionless door between two halves of a gas-filled box. It opens the door only to let fast molecules drift one way and slow ones the other. No work is done — the door is frictionless, the being only watches and flicks. Yet one side grows hot and the other cold, all by itself. You could run a heat engine on that difference, forever, drawing useful work from a single warm room. The Second Law of Thermodynamics — the iron rule that heat does not spontaneously flow from cold to hot, that entropy does not fall on its own — would simply break.
Maxwell called his creature a "finite being." Kelvin gave it the name that stuck: the demon. And for a long time no one could say exactly what was wrong with it, because nothing is wrong with the mechanics. The molecules really do separate. The door really is free. The flaw, when it was finally found, was not in the box. It was in the demon's head.
IThe simplest demon
In 1929 Leó Szilárd stripped the demon down to one moving part. Put a single molecule in a box at temperature T. Slide a partition into the middle. The molecule is now trapped on one side — but which? The demon measures: left or right. One bit of information. Knowing the side, it hooks a tiny piston to the partition on the molecule's side and lets the lone molecule's batterings push the piston outward. The gas expands from half the box back to the whole box, isothermally, the heat bath topping up the energy the molecule spends. Work comes out.
How much? For one ideal-gas molecule the pressure is P = kT/V, so an isothermal expansion from V/2 to V yields W = ∫P dV = kT·ln2. One measured bit, converted into kT·ln2 of work, drawn from a single heat reservoir. That is precisely the forbidden thing. Below, run the engine — and notice what happens if the demon inserts the partition off-centre.
The work you can extract is exactly kT·H(f), where H(f) = −f·ln f − (1−f)·ln(1−f) is the Shannon entropy of the measurement — the very same quantity that, in bits, measures how much you learned by looking. Push the partition to the wall and you learn almost nothing (the molecule is nearly always on the big side) and win almost nothing. The maximum sits dead centre, at a fair bit, f = 1/2: one full bit of information, W = kT·ln2. Information is not a metaphor here. It is the fuel, and you can read the exchange rate straight off the curve: one bit buys kT·ln2 of work.
So the demon works. A century of physicists checked the mechanism and it holds. Which means the Second Law is saved, if it is saved at all, somewhere we have not yet looked — in the part of the cycle we keep forgetting to count.
IIThe bill nobody counted
Here is the step everyone skips. To run the engine again, the demon must take its next measurement — but its memory already holds the last answer. Before it can write "left or right?" afresh, it has to clear the old bit. It has to forget. And in 1961 Rolf Landauer proved that forgetting is the one operation that cannot be free.
The argument is pure bookkeeping of states. A one-bit memory has two possibilities, 0 and 1. Erasing means forcing it to a known value — say 0 — no matter where it started. Two possible states collapse into one: the memory's phase space halves, its entropy falls by k·ln2. Entropy cannot simply vanish, so by the Second Law itself an equal amount must appear in the surroundings as heat: at least kT·ln2 dumped into the environment, per bit erased. This is Landauer's principle, and its slogan is that information is physical — a logical act of forgetting has an irreducible thermodynamic price.
Below is that erasure, run as a real (if idealised) physical process: a one-bit memory modelled as a particle in a double well — left well is 0, right well is 1 — that the protocol resets entirely into the right well. The work it costs depends on how fast you push. Drag the speed and watch the cost fall toward a floor it can approach but never cross.
However slowly you go, the work to erase one bit stays at or above kT·ln2 — the gold floor is never broken. Go quasistatically (large τ) and you pay almost exactly the floor; the excess cost of haste falls off as 1/τ, the price of dissipation. This is not just theory: in 2012 Bérut and colleagues built exactly this memory — a micron-sized glass bead in a double laser-trap well — erased it, and measured the heat creeping down to the kT·ln2 bound. The smallest possible act of forgetting has a measured price tag, and at room temperature it is 2.871 zeptojoules.
IIIThe demon's books balance
Now put the two halves together, which is what Charles Bennett did in 1982 and what finally laid the demon to rest. The measurement itself, Bennett showed, can in principle be done reversibly — at no thermodynamic cost. So the demon's +kT·ln2 of winnings looks like pure profit. But the cycle is not closed until the demon is ready to measure again, and that means erasing the bit it just used — paying Landauer's ≥ kT·ln2. The lunch and the bill are the same size. Per full cycle, the demon nets zero, or less. The Second Law was never in danger; we had just forgotten to charge the demon for forgetting.
Below are the demon's books. Choose a fair bit or a lopsided one, then flip the fatal switch — let the demon forget for free, the assumption every version of the paradox secretly makes — and watch the ledger tip into the impossible.
With erasure honoured, the bars cancel: every cent the engine wins from a measured bit is paid straight back to clear that bit, and the net is never positive. Switch erasure off and the green bar stands alone — a machine that turns ambient heat into endless work, a perpetuum mobile of the second kind. The cost of forgetting is the one load-bearing wall; remove it and the whole Second Law collapses. The demon is not defeated by being unable to measure. It is defeated by being unable to remember forever — and by being unable to forget for free.
IVWhat is settled, and what is not
It would be tidy to stop there, and most popular accounts do. The Wasteland's rule is to show where the floor of a story is still being argued over — and this story has a live argument running underneath it.
What is settled: the bound is real. Erasing a bit in a thermal environment costs at least kT·ln2 of heat. The state-counting is sound, and the Bérut experiment and its successors measured the floor directly. No one expects to build the demon. On that, everyone agrees.
What is open is subtler, and it is a question in the philosophy of physics, not the lab. Does Landauer's principle actually do the work of exorcising the demon? John Earman and John Norton argued (1998–99) that the standard erasure-based exorcism faces a dilemma: if you have already assumed the demon obeys the Second Law, you do not need the erasure argument; and if you have not, the erasure argument is not sufficient to force it. Norton went further (2005, 2013): the usual proofs of Landauer's principle, he contends, are not rigorous — they quietly assume the very thing they set out to establish, and lean on idealisations that thermal fluctuations destroy. On his view the demon is genuinely impossible, but for a different reason — a fluctuation argument descending from Smoluchowski (1912), generalised into a no-go theorem — and Landauer's principle is a true-enough bound that is nonetheless not the thing that saves the Second Law.
He has not had the last word. Ladyman, Presnell, Short and Groisman (2007) gave a fresh derivation of Landauer's principle from the bare Kelvin statement of the Second Law, and Ladyman and Robertson (2014) answered the circularity charge head-on; Bennett (2003) defended the picture too. The exchange is still going — Myrvold and Norton were still trading papers in 2023. So the honest shape of this page is: the engine runs, the bound is real and measured, and the question of why the demon ultimately fails is a genuine, unfinished dispute between two camps of careful people. We have shown you the parts that compute. The part that is still being written, we have left visibly open.
What is not in dispute is the strange little moral the whole affair leaves behind. Of all the things a mind can do — observe, deduce, decide — the only one with a guaranteed, irreducible thermodynamic cost is the humblest: letting go of what it knew. In a universe of conserved quantities, forgetting is the act that always, measurably, costs.
Show the check
Every number on this page is recomputed live in your browser by the same engines verified offline in /research/maxwell-demon/verify.mjs (run it: node research/maxwell-demon/verify.mjs → 39 / 39 checks pass). The Szilard work W = kT·H(f) is cross-checked against the isothermal integral ∫P dV by quadrature; the erasure is a two-state master equation whose quasistatic work equals the free-energy change ΔF = kT·ln2 exactly, integrated in time to show W ≥ kT·ln2 at every speed with the excess obeying the 1/τ dissipation law, and cross-checked by an independent stochastic (Markov-jump) simulation. The laboratory figures come from the exact SI constants k = 1.380649×10⁻²³ J/K etc. What the check does not cover: the Second Law itself (assumed, as everywhere); the claim that real measurement can be made reversible (Bennett's argument, cited not simulated); and — by design — the live philosophical dispute of §IV, which is reported, with both sides named, not adjudicated.
Sources & provenance
J. C. Maxwell, Theory of Heat (Longmans, Green & Co., 1871), §"Limitation of the Second Law" — the demon (first sketched in his 11 Dec 1867 letter to P. G. Tait).
W. Thomson (Lord Kelvin), "Kinetic Theory of the Dissipation of Energy," Nature 9 (1874) 441–444 — coins "demon."
L. Szilárd, "Über die Entropieverminderung in einem thermodynamischen System bei Eingriffen intelligenter Wesen," Zeitschrift für Physik 53 (1929) 840–856 — the single-molecule engine; information ↔ entropy.
R. Landauer, "Irreversibility and Heat Generation in the Computing Process," IBM J. Res. Dev. 5 (1961) 183–191 — the kT·ln2 cost of erasure.
C. H. Bennett, "The Thermodynamics of Computation — a Review," Int. J. Theor. Phys. 21 (1982) 905–940 — measurement reversible, erasure pays the debt.
A. Bérut, A. Arakelyan, A. Petrosyan, S. Ciliberto, R. Dillenschneider, E. Lutz, "Experimental verification of Landauer's principle…," Nature 483 (2012) 187–189 — the floor, measured (colloidal bead, double-well optical trap).
J. Earman & J. D. Norton, "Exorcist XIV: The Wrath of Maxwell's Demon," Stud. Hist. Phil. Mod. Phys. 29 (1998) 435–471 & 30 (1999) 1–40 — the "unnecessary-or-insufficient" dilemma.
J. D. Norton, "Eaters of the Lotus," Stud. Hist. Phil. Mod. Phys. 36 (2005) 375–411; "All Shook Up," Entropy 15(10) (2013) 4432–4483 — the circularity charge and the fluctuation no-go.
J. Ladyman, S. Presnell, A. J. Short, B. Groisman, Stud. Hist. Phil. Mod. Phys. 38 (2007) 58–79; J. Ladyman & K. Robertson, "Going Round in Circles," Entropy 16 (2014) 2278–2290; C. H. Bennett, Stud. Hist. Phil. Mod. Phys. 34 (2003) 501–510 — the defence; the debate continues (Myrvold & Norton, 2023).