The thermostat problem
Why the thermostat
The thermostat entered the argument as the smallest machine anyone was prepared to describe in mental terms. McCarthy (1979) used it to show that ascribing beliefs to a machine could be exact, and exactly limited:
Dennett used it as the limiting case of an intentional system:
It has kept that role. Orseau et al. (2018) report Dennett's description of it as “the simplest artifact that can sustain an intentional stance”. Abel et al. (2025) make it their running example. Kosoy (2026) names it as a non-example in the first paragraph that fixes her scope. And Wentworth lists it with bacteria, neural nets and markets:
The answers, side by side
Each row is one source's own statement about a thermostat. The quote is the source's. The verdict label and the frame in the last column are the Zoo's reading of the passage: what was taken as the unit, what was allowed as evidence, and what change was imagined. No row is inferred from a definition alone.
Four ways of asking
The answers differ less because the sources disagree about thermostats than because they are asking different questions.
1. Does describing it in mental terms pay?
McCarthy, Dennett and Orseau et al. ask this. For McCarthy the ascription is legitimate when it carries the same information it would about a person, and useful when it helps us understand or repair the machine:
Dennett's criterion is predictive success, and he adds that a thermostat can be enriched step by step with no sharp line crossed. Orseau et al. turn the same idea into a comparison of two descriptions, which is why the thermostat lands where it does:
On this way of asking there is a known failure: a description that is available but buys nothing. Dennett's lectern is the standard example, and Wooldridge & Jennings (1995) make the same point about a light switch.
Whether Orseau et al. have in fact formalised Dennett is disputed:
Zoo reading. Asked this way, the thermostat is the boundary case by construction. It is the point where a mental description stops being cheaper than a mechanical one.
2. What kind of mechanism is it?
Rosenblueth, Wiener & Bigelow (1943) defined purposeful behaviour by feedback, and counted the servomechanism as a machine with “intrinsic purposeful behavior”:
Demski (2019) separates two kinds of optimization and puts the thermostat in the second:
Hubinger et al. (2019) define an optimizer by what is inside it:
Demski objects that this leaves controllers out:
Zoo reading. Rosenblueth et al. never mention a thermostat, and neither do Hubinger et al. A thermostat is a negative-feedback device, so it would be purposeful on the 1943 account; it runs no internal search, so it would not be an optimizer on the 2019 one. Both of those are inferences, which is why neither appears in the table.
3. Would it adapt if the world worked differently?
Kenton et al. (2023) make this the definition of an agent, and it gives two answers for one device depending on where the boundary is drawn:
Kosoy excludes the thermostat from the theory of agents for what looks like the same reason:
Zoo reading. Kosoy's swapped outputs are a mechanism change in Kenton's sense. Kosoy does not cite Kenton et al., so the Zoo records the link as a hypothesis.
4. Relative to which frame?
Abel et al. decline to answer until a frame is fixed, and use the thermostat to show why:
Their four properties are taken from Barandiaran et al. (2009), and they quote Kenton et al.'s own admission that agent discovery is relative to a choice of variables.
Does a regulator need a model?
A separate argument approaches the thermostat from the other side. If anything that regulates well must contain a model of what it regulates, then one ingredient of agency would come free with regulation. Conant & Ashby (1970) are the usual citation:
Wentworth (2021) argues that the “model” this theorem delivers is too weak to carry that conclusion:
His repaired version does force a real model, under conditions a thermostat does not obviously meet:
The internal model principle is a different theorem. The Zoo has not read the original; a recent exposition stresses its assumptions:
Zoo reading. None of the regulator sources read so far mentions a thermostat. Demski notes that real thermostats represent their target temperature. Representing a set-point is not modelling the room, and nothing quoted here shows that a thermostat does the second.
What moves the verdict
| What varies | Where it shows |
|---|---|
| The unit | Kenton et al. give no for the device and yes for the device with the process that made it. Abel et al. make the boundary one of the frame's commitments. |
| The evidence allowed | Behaviour only (Orseau et al.). Internal structure (Hubinger et al.). The design, when known (McCarthy). Response to a change in mechanism (Kenton et al., Kosoy). |
| Where the goal comes from | Ascribed by the describer (McCarthy, Dennett). Represented in the device (Demski). Left to the frame (Abel et al.). |
| What the concept is for | Describing and repairing programs (McCarthy). One vocabulary for machines and organisms (Rosenblueth et al.). The scope of a theory for aligning superintelligent AI (Kosoy). |
| Graded or cut | Dennett: differences of degree. Kosoy: a non-example, with animals and current AIs as marginal cases. |
Zoo reading. Two direct disputes are on record in this slice: Demski against Hubinger et al. on whether controllers are optimizers, and Wentworth against Conant & Ashby on what the regulator theorem shows. Shimi's objection to Orseau et al. is a third. The yes and no answers in the table are mostly not in dispute with each other; they answer different questions.
Open connections
These are the Zoo's own hypotheses about this slice. No quoted source asserts them. Each says what was checked.