In the Discorsi of 1638, Galileo refutes two millennia of Aristotelian dynamics without dropping a single object. Suppose, as Aristotle held, that heavier bodies fall faster. Tie a heavy cannonball to a light musket ball. The slower musket ball should retard the cannonball, so the pair should fall more slowly than the cannonball alone. Yet the pair is heavier than either component, so it should fall faster. The hypothesis contradicts itself, and the natural escape is that, absent resistance, all bodies fall at the same rate.
Nearly four centuries later, the MICROSCOPE satellite confirmed the universality of free fall to roughly one part in a quadrillion. Galileo reached the right answer by imagination alone, and that is precisely the problem. Empiricism, the default epistemology of naturalised philosophy of science, holds that knowledge of contingent physical fact requires causal contact with the world. A thought experiment supplies no new data. How, then, can it tell us anything about nature that we did not already know?
This article examines three answers that together form an integrated account. The first frames the epistemic puzzle and the rival positions of rationalist and deflationist. The second treats thought experiments as executions of mental models that encode tacit physical knowledge acquired through experience. The third argues that their deepest work is modal: they map the space of physically possible laws, revealing which structural features any coherent physics must have, and sometimes converting those revelations into experiments the laboratory can actually run.
The Epistemic Puzzle: Knowledge Without Observation
Thomas Kuhn put the problem sharply in 1964: how can reflection on familiar data, with no new observations, lead to new knowledge of nature? The question bites hardest for a naturalist. If the mind has no special conduit to physical law, then a thought experiment seems confined to rearranging what we already believe. Yet the history of physics credits imagined scenarios with genuine discoveries, from Stevin's inclined plane to Einstein's equivalence principle. Something is being learned, and its source is unclear.
James Robert Brown offers the boldest resolution. On his Platonist view, a small class of thought experiments grants a priori insight into laws of nature, construed, following Armstrong, Dretske and Tooley, as necessitation relations among universals. Galileo's falling bodies is his paradigm: we do not infer the result, we see it. The view honours the phenomenology of the case, the sense of sudden revelation, but it purchases that fidelity with a cognitive faculty that no neuroscience or developmental psychology has identified.
John Norton takes the opposite line. Every thought experiment, he argues, is an argument in picturesque dress: it proceeds from explicit or tacit premises to a conclusion, and the imagery is eliminable without epistemic loss. Whatever reliability a thought experiment possesses is inherited from its premises, which are ordinarily empirical. This dissolves the mystery neatly. Nothing is learned that was not already latent in prior belief; the scenario merely makes the entailment vivid and psychologically tractable.
The deflationary picture has its own difficulty. Tamar Gendler observed that the force of Galileo's case depends on recognising that whether the tied bodies count as one object or two is physically arbitrary, and that the rate of fall cannot hinge on such a classification. That premise appears in no explicit reconstruction until someone contemplates the scenario. The argument can be written down afterwards, but the scenario seems to be what supplies the premise. Norton's reconstruction may be accurate yet retrospective.
A further constraint complicates both camps: thought experiments fail. In 1930 Einstein proposed a photon box designed to violate the energy-time uncertainty relation; Bohr rebutted it overnight by invoking gravitational time dilation from Einstein's own general relativity. Both men found their scenarios compelling. Any adequate epistemology must explain not only how thought experiments succeed but why felt compellingness is an imperfect guide, which suggests their reliability is an empirical matter rather than a guarantee of reason.
TakeawayThe real question is not whether thought experiments produce knowledge but where that knowledge was stored before the experiment was run. Any good answer must also explain why some equally vivid scenarios mislead.
Mental Model Execution: Running Physics in the Head
A third family of views, developed by Nancy Nersessian, Nenad Miščević and others, locates the knowledge in neither Platonic insight nor propositional premises. On the mental-model account, the reasoner constructs a structured, quasi-spatial representation of a physical situation and then runs it forward. The inference is model-based: conclusions are read off the evolving state of the simulation rather than derived line by line. The format of representation, not merely its content, does epistemic work.
Ernst Mach anticipated the idea when he credited thought experiments to 'instinctive knowledge' accumulated through experience. His favourite example was Simon Stevin's 1586 wreath of spheres: a closed loop of beads draped over a triangular prism. Since perpetual motion is impossible, the loop must stay at rest; since the hanging portion below is symmetric, it can be removed; therefore the beads on the two inclines balance, yielding the law of the inclined plane. The decisive premise was tacit, absorbed from a lifetime of watching things run down.
Cognitive science now gives this proposal empirical teeth. Battaglia, Hamrick and Tenenbaum showed in 2013 that human judgements about whether block towers will topple are well predicted by a noisy, probabilistic simulation of Newtonian dynamics. The mind appears to host something like an intuitive physics engine: a generative model, calibrated by perception and action, that can be run on hypothetical inputs. A thought experiment, on this reading, deliberately feeds that engine boundary conditions it has never encountered.
The naturalistic payoff is considerable. The knowledge a thought experiment yields is not a priori; it is empirical knowledge stored in a non-propositional format and extracted through novel execution. This vindicates Kuhn's suggestion that thought experiments reorganise familiar experience, and it explains Gendler's observation: the simulation surfaces constraints, such as the arbitrariness of object boundaries, that no one had written down as premises. Norton's arguments become faithful post hoc transcriptions of what the simulation computed.
The account also predicts failure where we observe it. Michael McCloskey's studies in the early 1980s found that many people expect a ball leaving a curved tube to continue curving, an impetus-like intuition closer to medieval physics than to Newton. The intuitive engine was trained on low-speed, friction-laden, everyday interactions. When thought experiments venture beyond that regime, to light-speed observers or entangled particles, the simulator must be disciplined by explicit theory, as when Einstein's chase of a light beam succeeded only because Maxwell's equations forbade the frozen wave he imagined.
TakeawayA thought experiment is an instrument, and like any instrument it is trustworthy only within the range over which it was calibrated. Before trusting where an intuition leads, ask what experience trained it.
Modal Exploration: Mapping the Space of Possible Laws
The mental-model account explains how thought experiments draw on stored knowledge, but it underdescribes what the most celebrated ones accomplish. Galileo did not merely predict how two tied balls would behave. He showed that an entire class of candidate laws is incoherent. This points to a further function: thought experiments probe physical possibility, testing which combinations of principles can jointly hold and thereby revealing features that any acceptable physics must possess.
Reconsider the falling bodies in modal terms. The argument shows that a law making fall rate increase monotonically with weight cannot coexist with two assumptions almost no one would surrender: that weight is additive, and that the dynamics are indifferent to how we parse composite objects. Given those constraints, universality of free fall is not one option among many; it is forced. The thought experiment discovers a conditional necessity, a structural feature of the space of admissible laws rather than a measured fact about this world.
Einstein's 1907 'happiest thought' operates the same way. An observer sealed in a windowless box cannot locally distinguish uniform acceleration from a homogeneous gravitational field. The equality of inertial and gravitational mass, a brute coincidence in Newtonian theory, is thereby reinterpreted as a principle constraining any adequate theory of gravity. Consequences follow by modal reasoning alone: if light crossing an accelerating box traces a curved path relative to its walls, it must bend in the equivalent gravitational field.
Maxwell's demon shows the method operating across more than a century. The tiny sorting agent forced the question of whether the second law of thermodynamics is necessary or merely statistical. Its eventual resolution, through Landauer's 1961 principle that erasing a bit dissipates at least kT ln 2 of energy and Bennett's subsequent analysis of the demon's memory, established that information is physical. In 2012 Bérut and colleagues measured the Landauer bound directly. An imaginary creature had identified a genuine constraint linking thermodynamics and computation.
The EPR argument of 1935 completes the picture. Einstein, Podolsky and Rosen explored what follows if quantum mechanics is complete and locality holds. Bell's 1964 theorem converted that modal exploration into an empirical test, and loophole-free experiments in 2015 delivered the verdict against local realism. The boundary between thought experiment and laboratory experiment proves porous: the former maps which theories are possible given certain principles; the latter determines which region of that map we actually inhabit.
TakeawayThe most powerful thought experiments do not tell us what happens; they tell us what cannot coherently happen. Knowing the shape of the possible is often the precondition for asking nature the right question.
Thought experiments have long served as a test case for the limits of empiricism, and the rival verdicts, Platonic insight or disguised argument, each capture something real. Close attention to the scientific details suggests a more integrated picture. Thought experiments run calibrated mental models that encode tacit empirical knowledge, and the most consequential ones use those models to chart the boundaries of physical possibility.
This is the naturalism Philip Kitcher recommends: an epistemology of science continuous with cognitive science and attentive to actual practice. The reliability of a thought experiment becomes an empirical question about the provenance of its intuitions and the theoretical constraints that discipline them, not a matter settled from the armchair about the armchair.
The lesson extends beyond thought experiments. Imagination in science is not the opposite of observation but one of its products: a compressed record of experience that, rightly constrained, lets us glimpse what any world like ours must be like.