There is an exception to this last response. In a most promising approach, Nersessian (1992) and Palmieri (forthcoming) call on the literature on mental modeling in cognitive science to explicate mental simulations in thought experiments. This literature accounts for the relevant cognition through the formation of mental models which guide our cognition. An extremely simple example—a further simplified version of Johnson-Laird, 1989, p. 472—is drawn from two assertions:
The fork is to the left of the plate.
The knife is to the right of the plate.
They allow formation of the mental model
Fork Plate Knife
This mental model in turn licenses further assertions such as
The plate is to the right of the fork.
The fork is to the left of the knife.
If these mental models can somehow be grounded properly in experience, then why should they not produce knowledge of the world if they are used in thought experiments? For me the real question is how they can do this reliably. Here's how. These models are built from templates into which we slot particulars. In the above example, the template is
Object 1 Object 2 Object 3
and we substitute fork/ plate/ knife for object 1/ object 2/ object 3. If this template correctly reflects the nature of space, then the resulting model can be used reliably. But now we see immediately that this reliability is purchased exactly by incorporating just the sort of generalized logic discussed above in the context of 4. The Reliability Thesis. The templates just are the schema of a generalized logic. Thus my response is (3b) Incorporation. Using this mental model literature implements exactly the sort of generalized logic I had in mind. Knowledge of the world enters the thought experiment in the factual templates that ground the physical scenarios imagined.14
In principle, mental models may function this way in thought experiments. However the case is yet to be made. That mental modeling might be a good account of ordinary cognition does not entail that it is a good account of thought experiments, a highly contrived activity within science with a deceptively simple name. What tempers my optimism is that I know of no example of a thought experiment in science that depends essentially on such mental modeling; or at least all the good candidates I have seen are indistinguishable from argumentation concerning pictures and schemes, much as proofs in Euclidean geometry are just arguments about certain figures. And there are many cases in which the core of the thought experiment is an explicit mathematical derivation in a physical theory and that is unambiguously an argument. (An example is Bohr's version of Einstein's clock-in-the-box thought experiment. Its essential content is the computation of a relativistic correction to the timing of a process. See Bishop, 1999.)
What of the remaining cases in which the decision is unclear? Mental modeling and traditional argumentation can be very close and thus hard to distinguish. Cognitive theorists allow that the supposed mental models of thought experiments can be reconstructed as arguments. The reverse may also be possible: the explicit argumentation of thought experiments can be simulated by mental models. Which is it? When they are close, I favor argumentation since thought experimentation is a far more highly refined activity than simple cognition about the placing of knives and forks on a table. It must be unambiguously communicable in a few words, and its outcome objectively verifiable—conditions that suggest a need for something more secure like argumentation. Or perhaps the evolutionary processes described above in 4. The Reliability Thesis have already culled out from the generalized logic of mental models those logics that might explicitly serve science.
I leave the matter to those with expertise in cognitive science, since either way I believe that characterization of thought experimenting as argumentation survives, if the activity is to be reliable. However the scope of the program is limited by the presence of explicit argumentation (as derivations within theories) at the heart of many thought experiments.
Thought experiments are so named since they mimic real experiments, the ultimate epistemic channel to the world, at least in the empiricists' view. Might thought experiments attain some epistemic power from their mimicking of this ultimate ideal? Experimentalism answers that they do. This mimicry is the factor X. My principal response is (3c) Epistemic Irrelevance. The reason is simple and obvious. Mimicking an experiment is just not the same thing as doing an experiment; one doesn't learn about the world by the mere fact of feigning contact with it. Here I set aside whether (3b) Incorporation is also an appropriate response since I do not want to decide what it takes for a thought experiment to be experiment-like in the way that (purportedly) gains it epistemic powers. Certainly if all that is required is to be experiment-like is that the thought experiment describe an imaginary experiment and even trace its execution, then that can be done by an argument.
There are two general accounts of how thought experiments gain epistemic powers from their experiment-like character. Sorensen (1992, p.205) defines a thought experiment as "…an experiment…that purports to achieve its aim without the benefit of execution." They have the power "to justify beliefs in the same fashion as unexecuted experiments." (p.213) And this last power derives from the fact that ordinary experiments persuade (justify?) in two ways: first by "injection of fresh information about the world"; and second by "…armchair inquiry: by reminder, transformation, delegation, rearrangement, and cleansing." (p. 251) Thought experiments avail themselves of the second mode only. In so far as the second is merely a synonym for argumentation, perhaps from tacit knowledge, then obviously Sorensen's view would accord, in the end, with the argument view. But Sorensen apparently does not accept that thought experiments are arguments.15 Recalling the discussion of the Reliability Thesis above, this invites the response of (3d) Unreliability, unless Sorensen can offer another reliable, truth preserving way of transforming or rearranging what we are reminded of.
The second account accords epistemic power to a thought experiment through their realizing some idealized limit of actual or possible experiments. (See Laymon, 1991, for such an account and for examples.) The immediate problem with this account is that it cannot provide a complete epistemology since many thought experiments are not idealized limiting cases.16 In so far as not all thought experiments manifest this factor X, the response is (3a) Denial. In those cases in which the factor X is present, the obvious response is (3c) Epistemic Irrelevance. Unless we are simply inferring the results from the properties assumed of the ideal limit, why should our imagining of this limit have any epistemic power? I will not pursue this line since Laymon (1991) seems to agree, in so far as he analyses thought experiments as tacit argumentation.