Whoever can think that John loves Mary can think that Mary loves John; whoever understands that the cat is on the mat understands that the mat is on the cat. For Fodor & Pylyshyn (1988), these capacities do not come separately: they are intrinsically connected, and the best explanation is that thoughts have constituent structure and that mental processes are sensitive to it, as in a "language of thought" (Russin et al., 2025, sec. 2.3; Buckner, 2024, p. 72).
Hence their famous dilemma against connectionism: either neural networks fail to exhibit systematicity, and are empirically inadequate, or they exhibit it because they implement a classical architecture, and then offer no new theory of the mind (Millière & Buckner, 2024a, sec. 3.1). The replies of the 1990s were technical (Smolensky showed how to encode structure in vectors without discrete symbols) and conceptual: several notions of compositionality were distinguished, and critics objected that the debate paid little attention to how systematic humans and animals actually are (Buckner, 2024, p. 72).
The debate was revived by data. Lake & Baroni (2023) trained a standard transformer on a stream of compositional tasks (meta-learning for compositionality, MLC) and compared it with people learning instructions in a pseudolanguage: against symbolic models that were systematic but rigid and networks that were flexible but unsystematic, only MLC combined both and even reproduced human error patterns; the authors present the result as an answer to the challenge (2023, p. 115). Russin et al. (2025) see meta-learning and in-context learning as a general route to reproducing the "behavioral signatures" of compositionality (2025, sec. 3). Woydt et al. (2026) reply that this works, if at all, only under a very narrow definition of the meta-learning setup, and that Fodor and Pylyshyn's criteria remain unmet.
Two philosophical questions remain. Is human systematicity all-or-nothing, or graded? And if a network learns it rather than having it built in, does it fall onto the second horn of the dilemma or dissolve it (see compositionality)?