Fodor and Pylyshyn's Systematicity Challenge Still Stands

A new analysis concludes that neural networks have not yet met the systematicity challenge posed decades ago by philosophers Jerry Fodor and Zenon Pylyshyn, directly disputing recent claims that a meta-learning protocol had bridged the gap between connectionist models and human-like compositional thought [1]. The argument from systematicity, advanced by Fodor and Pylyshyn, holds that human thought and language exhibit systematic biconditional dependencies. A person who understands “John saw Mary” will also understand “Mary saw John,” a pattern symbolic systems explain by default but neural networks do not [1][2]. Fodor, who died in 2017, was a Rutgers University philosopher whose language of thought hypothesis described cognition as operating on mental representations with a combinatorial syntax, a view that directly challenged connectionist approaches [6][7]. Several recent articles had argued that the challenge was being met. Brenden Lake and Marco Baroni proposed that their meta-learning for compositionality (MLC) protocol matched and perhaps explained human systematicity [1][3]. The new study, submitted on 12 Jun 2026, tested that claim and found it premature [1]. The authors identified three distinct levels of failure in systematic generalization within Lake and Baroni’s algebraic network [3][5]. The model struggled to learn rules that were even slightly out of distribution compared to its training data [1][2]. It also behaved unsystematically on many within-distribution problems [1][4]. While the network showed some success at later learning from extremely small datasets, the researchers concluded these successes cannot be described as instances of systematic generalization [3][5]. The paper addresses a potential counterargument: that neural networks need only be as systematic as humans are in performance. Human behavior can appear unsystematic for incidental reasons—a bad experience with someone named Bill might alter a response to a sentence involving that name—but cognitive science distinguishes between a faculty that is systematic and particular circumstances that mask it [4]. The study’s authors write that the MLC protocol, in its current shape, does not address the systematicity challenge even if the challenge is narrowly construed as being exclusively about generalization in behavior [3][5]. They conclude that current neural networks with generic architectures, lacking strong prior constraints tailored to emulate the core properties of symbolic systems, neither produce systematic behavior nor explain how systematicity can be a fundamental property of minds [3][5]. The debate traces to a foundational divide in cognitive science between functionalist views, which allow that mental states can be realized in multiple physical substrates including computers, and the specific claim that connectionist networks lack the compositional structure required for systematic thought [7][8].