ifc-0090

6.5 Prediction, explanation, and structural possibility

Chomsky, Roberts, and Watumull’s 2023 essay, “The False Promise of ChatGPT,” offers a deliberately sharp critique of the conception of language and knowledge embodied by statistical machine learning [ Chomsky et al. , 2023 ] . Its central concern is not that language models lack practical uses. It is that increasingly successful prediction of probable continuations may be mistaken for the human capacities to construct explanations, distinguish possible systems from impossible ones, and reason within causal and normative constraints. On their account, a compact theory that explains why an event can occur is categorically different from a model that becomes proficient at reproducing its statistical appearance.

This criticism converges with McCarthy’s distinction between appearance and reality and with Zahavy’s argument that scientific creativity may require an abductive jump beyond a fixed predictive language [ McCarthy , 2007 , Zahavy , 2026 ] . All three positions resist the identification of intelligence with compression or curve fitting alone. They ask how a learner obtains the theoretical entities, mechanisms, and laws that make observations intelligible and support questions not already implicit in the original representation.

The distinction between improbable and impossible is especially important for DIAL. A probabilistic model may assign a construction a small value \(p(x\mid D)\), but low probability does not establish that the construction is ill-typed, inconsistent, or unrealizable. Relative to a sketch \(\mathbb S\) and a registered semantic class, structural impossibility means that no model in that class satisfies the declared equations, cones, and preservation doctrine. Failure of a search procedure to find such a model is not by itself a proof of nonexistence. The resulting implications should not be collapsed:

\[ \begin{aligned} \text{unlikely observation} & \not\Rightarrow \text{poor fit of the current realization},\\ \text{poor fit of the current realization} & \not\Rightarrow \text{violation of the current theory},\\ \text{violation of the current theory} & \not\Rightarrow \text{inadequacy of the theory itself}. \end{aligned} \]

The first two conditions may be assimilated by changing beliefs or model parameters. A violation may expose bad data, a failed observer, or an inadmissible proposal. Persistent, localized failure supported by independent evidence may warrant opening a candidate extension; the declaration changes only after the proposed extension and its transport pass admission.