6.5.2 An empirical target, not a settled verdict

6.5.2 An empirical target, not a settled verdict

The 2023 essay should be read as a motivating position rather than a settled capability theorem. Language-model architectures and evaluations continue to change, and experiments can challenge particular claims about their structural biases. Wang and Ma, for example, compare learning of constructed “possible” and “impossible” languages by small transformer and recurrent models and report architecture-dependent behavior [ Wang and Ma , 2026 ] . Their reported comparisons are descriptive single runs (\(n=1\) per condition), so they do not support formal statistical inference. Such evidence can weaken an assertion that all statistical learners are indifferent to structural possibility without showing that their learned preferences constitute explicit causal theories or human-like linguistic knowledge.

The corresponding DIAL experiments must therefore go beyond asking whether a model prefers one string distribution to another. They should test whether a system can declare the relevant constraints, produce countermodels for violations, identify when no admissible realization exists, and decide when to repair a model rather than extend its theory. The strongest test is transfer: does the proposed explanation support new deductions, interventions, or experimental designs outside the data that elicited it?

Chomsky and colleagues also connect intelligence to moral reasoning. DIAL does not derive ethical principles from double involution. Its admission contracts can make normative constraints, responsibility for interventions, and rejected alternatives explicit, but the source and legitimacy of those constraints remain separate philosophical and institutional questions.

Boundary: What the critique contributes. The Chomsky–Roberts–Watumull critique supplies a demanding target: move from statistical continuation toward explicit possibility constraints, causal explanation, and responsible judgment. DIAL offers a candidate architecture for declaring, diagnosing, and revising structure. It does not yet establish that present language models lack those capacities in principle or that double involution is sufficient to produce them.