Language Is an Interface, Not Proof of Thought: Toward Warranted Explanation in Artificial Intelligence

The success of large language models has encouraged an increasingly close association between language and artificial intelligence. Yet the ability to acquire extensive structure from language does not establish that intelligence must itself be linguistic. Drawing on Michael Polanyi’s account of tacit knowledge, this paper proposes an alternative framing: language may function partly as a compatibility interface through which differently constituted systems communicate, without necessarily reproducing the native representations or processes being communicated. This framing does not exclude language from computation; generated linguistic sequences may themselves participate in producing an answer. The distinction has consequences for artificial-intelligence explainability. Research has demonstrated that chain-of-thought explanations can omit experimentally introduced influences on model predictions. Related work distinguishes the legibility of generated reasoning from its faithfulness, while intervention-based studies distinguish representational alignment from causal contribution. These findings motivate a distinction between explanation as a linguistic claim and faithfulness as an empirically testable relationship between that claim and system behavior. Building on research in chain-of-thought faithfulness, counterfactual simulatability, causal intervention, and shortcut learning, this paper proposes a seven-part evaluation program. The program distinguishes intermediate reasoning, retrospective causal accounts, and task-grounded justifications, and evaluates their claims through controlled interventions, counterfactual commitments, invariance tests, and provenance challenges. The proposed assurance objective is warranted explanation: an explanation whose evidentiary weight depends on independent evidence concerning specified claims, rather than linguistic plausibility alone. Warrant remains conditional on the explanatory scope, intervention design, evidence quality, task domain, and system configuration evaluated.