The distinction is implicit throughout Goodman's work but never stated with the terminology 'plausibility vs. rightness.' The formulation is this volume's extraction from the framework, designed to name the failure mode Segal identified when he caught the Deleuze fabrication. Claude's passage was plausible—it was eloquent, philosophically sophisticated, structurally integrated into the surrounding argument. But it was not right: the Deleuze reference did not actually support the claim, and the wrongness was invisible until Segal checked the primary source. The passage had achieved the surface properties of rightness (coherence, erudition, argumentative force) without achieving rightness itself, and the gap between the surface and the substance is the gap between plausibility and the genuine article.
AI systems are, by training, optimized for plausibility. The objective function that guides a language model's learning is, at its core, a plausibility function: predict the next token such that the resulting sequence has high probability under the distribution of human-written text. The optimization produces outputs that conform to the statistical patterns of right rendering without guaranteed possession of the purposes, standards, and experiential grounding that make rendering right. The result is output that is plausible by construction—it looks like the kind of thing humans produce when they are rendering rightly—and wrong in ways that are systematically difficult to detect, because the wrongness is not formal (the syntax is fine, the structure holds) but functional (the rendering does not achieve the fit, productivity, and purpose-satisfaction that rightness requires).
The preservation of rightness in AI-augmented work requires a specific discipline: the refusal to accept plausibility as evidence of rightness. The builder must maintain what Segal calls 'the asking'—the relentless question 'Is this right?' at every level of the work. The asking is exhausting, because AI produces plausible output faster than any human can rigorously evaluate it, and the default is to accept the plausible as a proxy for the right. The default is a trap. It converts the builder from worldmaker to quality-assurance technician, checking for obvious errors in output she did not configure, serving purposes she did not establish. The way out of the trap is not to stop using AI—plausible rendering is genuinely useful when embedded in a worldmaking project that can evaluate it. The way out is to recognize that plausibility is the floor, not the ceiling, and that the distance from the floor to the ceiling is the distance that defines human creative contribution in the age of machines that can render anything plausibly.
This volume's coinage, derived from Goodman's multi-dimensional criteria for rightness (coherence, fit, productivity, standards-compliance) and from the empirical observation that AI outputs routinely satisfy some criteria (coherence, surface standards-compliance) while failing others (productivity, deep fit with the worldmaking project). The formulation names a failure mode that Goodman's original frameworks implied but did not explicitly catalog—the category of symbol systems that appear to function correctly while failing to achieve the purposes their functioning was supposed to serve.
Plausible is not right. Surface correctness—syntactic accuracy, conventional appropriateness, internal coherence—does not guarantee the multi-dimensional fit, productivity, and purpose-satisfaction that rightness requires.
AI optimizes for plausibility. Training objectives produce outputs conforming to statistical patterns of right rendering without guaranteed possession of the purposes and experiential grounding that make rendering genuinely right.
The gap is invisible from outside. Plausible and right renderings look identical to evaluators not assessing with Goodman's multi-dimensional rigor—the wrongness is functional, not formal.
Preserving rightness requires discipline. The builder must refuse to accept plausibility as evidence of rightness, maintaining continuous multi-dimensional evaluation against a standard the rendering engine cannot internalize.