Plausibility Corruption

In the AI Story

Crawford's framework reveals plausibility corruption as structurally distinct from older epistemic hazards. Deliberate deception requires an agent with intent to deceive. Bullshit, in Harry Frankfurt's technical sense, requires indifference to truth. Plausibility corruption requires neither. An AI system produces plausible output not because it intends to deceive or is indifferent to truth but because it has been optimized to match the distribution of competent human output, and the match is primarily at the surface. The confident tone, the well-structured argument, the appropriate vocabulary — these are exactly what the training process rewards, and they are exactly what the evaluator reads as evidence of competence.

The diagnostic test Crawford's framework proposes is whether the output's correctness depends on features the evaluator could verify. A plausible legal brief citing real cases may still misread those cases in ways only deep engagement with the precedents would reveal. A plausible architectural design may meet all stated requirements while containing structural vulnerabilities that would be obvious to a practitioner who had built similar structures by hand. A plausible medical recommendation may be technically defensible and clinically disastrous. In each case, the plausibility is real — the output would pass review by any evaluator whose standards are the surface standards plausibility satisfies. The corruption lies in the gap between what plausibility guarantees (surface competence) and what users assume it guarantees (underlying understanding).

The cumulative dimension is what makes plausibility corruption particularly dangerous. Individual instances of corrupted output may be caught by individual evaluators with deep expertise. But as AI-mediated production scales and as practitioners increasingly rely on AI for work they would previously have done themselves, the pool of human expertise capable of detecting plausibility corruption progressively thins. The evaluators themselves become practitioners shaped by AI-mediated work, and their capacity to detect the gap between plausible output and grounded output erodes. The corruption compounds as the detection capacity it depends on diminishes.

The connection to ersatz expertise is direct. Plausibility corruption is the specific mechanism by which ersatz expertise passes the evaluation structures that would, under older conditions, have distinguished it from the genuine article. The mechanism is not malicious — no one is trying to pass off bad work as good. The mechanism is structural: AI produces output optimized for the properties evaluation measures, evaluators measure properties plausibility satisfies, and the gap between plausibility and understanding gradually expands until reality administers a test the structure did not anticipate.

Origin

The concept emerges implicitly in Crawford's treatment of AI-generated output across his 2024-2025 writings. Crawford does not typically use the phrase "plausibility corruption" but the concept is what his framework requires to name the specific hazard AI introduces.

The philosophical antecedents include Harry Frankfurt's distinction between lying and bullshit (plausibility corruption is a third category, produced without intent to deceive and without indifference to truth), Plato's attack on rhetoric in the Gorgias (plausibility as persuasion independent of truth), and the contemporary literature on AI hallucination and its distinctive epistemic properties.

Key Ideas

Structural, not intentional. Plausibility corruption does not require an agent with deceptive intent — it emerges from the structure of training optimized for surface match with competent human output.

The surface-depth gap. Plausibility is a surface property; understanding is a depth property; AI optimization targets the surface, producing a gap that only tests beyond the surface can detect.

Cumulative compounding. Individual instances of corrupted output may be caught by expert evaluators, but as AI-mediated work scales, the pool of expertise capable of detection progressively thins, and the corruption compounds.

Third category of epistemic hazard. Distinct from both deliberate deception (requires intent) and bullshit (requires indifference to truth), plausibility corruption emerges without agency — a structural feature of the medium rather than a choice of a speaker.

The detection gap. Distinguishing plausible output from grounded output requires evaluators with deep expertise in the relevant domain — expertise itself being eroded by the AI-mediated work it is supposed to evaluate.

Debates & Critiques

Critics of the concept argue that plausibility has always been how knowledge work is evaluated — peer review, expert judgment, and professional accreditation have always relied on surface indicators of competence. The AI transition, on this view, does not introduce a new hazard but merely scales an old one. Crawford's response is that the scale change matters because it changes the ratio between plausibility-producing output and depth-producing output in the broader culture. When most output is produced through processes that optimize for plausibility, the evaluation systems calibrated for lower volumes of plausibility-optimized output begin to fail in ways their designers did not anticipate. Whether this structural shift amounts to a different hazard or an amplified version of the old one is partly a terminological question — but the practical consequences are the same either way.