Corruptible Testing

In The You On AI Field Guide

The concept emerges by contrast with the motorcycle test. The motorcycle is incorruptible because its verdict is administered by the physical system itself — the engine runs or does not, and no rhetorical sophistication can change that fact. Corruptible tests operate differently. They are conducted by evaluators whose frameworks may share the assumptions being tested. They are conducted against specifications defined in advance that may or may not capture the full complexity of the situation. They are conducted under conditions that differ from the conditions of actual use. Each of these features is a form of mediation that introduces room for the tested output to pass while lacking the understanding the test was supposed to verify.

The plausibility problem is the specific form corruption takes in AI-mediated work. AI-generated output is optimized for plausibility — sounding right, reading well, matching the patterns human evaluators associate with competent work. Plausibility is a surface property, a property of how the output presents rather than of what the output means. A plausible legal brief may contain a fatal misreading of precedent. A plausible architectural design may contain structural vulnerabilities invisible to anyone who has not built similar structures by hand. A plausible medical recommendation may be technically defensible and clinically disastrous. In each case, the plausibility passes the evaluator's review, and the evaluator's review is the only test available.

The circularity at the heart of corruptible testing is what makes it structurally different from the motorcycle test. The specification against which AI output is evaluated is defined by the same human understanding that the test is supposed to verify. If the specification is incomplete — if it fails to capture features of the situation that matter — then the test passes output that meets the incomplete specification while missing features that only embodied engagement with the situation would have revealed. The motorcycle test avoids this circularity because its standard is administered by the engine's behavior rather than by any human-constructed specification.

The cultural consequences Crawford identifies follow from the progressive expansion of corruptible testing as AI enters more domains. As more work is produced through AI-mediated processes and tested against corruptible standards, the culture's capacity to recognize the difference between genuine understanding and persuasive simulation diminishes. Practitioners trained in environments where corruptible tests dominate lose the calibration that comes from sustained encounter with incorruptible standards. The loss is gradual and self-concealing, because the outputs continue to pass the tests that remain.

Origin

The concept is implicit throughout Crawford's work but becomes explicit in his AI writings of 2024-2025, where the contrast between AI-mediated evaluation and the motorcycle test sharpens into a specific diagnostic category. Crawford does not typically use the phrase "corruptible testing" but the concept is what his framework requires to make the distinction with the incorruptible standard work diagnostically.

The philosophical antecedents run through the epistemological tradition concerned with what distinguishes genuine knowledge from its plausible double — a tradition that includes Plato's allegory of the cave, medieval debates about the distinction between scientia and mere opinion, and contemporary epistemology's attention to the problem of "Gettier cases" where justified true belief fails to qualify as knowledge.

Key Ideas

Surface optimization. AI-generated output is optimized for plausibility — the surface properties evaluators associate with competent work — producing output that passes corruptible tests while lacking the depth that would distinguish it from genuine understanding.

The circularity problem. Specifications against which AI output is evaluated are defined by the same human understanding the test is supposed to verify, creating a self-referential structure that the motorcycle test's material administration avoids.

Evaluator framework dependence. Tests mediated by human evaluators can be passed by output that matches the evaluator's framework even when the framework itself is incomplete or mistaken — a form of failure the material test cannot produce.

The cultural consequences. Progressive expansion of corruptible testing as AI enters more domains erodes the cultural capacity to recognize the difference between genuine understanding and persuasive simulation.

Self-concealing degradation. The loss of calibration against incorruptible standards is invisible from inside environments where corruptible tests dominate, because outputs continue to pass the tests that remain.