Smooth Failure

In the [YOU] on AI Field Guide

The cycle that [YOU] on AI inaugurates is, among other things, a manual for productive collaboration with a machine that is sometimes wrong with perfect confidence. The Deleuze misattribution—a passage connecting Csikszentmihalyi’s flow state to a Deleuzian concept of “smooth space” in a way that would have been obviously wrong to anyone who had read Deleuze—is the cycle’s specimen smooth failure. The passage worked rhetorically. It was eloquent, well-structured, and convincing. The idea beneath the eloquence was hollow. And the hollowness was undetectable without the domain knowledge to recognize it—which means it was undetectable to every reader who lacked that knowledge and perfectly legible to every reader who possessed it.

Smooth failure is the specific pathology that the decorrelation of fluency from authority produces in practice. A system trained to generate fluent text generates fluent text whether the content is accurate or fabricated. The surface quality of the output is not a signal about the substrate quality; it is a property of the model’s generation process that applies uniformly. This structural feature of current AI means that the most dangerous errors are precisely the ones that feel most like successes—the ones that read well, that fit the expected register, that arrive without the friction that ordinarily triggers re-examination.

The cycle’s response to smooth failure is the practice of interrogative vigilance: not the wholesale rejection of AI output, which would be the Swimmer’s posture, but the disciplined maintenance of skepticism toward output that feels correct, the deliberate generation of the question “is something wrong here?” even when nothing in the output’s surface suggests it. The practice is psychologically expensive because it runs counter to the cognitive ease that smooth output is designed to produce. It requires the collaborator to sustain an orientation that the tool’s design actively undermines—and it depends, ultimately, on the domain knowledge that makes undetected error detectable.

Origin

The concept emerges from the intersection of Dweck’s research on failure response with the specific properties of large language models. Dweck’s framework assumes visible failure: the growth-mindset response—engage, examine, learn, adjust—is triggered by a signal that something has gone wrong. The signal must exist for the signal to be read. The AI transformation introduced a category of error that produces no external signal, because the quality of the surface presentation is entirely decoupled from the quality of the underlying content.

Kahneman’s work on processing fluency provides the mechanism. Fluent text is judged more credible than disfluent text across a wide range of domains, including factual claims, logical arguments, and personality assessments. The fluency heuristic is adaptive under normal conditions, where the effort invested in clear presentation does tend to correlate with the care invested in the underlying content. It becomes a liability in any environment where surface quality and content quality are decoupled—which is precisely the environment that language models create. The same generation process that produces accurate, clearly expressed information also produces inaccurate, clearly expressed information; the fluency is not a reliability signal but an artifact of the model’s training on fluent human text.

Key Ideas

The detection asymmetry. Smooth failure is not equally detectable across all users. It is detectable only by those who possess domain knowledge sufficient to recognize that the plausible thing is not the true thing. The junior practitioner who lacks that knowledge is precisely the person most likely to accept the smooth failure, and the senior practitioner who possesses the knowledge to catch it is the person whose expertise the AI is simultaneously rendering economically less scarce. The corrective therefore has a distributional dimension: the people most at risk of smooth failure are those with the least independent knowledge, and the infrastructure that makes interrogative vigilance possible—domain education, independent engagement with primary material, unassisted practice—is precisely what smooth AI interfaces tend to displace.

Smooth failure and the growth-mindset extension. The standard growth mindset responds to visible difficulty. Smooth failure requires an extension of that orientation: the capacity to interrogate success with the same rigor that the growth mindset brings to failure. Not because every AI output is wrong, but because the smoothness of the output means that the distinction between genuine success and concealed failure cannot be made without active investigation. This extension produces what the cycle calls interrogative vigilance, and it is a harder discipline than the standard growth-mindset response because it must be self-generated rather than triggered by an external signal.

Productive distrust. The appropriate orientation toward AI output is neither wholesale trust nor categorical rejection but productive distrust: the capacity to hold simultaneously an openness to the machine’s contributions and skepticism toward those same contributions. This dual orientation requires the ambiguity tolerance that the growth mindset cultivates and that the fixed mindset resists—the ability to maintain contradictory stances without resolving them prematurely into either wholesale trust or wholesale rejection.