Which workplace cues survive AI mediation and which disappear?

"The Fabricated Front" reads workplace communication through Goffman's idea of a "front," a patterned performance through which workers make competence and commitment available for interpretation. Using 1,250 interview transcripts from Anthropic's AI Interviewer dataset, it names five opacity mechanisms through which generative AI reorganizes that front: voice, provenance, vulnerability, attention, and investment. Its "most consequential finding" is an asymmetry. Workers "did not treat AI mediation as uniformly problematic"; they were "more likely to protect identity-bearing cues, especially voice and provenance, while allowing labor-bearing cues such as effort, attention, and uncertainty to disappear into otherwise acceptable outputs."

The paper explains the asymmetry by what each cue is attached to. Voice and provenance "attach an artifact to a recognizable source," so they stay contestable: a colleague can notice that a message no longer sounds like its sender, and a client can ask whether a designer really produced an image. Effort, attention, and uncertainty have no such anchor. They "receded into the completed task and became accountable mainly when something went wrong." The authors trace this to the "output-centered organization of contemporary work," where the deliverable (the email sent, the ticket closed) is commonly taken as evidence that work occurred. On this account the erosion of labor-bearing cues goes unchallenged not because workers endorse it but because the workplace has no ordinary occasion to ask about them.

The finding sharpens a theme in nearby notes by moving it from the individual to the exchange between coworkers. Does polished AI output trick audiences into trusting it? describes audiences inferring accuracy from a finished look, while this paper describes colleagues inferring that engagement happened from a finished deliverable. How do AI tools trick users into overestimating their own skills? concerns how a user's own sense of capability inflates; the present paper concerns which cues others still hold that user accountable for, and its abstract says weaker cues undermine "the reciprocal exchange that sustains collaborative trust." The provenance mechanism ("who can stand behind the artifact") also bears on Do users truly own the AI-generated content they produce?: authorship as a social claim gets policed, while the process-level engagement behind it mostly does not.

The excerpt does not report how many workers raised each cue, how "more likely" was measured, which occupations or tools are involved, or whether the interview accounts were checked against actual work. It also does not test whether hidden effort or attention leads to worse outcomes, and the correspondence I draw between the discussion's effort, attention, and uncertainty and the abstract's investment, attention, and vulnerability mechanisms is my reading, not something the excerpt states. At the strength the evidence allows, the paper suggests that the cues most exposed to AI mediation are the ones a deliverable-centered workplace was least likely to be checking already, so any attempt to make AI-assisted work legible has to make those labor-bearing cues visible on purpose rather than rely on colleagues to notice their absence.