Do unfounded AI accusations harm human writers instead?

The paper's second claim is about who is harmed. It argues that accusations of AI use may invert the direction of testimonial injustice that the literature has traced to AI systems. The authors cite Kay et al. (2024), who position "the AI system as a perpetrator of testimonial injustice, with harm flowing from machine to human," and Hauswald (2025), who theorizes AI systems as artificial epistemic authorities. The paper's own claim runs the other way: "lay AI literacy, operating in degraded epistemic conditions, produces testimonial injustice between humans at population scale, in this case directed towards writers by readers."

The mechanism leans on Fricker's account of epistemic injustice as a speaker being "wronged in their capacity as a knower." A reader who labels a comment machine-written does not necessarily dispute its content; the label discounts the writer as its source, whether or not it is accurate. The paper frames the choice that AI-mediated communication forces as one "between gullibility and blanket distrust" (citing Sahebi & Formosa 2025). Accusation is the reader-side act that makes the second option cheap to take. The evidence offered for the mechanism is the accusation register itself: the matched-control result shows that accused comments are not distinguished by the features that separate AI text from human text.

Within the library, this turns the AI-as-perpetrator framing around. The Thin Line note finds that audiences who suspect AI involvement grow more critical yet are still swayed Does telling people an AI wrote something actually stop them from believing it?; this paper shows the same suspicion can be aimed at a human writer and recorded as a judgment about the text. The paper also argues for a change of focus: "Future work needs to take the receiving side seriously as the primary site of analysis, with production effects as downstream consequences of policing intensity." That is a direct contrast with production-side studies such as Does AI writing assistance change how readers perceive the writer?, which the paper says treat the writer as the locus of action. The gatekeeping evidence in Do AI slop accusations actually detect AI text? is what makes the harm claim plausible in the first place.

The excerpt does not establish the injustice directly. It measures what accusations say and the register they form, not who wrote the accused comments. It does not report how many accused comments were human-written, nor any consequence for writers' credibility, hiring or reception. The "unfounded" premise therefore rests on the matched-control result, which shows accused text does not carry AI-distinguishing features, rather than on verified authorship. The discussion hedges with "suggesting," while the conclusion says the paper documents the inverted form; the writer-side harm itself is not measured here. A test would need authorship verified and writers' outcomes tracked, and the English-only scope limits how far the pattern generalizes, as the limitations section notes.