Do readers trust unlabeled AI-written messages as much as human ones?

In a pre-registered online experiment (N = 647, Prolific), recipients who were given no information about how an email was written judged its sender in a way the authors describe as "virtually indistinguishable" from a message known to be human-written. Disclosure is what moved them. When the message was labeled as entirely AI-generated, the authors report "strong negative effects" on social impressions, in both the overall rating and the net valence of open-ended impressions. When AI involvement was raised as a possibility but left unresolved, impressions were "overly positive," though still closer to the human-written baseline than to the AI-generated one. The disclosure effect appeared in all four scenarios, and 46% of the sample said they had used such tools to write messages within the past two weeks.

The authors ground this in signaling theory. Generative AI lowers the cost of writing and makes authenticity hard to verify, which weakens the qualities that make a written signal credible: difficulty to fake, verifiability, and self-sacrifice. That account predicts skepticism once AI use is suspected, and the disclosed condition shows it. It does not explain the uninformed result on its own, so the authors invoke attention. Explicit ratings differed between the uninformed and uncertain conditions, which they read as participants who "by default, did not even consider the possibility of AI involvement" until it was raised. Neither condition revealed the message's origin, so they conclude the gap "can only be explained by different levels of attention to AI involvement." The paper states that which mechanism drives the main effects is beyond its scope.

The nearest note, Does telling people an AI wrote something actually stop them from believing it?, finds the same modulation pattern for persuasion: awareness raises scrutiny without switching the effect off. This excerpt applies that pattern to impressions of a sender. Does AI writing assistance change how readers perceive the writer? measures what assistance changes about the writer; this excerpt asks how readers judge when no assistance is flagged. The authors also argue that prior disclosure studies (Glikson & Asscher 2023; Hohenstein et al. 2023; Lim et al. 2025; Weiss et al. 2022) look different once the origin is treated as uncertain. The outcome also differs from Does telling people they are talking to AI change how persuaded they become?, where an AI-identity label alone changed nothing. Labeling a message as AI-written is a different manipulation, and the two results should not be pooled.

The excerpt does not establish how often recipients meet unlabeled AI-assisted writing. The 46% figure is self-report from one sample, the scenarios were hypothetical (readers imagined an email from "Alex"), and the excerpt reports no effect sizes. What it supports is narrow: for single, hypothetical messages, the default reading of an unlabeled message is trust rather than suspicion, and the authors call that default blissful ignorance. Whether the default holds as awareness grows is the open question in Does trust in unlabeled AI messages decline as awareness grows?.