Why a Believed Label May Not Change a Belief

The legal and technical chapters explain how a disclosure can be required and how a provenance record can be checked. The audience question is different: once people know that content came from AI, do they change their judgment or behavior? This chapter interprets a survey experiment that tested that question with persuasive policy messages. Its value is not a universal answer about every label; it is a clear demonstration that communicating origin and reducing influence are separate outcomes.

What the experiment actually tested

Gallegos and colleagues conducted a preregistered survey experiment with 1,601 participants remaining after attention-check exclusions. Participants were U.S. residents recruited through Prolific. They read an AI-generated persuasive message about one of four relatively non-polarized policy proposals: geoengineering, drug importation, college-athlete salaries, or social-media platform liability.

The researchers used relatively unfamiliar or less polarized issues for a reason. If people begin with intense, settled views, a short message may have little room to move them. The study instead created a setting in which a persuasive appeal could plausibly change policy support, making it possible to test whether the authorship label altered that change.

The messages were held constant within the experiment. Participants were randomly assigned to one of three authorship conditions:

  • an expert AI model generated the message;
  • a human policy expert wrote the message; or
  • no authorship label was supplied.

The intervention therefore varied the disclosure while keeping the persuasive content the same. That is the central strength of the comparison. It asks what difference the label makes when the argument itself is not changed at the same time.

Participants reported their support for the policy before and after reading the message. The researchers also measured confidence in that support, judgments of the message’s accuracy, and intentions to share it. The primary persuasion measure was the change in policy support from before to after exposure.

sequenceDiagram participant P as Participant participant S as Survey S->>P: Measure prior knowledge, support, and confidence S->>P: Randomly assign AI, human, or no authorship label S->>P: Present the same persuasive policy message S->>P: Measure post-message support, confidence, accuracy, and sharing intention P-->>S: Report outcomes and label perception

The design does not test whether the content is factually correct, whether a legal disclosure is adequate, or whether a C2PA manifest is complete. It tests the behavioral and attitudinal consequences of one kind of authorship information in one text-based setting.

The labels were believed, but persuasion remained

The messages were persuasive overall. Policy support increased by 9.74 percentage points on average from before to after exposure. That average change establishes that the intervention had something to measure: participants were not simply ignoring the messages.

The authorship manipulation also reached most participants. In the AI-label condition, 94.6% believed the assigned authorship. In the human-label condition, 89.3% believed the human authorship information. The label therefore did not fail merely because people missed it or rejected it as implausible.