The Unintended Consequences of Labeling AI-Generated Media Online

Contributions

• First study examining how AI-generation labels (distinct from misinformation warnings) affect user perceptions of political content
• Distinguishes three outcomes: perceived authenticity, belief in claims, and sharing intentions—prior work conflates these
• Identifies "implied authenticity" effect: absence of a label signals authenticity, not just absence of detection
• Demonstrates that people conflate signals about how content was created with signals about whether claims are true
• Evidence from pre-registered experiments with diverse political headlines and image types

Method

Two pre-registered survey experiments with 11,044 U.S. respondents (Cint Theorem platform) in which participants rated political news posts. Design:

• Stimuli: 24 political news headlines (varying in veracity: true/false) paired with images (authentic or AI-generated). Headlines drawn from fact-checking sites or established news outlets.
• Image generation: Authentic images sourced from news outlets; AI images generated using Ideogram 2.0 or OpenAI Sora to match headline claims.
• Conditions:
• Control: all posts unlabeled
• Treatment: ~75% of posts with AI-generated images labeled; unlabeled content remains visible
• Outcomes measured (0–1 scale):
• Perceived authenticity of image: "To what extent was this image generated using AI?"
• Belief in claim: "Do you believe the events described in the headline actually occurred?"
• Likelihood of sharing: "How likely would you be to share this post on social media?"
• Heterogeneous effects tested: Partisanship, partisan concordance with headline, headline-image concordance, political knowledge, cognitive test scores

Results

Direct labeling effects: Labels effectively reduce perceived authenticity of AI images but produce negative unintended consequences:

• Perceived authenticity of AI images drops 8–15 percentage points when labeled (depending on headline veracity); largest effect when labeling true claims (−15pp, CI: [−0.17, −0.13])
• Belief in accompanying claims decreases 4–7 percentage points (effect size ~half of authenticity effect)
• Sharing intention drops ~6 percentage points for both true and false headlines
• People conflate authenticity judgments with veracity judgments: respondents infer labeled "AI-generated" as "false" even when the claim is factually true

Implied authenticity effect: When some content is labeled, unlabeled content appears more authentic:

• Exposure to partial labeling increases perceived authenticity of unlabeled content by 3–5 percentage points (regardless of whether the claim is true/false or image is authentic/AI)
• ~37% of respondents explicitly report inferring that unlabeled content had "been checked and verified as not being AI-generated"
• ~24% assume unlabeled images "had not been checked by professionals"

Belief and sharing: Limited implied effects on belief (≈0pp) but modest reductions in sharing intentions when exposed to labeling (2–4 percentage points)

Heterogeneous effects: Individuals with higher digital literacy and risk literacy show larger reductions in perceived authenticity of labeled content; moderating effects of partisanship and concordance are significant

Notes

Strengths:

• Pre-registered design with large sample and two experiments for replication
• Explicit measurement of implied effects—prior work wouldn't detect this
• Separates authenticity from veracity, revealing that people conflate these signals
• Real, balanced stimuli with crowdsourced vetting of believability
• Addresses a timely policy question as platforms scale AI-labeling initiatives

Limitations:

• Survey setting (not real social media feed); information overload and competing content incentives differ in deployment
• Limited to political news; generalization to other domains (marketing, art, journalism) unclear
• AI generation limited to image; audio/video AI labels may operate differently
• Detection coverage scenarios modeled artificially; real-world partial labeling coverage dynamics unknown

Open questions:

• How do label design and wording (specificity, placement, visibility) affect the implied authenticity effect?
• Do effects differ across platforms and user bases?
• Can labeling be redesigned to reduce authenticity-veracity conflation (e.g., clearly distinguishing "method of creation" from "accuracy")?
• How do effects change with exposure frequency and over time?