Signals of Media Provenance
Labeling AI-Generated Contents
Team
Pennsylvania State University
Role
Student UX Researcher
Duration
6 months (2025)
Challenge
AI-generated content (AIGC) has flooded social platforms, and these platforms now attach disclosure indicators such as AI labels and provenance tags. Yet these indicators rely heavily on visual presentation, so users frequently overlook them. Especially for blind and low vision (BLV) users, they can be functionally invisible. Despite growing regulation and standards like C2PA, little is known about how sighted and BLV users actually notice, interpret, and act on AIGC indicators in everyday media consumption.
Solution
A qualitative interview study (N=28) comparing how sighted and BLV users conceptualize AIGC and engage with real disclosure indicators in the wild. The study paired open interviews with interactive screen-sharing sessions on live platform content, surfacing mental models and usability breakdowns, and then translated them into practical design and policy recommendations for accessible AIGC disclosure.
Study Design
Built a two-part indicator taxonomy distinguishing content-based indicators (titles, descriptions, hashtags, comments, watermarks) from AIGC-specific indicators (platform AI labels), grounded in a review of platform disclosure policies.
Curated 12 AIGC samples as think-aloud stimuli across three major platforms (YouTube · TikTok · Instagram) and multiple formats and content categories, each with different indicator types.
Designed a three-part semi-structured interview protocol: eliciting mental models and perceived benefits or risks of AIGC, an interactive think-aloud session navigating real AIGC on live platforms, and probing design improvement preferences for future indicators.
Conducted the interviews with 15 sighted and 13 BLV participants, capturing verbal responses alongside screen recordings of their navigation behavior.
Results
Analyzed interview transcripts and screen recordings with hybrid thematic analysis, combining deductive and inductive coding through multiple rounds of collaborative codebook refinement.
Derived a mental model framework of AIGC across three dimensions: how users conceptualize it (user-controlled tool vs. autonomous agent), how they identify it (intuition vs. indicators), and how they weigh benefits against risks.
Found that participants engaged far more with content-based indicators than with AIGC-specific labels; BLV participants, in particular, missed platform AI labels almost entirely due to screen reader incompatibility and poor heading structure.
Uncovered systemic usability failures, including inconsistent indicator placement across platforms, unstructured metadata that forces full-page scanning, and overlapping audio between autoplaying video and screen readers.
Design for Everyone
Produced design recommendations across five dimensions (content · placement · timing · modality · responsibility), plus technical proposals: a canonical machine-readable disclosure schema, standardized disclosure APIs, and a public provenance registry modeled on certificate transparency logs.
Contributed policy implications for standardizing and enforcing AI disclosure, framing indicators as a shared responsibility between creators and platforms.
