DISO Framework: How Douyin and TikTok Operationalise AI Disclosure Through User-Visible Design

Abstract

AI disclosure policies become meaningful only when they are exposed through interface elements that users can see, read, and revisit. This paper presents the Disclosure Interface Salience Optimizer (DISO), a reproducible audit algorithm for translating policy clauses into measurable interface features. We instantiate the algorithm in a controlled computational experiment with 60 paired interface stimuli: 30 Douyin-style and 30 TikTok-style renderings of the same scenes. A browser runner captures each stimulus at a fixed 360 x 640 viewport and extracts label geometry, contrast, opacity, persistence, interaction depth, lexical coverage, and bounded salience and semantic-completeness proxies. The benchmark archives public policy snapshots used to set the two interface priors, fixes seed 20260831, stores every row-level measurement, and verifies deterministic hashes across reruns. Douyin renderings have mean salience 0.1804 (SD 0.0109) and semantic proxy 0.9910 (SD 0.0043); TikTok renderings have 0.0136 (SD 0.0008) and 0.5203 (SD 0.0023). Paired bootstrap intervals are [0.1632, 0.1705] and [0.4700, 0.4714], respectively, and paired tests reject a zero difference. These are interface proxies, not human outcomes or platform-wide prevalence estimates. The contribution is an auditable algorithm and experiment design that makes assumptions, measurements, and evidence boundaries explicit.