Multimodal Semantic Integrity and Safety: A Layered Framework for Evaluating AI-Generated Human Representations

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This concept document proposes a multidisciplinary framework for evaluating the semantic integrity and safety of AI-generated human representations across linguistic, audio, visual, cross-modal, provenance, consent, and human-review layers.The framework begins from the premise that many harms associated with generative AI did not originate with AI itself. Existing vulnerabilities such as stolen identity data, weak verification, misleading advertising, manipulated media, and slow reporting or removal systems already existed. Generative AI can amplify these vulnerabilities through increased scale, speed, realism, personalization, and automation.The framework combines a seven-stage harm chain with six multimodal evaluation layers and emphasizes that safety systems must distinguish observation from inference, preserve uncertainty, avoid profiling normal human variation as deception, and maintain meaningful pathways for human review, correction, and accountability.An initial pilot is proposed using AI-generated user-generated content and synthetic identity as a bounded test case. The document is intended for multidisciplinary review and further development across linguistics, audio and media forensics, computer vision, human-computer interaction, accessibility, privacy, governance, legal and evidentiary standards, AI evaluation, and red teaming.