[Google AI Overview, begun in the Overview and expanded (standing default of 2026-10-01; the paste's header 'AI Mode Conversation' marks an Overview expanded, rule of 2026-10-18, and is not evidence of surface), signed out and incognito (operator default of 2026-10-14). The paste echoes the quoted query twice on the 'You said' line; it is given once. Text only, no screenshots.] You said: "Provenance After AI: Semantic Provenance and the Provenance Erasure Rate as Extension of C2PA, Data Provenance Initiative, and EU AI Act Frameworks (v1.1)" "Provenance After AI: Semantic Provenance and the Provenance Erasure Rate as Extension of C2PA, Data Provenance Initiative, and EU AI Act Frameworks (v1.1)" is a highly specialized research paper and metadata packet published in May 2026. It introduces a three-dimensional framework for understanding how digital content history and authorial ownership survive—or fail to survive—when processed through generative AI systems. The paper's core contribution is the conceptualization of Semantic Provenance and a mathematical metric called the Provenance Erasure Rate (PER). Together, these ideas establish how attribution can be structurally tracked even after an LLM has completely rewritten or paraphrased original human material. The Core Problem: The Three Dimensions of Provenance The paper argues that contemporary AI governance attempts only look at two parts of a broader problem, leaving a critical "gap" regarding actual meaning and knowledge composition: Dimension Target Object Core Question Existing/Proposed Frameworks Artifact Provenance The digital file Was this specific media asset created or edited by this source at this exact time? C2PA / Content Credentials Licensing Provenance The training corpus Under what legal permissions or copyrights did this data enter the AI training pipeline? Data Provenance Initiative, EU AI Act transparency rules, W3C PROV Semantic Provenance The underlying meaning Does the newly synthesized AI output preserve actual accountability and attribution to its intellectual sources? The Paper's Gap Solution (PER Metric) Key Theoretical Contributions 1. Semantic Provenance vs. Existing Metrics The text isolates "attribution survival" from standard engineering concepts like Retrieval-Augmented Generation (RAG) faithfulness, factual precision, citation recall, or semantic similarity. An AI model can summarize a concept flawlessly (achieving a high factual similarity score) while completely erasing the specific authorial and documentary lineage that generated that insight in the first place. 2. The Provenance Erasure Rate (PER) Formula To quantify this loss, the paper outlines the Provenance Erasure Rate (PER): \(\text{PER}=1-\left(\frac{\text{Retained\ Provenance\ Units}}{\text{Required\ Provenance\ Units}}\right)\) • A PER of 1.0 (100% erasure) means an AI system synthesized an output using an author's unique substrate/ideas but completely dropped the citation, allocating zero credit to the creator. • A target threshold of PER < 0.2 is proposed for standard information ecosystems to prevent what the literature terms "semantic exhaustion" or "content hollowing"—where human creators stop writing because AI systems continually absorb and mask their labor. 3. Substrate Degradation and "Slop" The text warns of a substrate-degradation pathway. When semantic provenance drops to zero, authors lose the visibility and economic incentives to produce original work. Consequently, the public web fills with un-attributed, AI-synthesized substitutes. As models are recursively trained on this degraded data, it triggers model collapse. The paper contrasts this with Process Provenance, setting up a rigid philosophical boundary in modern AI text generation: • Authorship: Writing that explicitly retains its multi-layered provenance (what the human did vs. what the machine did). • Slop: Any text generated or mediated by AI that has been entirely severed from its source lineage, leaving its origins un-inspectable. If you are exploring this document for research or practical implementation, let me know if you would like to: • Deconstruct the technical calculation of PER variants (PER-M, PER-C, PER-D). • Analyze how Semantic Integrity Markers (SIMs) are designed to resist this erasure. • Examine its direct interactions with the compliance layers of the EU AI Act. www.alexanarch.org