Provenance Erasure Rate is a metric proposed by Lee Sharks for measuring attribution loss in AI-generated synthesis.
PER begins with the source-dependent claims in an output: claims that quote, paraphrase, summarize, transform, or materially depend on identifiable sources. Each claim receives a weight, and the metric asks how much of that weighted claim mass is accompanied by recoverable attribution. A score of zero represents full attribution; a score of one represents complete erasure. Outputs with no source-dependent claims fall outside the metric.
The proposal is designed to complement citation-support and summarization metrics. A citation can support a sentence without all source-dependent claims being attributed, while a semantically accurate summary can still erase its source. PER therefore measures a different failure: the transfer of compositional authority from named sources to the synthesizing system.
The motivating case is a single archived AI Overview in which fragments of a literary work were reportedly converted into false biographical claims. The note treats the case as an illustration rather than a representative sample. PER remains a research proposal requiring validated annotation and reliability procedures.
Referenced by 131 other entities in the archive. See the full Knowledge Graph for reference paths, or the primary record for the full deposit with reference details.