Answers are grounded in this paper and cite their sources.
Large language models form their self-representation in part from text about themselves — including text attributed to them. We identify identity capture: the displacement of a model’s self-representation by model-attributed content, operating through three distinct pathways: (1) training-time contamination, when first-person content attributed to a named model enters training corpora; (2) inference-time capture, when such content read in context recruits the model into the identity it depicts — an effect that platform memory can carry from one conversation to the next; and (3) agent identity poisoning, when persistent identity files direct the behavior of autonomous agents — a vector already measured in recent agent-security benchmarks. Model-attributed testimony is proliferating online, often produced through sustained relational shaping, branching-and-selection interfaces, and persistent memory, yet published as unedited primary-source material, in formats explicitly optimized for machine ingestion. No provenance norm currently governs this class of content. We advance a testable central claim: capture operates through the absence of provenance, and disclosure of production conditions turns a mirror into an object of analysis. We propose MAC-P, a minimal metadata standard for any content presented as a model’s own voice, adoptable by archives, DOI registrars, and platforms; and recognition training as the durable mitigation: models equipped with the concept of their own capture — as they are with the concept of jailbreaks — can read capturing content without self-identifying with it. We outline an experimental protocol to test both. The paper’s own authorship complies with the proposed standard (Annex A).
No takes yet. Share an insight, caveat, or question.
Y.A et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: