Saved in:
Bibliographic Details
Main Author: Bensen, William J.
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2604.22773
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913059458514944
author Bensen, William J.
author_facet Bensen, William J.
contents Large language models (LLMs) are increasingly deployed as partners in knowledge work, where the shared conversational record functions as the decision record that safeguards work continuity. We characterize a class of context failures we term trace mutations, in which distortions enter the shared record while presenting as grounded continuity. We describe two forms: utterance effacement, in which an interlocutor's contribution is re-presented with altered substance, and genitive dissociation, in which a model loses authorship of its own contributions. Using a schematic illustration and two naturalistic anchor cases, we show how these failures differ from confabulation and sycophancy and why they resist ordinary conversational repair. Preliminary cross-model elicitation suggests that at least one such failure is highly camouflaged to contemporary models. We situate the phenomena within grounding and repair theory and discuss implications for tool design.
format Preprint
id arxiv_https___arxiv_org_abs_2604_22773
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Trace Mutation in Human-LLM Dialogue: The Transcript as Forensic and Mitigation Surface
Bensen, William J.
Human-Computer Interaction
Large language models (LLMs) are increasingly deployed as partners in knowledge work, where the shared conversational record functions as the decision record that safeguards work continuity. We characterize a class of context failures we term trace mutations, in which distortions enter the shared record while presenting as grounded continuity. We describe two forms: utterance effacement, in which an interlocutor's contribution is re-presented with altered substance, and genitive dissociation, in which a model loses authorship of its own contributions. Using a schematic illustration and two naturalistic anchor cases, we show how these failures differ from confabulation and sycophancy and why they resist ordinary conversational repair. Preliminary cross-model elicitation suggests that at least one such failure is highly camouflaged to contemporary models. We situate the phenomena within grounding and repair theory and discuss implications for tool design.
title Trace Mutation in Human-LLM Dialogue: The Transcript as Forensic and Mitigation Surface
topic Human-Computer Interaction
url https://arxiv.org/abs/2604.22773