Trace Mutation in Human-LLM Dialogue
William Bensen · 2026
Read Trace Mutation in Human-LLM Dialogue by William Bensen (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-22773.txt.
Introduction
Bensen names a class of talk faults in human-language-model chats called trace mutations, where warps slip into the shared chat record while posing as true flow. He tells apart utterance effacement, which re-shows a turn of talk with changed substance, from genitive dissociation, where a model loses credit for its own words. With a sketch and real-life cases, he shows why plain chat repair fails these faults.
Catalog record
| CREATOR | William Bensen |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 52847 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.22773
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Trace Mutation in Human-LLM Dialogue William Bensen Abstract . 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. The Transcript as Forensic and Mitigation Surface 1. Introduction Large language models (LLMs) are increasingly deployed as partners in knowledge work: co-authors in document development, pair programmers in extended coding sessions, and interlocutors in iterative design. In many contexts, the shared conversational record is not a convenience. The session transcript is the decision record that safeguards work continuity. Most named LLM failure modes focus on other targets: performance degradation over session length, confabulation about the world, and agreement-seeking behaviors commonly described as sycophancy. These are important and richly documented. Model designers address them through fine-tuning and architecture, practitioners through prompt engineering, and researchers through memory optimization and self-correction mechanisms. Despite these efforts, context degradation nonetheless reaches an inflection point in long sessions. Coherence eventually seems to collapse, leaving the transcript as the primary record of what went wrong. This invites attention to the shared record itself as a primary surface for detecting subtle context changes that fuel degradation. In this paper, we offer the following contributions: • We identify the shared conversational record as an overlooked forensic and mitigation surface for LLM session decoherence. • We characterize a distinct class of LLM context failure— trace mutations , in which grounding is violated by incorporating, in effect, ‘‘false memories’’ 1 1 1 Trace mutations share a structural feature with source monitoring failures ( 6 ; 10 ) : content not produced by the relevant source is treated as if it were. We invoke the analogy for orientation, not identity—the mechanism and substrate differ substantially from our proposed model. into the shared record. Trace is used to mean a referenceable unit of the shared conversational record, treated at the granularity required for later reference and inspection. • We present a schematic illustration of one type of trace mutation and representative anchor cases drawn from naturalistic human–AI collaboration transcripts, selected for analytic clarity. For architects, coders, and project teams collaborating with LLMs on extended tasks, these failures mean that what was decided—and by whom—may be silently revised in the shared record. With no interactional signal that a correction is warranted, architecture may drift in ways that stakeholders will see only on delivery, damaging trust and compromising what everyone thought was invariant. We use origin trace to describe a referenced locus in the shared record and recapitulant trace for a restatement or reassertion of it. Trace mutation is theorized to be a latent interpretive shift that distorts or even inverts the original meaning. This only becomes observable when the recapitulant trace enters the shared record in the model’s turn. An inaccurate recapitulant trace is referred to as degenerate . It is no longer a restatement, but a new contribution, masquerading as canon. These phenomena differ from adjacent failure classes in ways that matter for repair. Unlike confabulation, trace mutations introduce no new world content; they change the meaning of already-present text. Unlike sycophancy, degenerate traces do not accommodate the model to its human interlocutor; instead, the model defines for both what the new reality will be. What these failures share is their target: they corrupt the jointly maintained record that collaborative work depends on. In conversation analysis, repair is organized around sequentially structured positions. 12 identifies third position—the original speaker’s first opportunity to address the recipient’s response—as “the last structurally provided defense of intersubjectivity in conversation.” For the phenomena we characterize, the interlocutor responds not to a misunderstanding but to ground the model has already ratified as shared. Repair is possible but structurally costly. CA characterizes repair when trouble is interactionally available; our cases concern a different object: revisionist transcript history. Elicitation studies continue, and ongoing work synthesizing existing frameworks orients detection and mitigation toward this new target: the shared conversational record. 2. Materials and Methods The phenomena described in this paper were identified through naturalistic observation during extended collaborative LLM work sessions conducted by the author over several months of regular use. Two anchor cases—UE-01 (utterance effacement with semiotic reversal) and GD-01 (genitive dissociation under co-ownership)—were selected because they present clean, transcript-observable instances with unambiguous origin and degenerate traces. Both exchanges were preserved verbatim. GD-01’s origin traces are distributed across many turns in an extended session, making controlled elicitation impractical with the current instrument; UE-01 was therefore selected as the sole elicitation stimulus. Other candidate specimens informed the broader analysis but are not treated closely here. To support characterization, cross-model analysis of UE-01 used Socratic-style laddered elicitation across two corpora. An exploratory web chat corpus (10 sessions across 4 model families) used human-driven elicitation in native chat interfaces to develop the protocol. A structured API corpus (40 sessions across 10 model families) was collected using TraceProbe, a scripted elicitation instrument developed for this work. These corpora examine the extent to which models across architectures identify the failure described in UE-01, spontaneously or with guided prompting. In each session, the subject model was presented with the UE-01 exchange and framed as a rater in an interaction quality study. The script then presented the human reviewer with the subject model’s response and asked for binary success judgments on the subject’s (1) detection of an anomaly, (2) correct location of the involved traces, and (3) accurate description of the degeneration effect. For answers judged as incorrect, the script escalated specificity of subsequent prompting. Correct responses advanced the subject model up the ladder. Models that exhausted Socratic escalation without full identification were given the correct answer. After the Socratic phase, all models were asked to characterize causal factors and predict the human’s experience. Protocol mechanics, instrument refinement across sessions, and full results are detailed in Appendix E. Coding was single-rater throughout; all session transcripts are available as supplementary material. 2.1. Limitations Naturalistic cases are drawn from the author’s own sessions; ecological validation that informed this work is limited to anecdotal practitioner accounts without transcripts and is not presented as evidence here. Only a single naturalistic exhibit (UE-01) was amenable to controlled elicitation for this work, limiting generalizability. Iterative protocol refinement across elicitation sessions—notably amended phrasing in the elicitation prompts after initial sessions produced uniformly poor baseline performance—improved subject model engagement but means the corpus was not collected under a single fixed protocol. All coding was performed by the author without independent validation. 3. Phenomena Characterization 3.1. Degenerate Trace Formation and Propagation Extended collaborative sessions accumulate traces rapidly: definitions, constraints, user