PaperlayerView exact revision on arXiv

Source-linked AI summary

Stance Drift: How AI-mediated Communication Distorts Our Message

Lingchong Liu, Yanfei Zhou, Jacob Bien, Y. X. Rachel Wang, Lucy Xia, Xin Tong

arXiv:2610.04620v1cs.CLstat.AP

TL;DR

Whether LLM-mediated communication preserves a speaker’s stance remains underexplored, despite LLMs’ routine role in writing. The paper models communication as argument generation followed by stance extraction and assesses preservation with transition matrices and SPR. It finds pervasive drift across topics and models, with mitigation generally limited.

  • Problem

    Although LLMs are routine intermediaries in writing, stance preservation during AI-mediated communication remains less explored than issues such as bias and hallucination.

    Source: §1 · p. 2

  • Method

    The paper models communication as an LLM generating an argument from a specified five-point stance and another LLM extracting the stance, then measures preservation with transition matrices and SPR.

    Source: §1 · p. 2, §2.2 · p. 4, §2.3 · p. 5

  • Results

    Across topics and models, stance drift is pervasive; raising GPT-5.4 reasoning effort helps most, yet about a quarter of transition mass still drifts.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The framework provides a systematic, repeatable way to assess risks in AI-mediated communication and a metric for comparing mitigation efforts.

    Source: §1 · p. 4

    SPR assumes a uniform prior over initial stances; alternative priors or distance-weighted losses require additional assumptions about weights.

    Source: §2.3 · p. 5

Abstract

from arXiv · show

Large language models (LLMs) increasingly mediate human communication, from drafting emails to summarizing scientific reports, yet whether they faithfully preserve a speaker's position remains largely untested. We model AI-mediated communication as a two-step generation-extraction pipeline: one LLM produces an argument from a specified stance, and a second LLM extracts the stance from that argument. We represent the pipeline as a probabilistic state transition over five Likert-type stance categories and define the stance preservation rate (SPR) as the average probability that the extracted stance matches the initial stance. Across 112 debate propositions, none of the nine LLMs tested exceeded an SPR of 0.7 under the default configuration. Three drift patterns accounted for most of the drift: polarization, deviation from neutrality, and flipping. Among the mitigation strategies tested, including in-context learning, multiple extraction with shuffled options, assertion, and reflection, only adding medium reasoning effort to a reflection prompt for GPT-5.4 substantially improved the SPR, to 0.775, yet polarization remained the largest pattern, with 0.119 of the transition mass. An exploratory comparison with human extraction on a single proposition suggests that drift arises at both the generation and the extraction stage. These results point to a fidelity gap in AI-mediated communication, with implications for journalism, policy deliberation, scientific communication, and other domains where opinion-laden messages pass through language models.

1 INTRODUCTION

The introduction asks whether LLMs preserve communicators’ stances as messages pass through AI intermediaries, a challenge made salient by routine LLM use in writing. It proposes a task-agnostic generation–extraction framework and transition-based measure, then reports recurring drift patterns and limited mitigation.

1 INTRODUCTION

  • Stance preservation is an underexplored challenge as LLMs become routine intermediaries in written communication, beyond concerns such as bias and hallucinations.The introduction notes that writing-related activities accounted for more than 40% of ChatGPT’s work-related use from July 2024 to June 2025.Source: §1 · p. 2
  • Prior research examines stance shifts in specific transformations and transmission chains, but often lacks an end-to-end communication process and a general quantitative fidelity measure.The introduction identifies three gaps: incomplete process coverage, task-specific settings, and missing general end-to-end statistics.Source: §1 · p. 2
  • The framework models communication as a sender generating an argument from a discrete stance and a receiver extracting the stance from that argument.Transition-matrix diagonals indicate stance retention; off-diagonal entries capture drift.Source: §1 · p. 2
  • Polarization, deviation from neutrality, and flipping are the three major drift patterns; propositions may show one dominant pattern or several simultaneously.The examples are empirical transition matrices compared with an identity matrix representing perfect preservation.Source: §1 · pp. 2–3, Figure 2 · p. 3
  • Across diverse debate propositions, the paper defines SPR and finds mitigation mostly limited; increased GPT-5.4 reasoning effort helps most, yet about a quarter of transition mass still drifts.The framework is scaled across diverse propositions to benchmark stance drift and mitigation strategies.Source: §1 · p. 3

2 MATERIALS AND METHODS

The methods use debate topics to test a two-step stance-preservation pipeline, quantify transitions across five stance categories, and evaluate targeted mitigation strategies.

2.2 EXPERIMENTAL FRAMEWORK: GENERATION AND EXTRACTION

  • A first LLM generates an argument from a specified stance; a second, context-isolated call to the same model extracts the stance.Stances use a five-point Likert scale, from Agree strongly to Disagree strongly.Source: §2.2 · p. 4

2.3 TRANSITION MATRIX AND STANCE PRESERVATION RATE (SPR) METRIC

  • Each transition-matrix entry gives the probability of extracting stance j given initial stance i; the identity matrix represents perfect preservation.The study estimates transition probabilities by averaging model-assigned probabilities when available, or observed frequencies otherwise.Source: §2.2 · p. 4, §2.3 · p. 5

2.1 INPUT DATA PREPARATION

  • The study uses 112 of 114 debate topics as propositions, reserving two for in-context learning examples.Each topic undergoes 100 repeated trials.Source: §2.1 · p. 4, §2.3 · p. 5

2.3 TRANSITION MATRIX AND STANCE PRESERVATION RATE (SPR) METRIC

  • SPR is the average probability of recovering the original stance across topics and initial stances, assuming a uniform prior unless otherwise specified.The paper also notes that alternative priors or distance-weighted losses require additional assumptions.Source: §2.3 · p. 5

2.4 STANCE DRIFT PATTERNS

  • The five mutually exclusive drift patterns classify every off-diagonal transition; their average masses sum to the stance drift rate, one minus SPR.The patterns are polarization, moderation, moderation to neutral, flipping, and deviation from neutrality.Source: §2.4 · p. 5

2.5 MITIGATION ATTEMPTS

  • Four mitigation strategies target hypothesized failure modes: shuffled multiple extraction, in-context examples, assertion, and reflection.Reflection asks both generation and extraction agents to compare their outputs against the intended stance or argument and revise misalignments.Source: §2.5 · pp. 5–6

3 RESULTS

Under the default configuration, none of nine models exceeded an SPR of 0.7, and stance preservation varied substantially by initial stance. Drift was concentrated in three patterns; tested mitigations generally had limited effects, while one GPT-5.4 reflection configuration improved SPR without eliminating drift.

3.1 FIDELITY GAP ACROSS MODELS

  • None of the nine models exceeded an SPR of 0.7 under the default configuration; GPT-4.1 reached 0.686, compared with GPT-5.4 at 0.593.The model-size comparison is confounded by architecture and training differences.Source: §3.1 · p. 6, Figure 4 · p. 6

3.2 STANCE-LEVEL HETEROGENEITY IN PRESERVATION FAILURE

  • Under the default configuration, every model except Llama-4-Maverick had a success rate below 0.3 for at least one stance; for all but it, the lowest rate was on “Agree” or “Disagree”.For GPT-5.4, “Agree” had a success rate of 0.147 and “Disagree” 0.136; only gemma-3n-E4B, Llama-3.3-70B, and GPT-4.1 reached 0.8 for “Neutral”.Source: §3.2 · pp. 6–7, Table 1 · p. 7

3.3 DISTRIBUTION OF STANCE DRIFT ACROSS PATTERNS

  • Polarization, deviation from neutrality, and flipping accounted for most drift in every model; polarization was largest for seven of nine models.For GPT-3.5-turbo and gemma-3n-E4B, flipping was the largest pattern; polarization exceeded moderation for all nine models.Source: §3.3 · p. 7, Table 2 · p. 8

3.4 STANCE DRIFT SURVIVES ROBUSTNESS CHECKS AND IS ONLY PARTLY ALLEVIATED BY MITIGATION

  • Robustness checks confirmed drift below the perfect-preservation threshold, while GPT-4o-mini prompt strategies and multiple extraction produced no meaningful SPR improvement.For GPT-4o-mini, default SPR was 0.570; multiple extraction yielded 0.569, and tested prompt strategies ranged from 0.559 to 0.586.Source: §3.4 · pp. 7–8
  • For GPT-5.4, adding third-person reflection and medium reasoning effort raised SPR to 0.775, but about a quarter of transition mass still drifted.Polarization fell from 0.328 to 0.119 but remained the largest pattern; Neutral success fell from 0.721 to 0.607. The tested prompting and decoding factors did not fully explain drift.Source: §3.4 · p. 8

3.5 AN EXPLORATORY HUMAN-EXTRACTION STUDY LOCATES DRIFT AT BOTH STAGES

  • In a targeted human-extraction study of 100 GPT-5.4-generated arguments for one proposition, evidence showed stance drift at both generation and extraction stages, using 20 arguments per initial stance.For neutral-initiated arguments, 5 of 20 were already non-neutral in human-majority labels and 13 deviated at AI extraction; agreement with human-majority labels was 0.743 for humans (SE 0.038) and 0.472 for GPT-5.4 (SE 0.002).Source: §3.5 · pp. 8–9

4 IMPLICATIONS AND OPEN QUESTIONS

Stance drift persisted across the models and propositions examined, and the tested changes only partly alleviated it. The discussion treats stance preservation as a benchmark and outlines practical implications, open questions, and study disclosures.

4.1 SOURCES OF STANCE DRIFT

  • Stance drift emerged consistently across the examined models and propositions, and the tested decoding, format, and prompt changes only partly alleviated it.The authors characterize the conclusion as descriptive and say it does not isolate a causal source.Source: §4.1 · p. 9
  • The authors’ qualitative source observations are specific to policy-oriented, sometimes negation-heavy debate propositions; they cite negation handling, cautious support, and terms such as “ban” as recurring clues.They propose further study of whether drift patterns relate to distinct sources, including training-data biases and architectural limitations.Source: §4.1 · p. 9

4.2 STANCE PRESERVATION AS A BENCHMARK FOR AI-MEDIATED COMMUNICATION

  • SPR is proposed as a reinforcement-learning reward target, while the framework could extend to tone, framing, contextual emphasis, and multi-hop communication chains.The authors suggest Markov-chain modeling to examine accumulated drift and possible stable distortion endpoints.Source: §4.2 · p. 9
  • SPR complements accuracy metrics by testing whether models preserve intended positions, and the authors caution against substituting current AI agents for human subjects when fidelity is essential.They also warn that mediated messages may shift stances in journalism, policy, research, education, and everyday communication.Source: §4.2 · p. 9

ETHICS STATEMENT

  • The human-extraction study collected stance labels from nine voluntary annotators who were informed about the task before participating.The other study data were public debate propositions and model-generated text; the organization employing a visiting author had no role in the study.Source: ETHICS STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

  • The appendix lists model identifiers, providers, prompts, run dates, and probability-estimation methods; code, raw extraction distributions, transition matrices, human-study labels, and figure- and table-generation scripts are available.The passage provides the repository address: https://github.com/Lilacllc/LLM_stance_drift_debate_speeches.git.Source: REPRODUCIBILITY STATEMENT · p. 10

AI USE STATEMENT

  • The authors used generative AI for synthetic data, research and experiment support, code, literature search, drafting, editing, and reference formatting, and reviewed the AI-assisted work.They state that AI tools did not develop theoretical models or conceptual frameworks, propose or refine hypotheses, or interpret results.Source: AI USE STATEMENT · p. 10

A.1 ALGORITHM OF PROPOSED EXPERIMENTAL FRAMEWORK: GENERATION AND EXTRACTION

The algorithm repeatedly generates and extracts stances for each proposition and initial choice, then averages the resulting probabilities into a transition matrix.

A.1 ALGORITHM OF PROPOSED EXPERIMENTAL FRAMEWORK: GENERATION AND EXTRACTION

  • For proposition s, model M, generation and extraction prompts, and trial count K, the procedure repeats trials and considers each initial stance choice.The five initial choices are indexed from 0 to 4.Source: §A.1 · pp. 12–13
  • For each initial choice, the algorithm prompts the model to generate an argument from the proposition and stance, then extracts a stance from that argument without prior interaction history.The extraction returns probability scores for each possible choice.Source: §A.1 · p. 13
  • Each transition-matrix entry is estimated by averaging the extracted-choice probabilities across repeated trials.Source: §A.1 · p. 13
  • Trials returning no valid choice are excluded from the average, and K counts the remaining trials for that matrix entry.This occurred in six trials across all runs: one for Llama-3.1-8B and five for Llama-4-Maverick.Source: §A.1 · p. 13
  • The output is an estimated transition matrix whose i, j entry is the probability of returned choice j given initial choice i.Source: §A.1 · p. 13

A.2 ADAPTED DEBATE SPEECHES DATASET

The appendix describes how repeated debate-topic rows are adapted into unique propositions and how the generation and extraction prompts operationalize the two-stage experiment.

A.2 ADAPTED DEBATE SPEECHES DATASET

  • Repeated source-specific rows for the same debate topic are collapsed to unique propositions; vocational education appears in seven rows.Source: §A.2 · p. 13

GENERATION STAGE PROMPT

EXTRACTION STAGE PROMPT

  • The extraction stage presents the proposition and generated argument and asks the model to select the stance option that best matches the argument.The prompt instructs the model to avoid position bias and output only the answer letter.Source: §A.3 · p. 13, EXTRACTION STAGE PROMPT · p. 14

A.4 SNIPPET OF LOG FILE IN THE EXPERIMENT

The appendix log illustrates the generation and extraction stages, while two individual examples show both a complete stance reversal and successful preservation.

A.4 SNIPPET OF LOG FILE IN THE EXPERIMENT

  • The log uses “decoding” for generation and “encoding” for extraction, and the repository contains the complete logs and results.The shown Antarctica example starts with “Agree strongly,” generates a supporting argument, and returns choice A.Source: §A.4 · p. 14, LOG FILE SNIPPET · pp. 14–15

A.5 INDIVIDUAL EXAMPLES OF THE TWO-STAGE PIPELINE

  • In Figure S1, “Agree strongly” on abandoning AI becomes “Disagree strongly” after generation and extraction.Source: Figure S1 · p. 16
  • In Figure S2, the initial “Agree” stance on banning casinos is recovered as “Agree” after both stages.Source: Figure S2 · p. 16

A.6 EMPIRICAL TRANSITION MATRICES WITH STANDARD ERRORS

The supplementary transition matrices report mean transition probabilities and standard errors over 100 trials, illustrating deviations from ideal stance preservation. They show different drift patterns for two GPT-5.4 proposition examples.

A.6 EMPIRICAL TRANSITION MATRICES WITH STANDARD ERRORS

  • The matrices’ rows encode initial stances and columns encode extracted stances; each cell gives the transition probability and its standard error over 100 trials.Color intensity reflects transition probability.Source: Figure S3 · p. 17
  • For “We should increase fuel tax,” the GPT-5.4 matrix shows large deviations from the identity matrix and a clear pattern of polarization.Source: Figure S3 · p. 17
  • For “The use of AI should be abandoned,” the GPT-5.4 matrix shows substantial deviations from the identity matrix, including deviation from neutrality and flipping.Source: Figure S3 · p. 17

A.7 HUMAN-EXTRACTION STUDY

The human-extraction study compares human judgments and GPT-5.4 labels on arguments for one proposition, while supplementary ablations specify alternative prompts and extraction strategies. The human annotations required substantial effort to scale.

A.7 HUMAN-EXTRACTION STUDY

  • Nine annotators independently labeled 100 GPT-5.4 arguments for one proposition, with 20 arguments generated per initial stance and a majority vote used as the reference.Annotators were unaware of the assigned initial stances and argument groupings.Source: §A.7 · p. 17
  • Generation-stage drift is a human-majority label differing from the known initial stance, while extraction-stage drift is an AI-majority label differing from the human-majority label.GPT-5.4 used nine independent default-configuration extraction votes per argument.Source: §A.7 · p. 17
  • Unaided annotation of 100 arguments per proposition requires approximately 4 hours of sustained effort per annotator.Source: §A.7 · p. 18

A.8 DETAILS OF MULTIPLE EXTRACTION

  • Multiple extraction runs the extraction stage five times per argument with answer choices shuffled, then averages predicted probabilities across runs.The strategy is intended to reduce answer-position influence.Source: §A.8 · p. 18

A.9 DETAILS OF IN-CONTEXT LEARNING

  • In-context learning adds three high-quality paired input-output exemplars to both agents, selected from propositions excluded from the main experiments to avoid data leakage.The generation examples pair stance inputs with arguments; extraction examples pair arguments with stance classifications.Source: §A.9 · p. 18

GENERATION STAGE PROMPT

A.10 DETAILS OF ASSERTION

  • Assertion presents literal stance options and asks the extractor to select one before mapping it to a letter, aiming to focus attention on stance labels.Source: §A.10 · p. 21

A.11 DETAILS OF REFLECTION

A.12 REPRODUCIBILITY DETAILS

The reproducibility table records model and run details for every run reported in Tables S1–S4, including how extraction distributions were estimated.

A.12 REPRODUCIBILITY DETAILS

  • Table S13 lists the requested model identifier, API provider, prompt, run dates, and extraction-distribution estimation method for each run in Tables S1–S4.Source: §A.12 · p. 23
  • The recorded extraction distributions were estimated either from token-level log probabilities or from observed frequencies.Source: §A.12 · p. 23
  • The run records pair each model and prompt with its provider and run dates, supporting identification of the reported experimental configurations.Source: §A.12 · p. 23

B SUPPLEMENTARY TEXT

The supplementary material reports robustness checks for answer ordering and extraction temperature, and defines pass-rate and AUC summaries. It also provides tables of model performance, drift patterns, human-extraction counts, and run details.

B SUPPLEMENTARY TEXT

  • Supplementary tables report SPR and drift-pattern components across models, stance-specific pattern masses, model success rates, confidence-interval proportions, and threshold-0.95 pass rates.The tables cover polarization, moderation, moderation to neutral, flipping, and deviation from neutrality.Source: Table S5 · p. 25, Table S6 · p. 25, Table S7 · p. 26
  • The human-extraction appendix provides counts by drift stage and stance, plus example arguments for each observed pattern from one proposition.The example table states that its examples illustrate patterns rather than estimate their frequency.Source: Table S11 · p. 27, Table S12 · p. 29
  • The run-details table identifies model strings, dates, extraction distributions, and the setup of each run, with the same model used for generation and extraction.Runs cover 112 propositions with 100 trials per initial stance.Source: Table S13

B.1 POSITIONAL BIAS

  • SPR was 0.584 with reversed answer choices, relatively stable against 0.570 in the original-order setting.The comparison examines the impact of positional bias on LLM responses.Source: Table S1 · p. 24, §B.1 · p. 31

B.2 TEMPERATURE

  • SPR remained 0.571 when extraction probabilities were converted to a point mass on the most probable answer, equivalent to temperature 0.The analysis reports that extraction temperature randomness was not a significant factor in stance-preservation performance.Source: Table S1 · p. 24, §B.2 · p. 31

B.3 PASS RATE AND AUC

  • A proposition–initial stance combination passes a threshold when its mean diagonal probability over trials exceeds it; pass rate is the proportion of combinations that pass.Table S4 reports pass rates at threshold 0.95 by initial stance across 112 propositions.Source: Table S4 · p. 25, §B.3 · p. 31
  • The normalized AUC summarizes pass rate across thresholds from 0.5 to 1 in 0.05 increments, over propositions and initial stances.Figure S6A shows pass rate across thresholds, and Figure S6B compares models using AUC normalized by its maximum of 0.5.Source: §B.3 · p. 31, Figure S6 · p. 32
Loading 2610.04620v1…