PaperlayerView exact revision on arXiv

Source-linked AI summary

"If You're Not Doing It, Somebody Else Is": Active Negotiation and the Invisible Labor of Sustained LLM Use

Matt Viana, Patrick Erickson, Shomir Wilson, Dana Calacci

arXiv:2609.30699v1cs.HC

TL;DR

Existing adoption frameworks do not explain how users continue using LLMs despite recognizing their costs. Based on interviews with graduate student workers, the paper develops the Active Negotiation framework and frames sustained use as invisible labor. It concludes that formalizing this process provides a foundation for understanding user behavior.

  • Problem

    Existing adoption frameworks offer little account of how users sustain LLM use despite recognized costs, a gap relevant to understanding continued reliance on a distrusted tool.

    Source: §1 · p. 1, §2.3 · p. 4

  • Method

    The authors draw on 36 interviews with graduate student workers to develop the Active Negotiation framework.

    Source: §10 · p. 21

  • Results

    Formalizing continued LLM use as a process makes visible the labor involved in AI adoption and provides a foundation for understanding user behavior.

    Source: §10 · p. 22

  • Takeaways & Limitations

    The framework offers a way to understand user behavior through the often-hidden labor of sustained AI adoption.

    Source: §10 · p. 22

    The framework is grounded in 36 graduate student worker interviews, and the authors flag its validation as bounded by field and institutional scope.

    Source: §9 · p. 21, §10 · p. 21

Abstract

from arXiv · show

Large language models (LLMs) have become fixtures of academic work even as their users describe them as degrading their writing, thinking, and skills. Dominant adoption frameworks read continued use as evidence of satisfaction, and cannot explain continued use of a distrusted tool. We interviewed 36 graduate student workers, balanced between English-as-a-foreign-language (EFL) and non-EFL speakers, and introduce the Active Negotiation framework: a model of sustained LLM use as a recurring cycle of risk, mitigation, and justification. A failure surfaces a risk, mitigation labor addresses it, and a justification renders the residual risk tolerable until the next failure reopens the cycle. The cycle runs across three dimensions: practical, auditing output; internal, auditing one's own cognition and identity; and social, managing how peers and institutions perceive use. EFL participants invoke linguistic parity as a further justification. We reframe continued adoption as compliance sustained by invisible labor.

1 INTRODUCTION

LLM use persists despite users’ recognition of its costs, while existing adoption frameworks struggle to explain continued use of a distrusted tool. The paper frames that use as recurring negotiation under structural pressure, involving practical, internal, and social work.

1 INTRODUCTION

  • Adoption frameworks explain use through usefulness, ease, or satisfaction, but not how users sustain engagement when they judge a tool negatively or face pressures beyond their control.The paper describes refusal as costly when it means falling behind colleagues and peers.Source: §1 · pp. 1–2
  • The study investigates how users navigate continued LLM use despite recognized costs, drawing on 36 semi-structured interviews with graduate student workers.The sample was stratified by English proficiency.Source: §1 · p. 2
  • The Active Negotiation framework models sustained LLM use as a recurring cycle in which failures surface risks, users mitigate them, and justifications make residual risks tolerable.New failures reopen the cycle as users’ mental models develop.Source: §1 · p. 2
  • The cycle spans practical assessment of technical failures, internal conflict over efficiency and cognitive costs, and social management of stigma, disclosure, and institutional perceptions.Participants may negotiate all three dimensions simultaneously.Source: §1 · p. 2
  • The paper reframes adoption as a coercive condition and uses EFL participants’ structural pressures to show how inequality shapes user–system negotiation.EFL participants invoke linguistic parity as a justification for continued use.Source: §1 · p. 2

2 RELATED WORK

Prior work documents widespread and varied LLM use, but leaves open how users sustain engagement despite its costs. The paper connects that process gap to limits in adoption theories, domestication accounts, and research on shifting mental models.

2.1 Understanding User–LLM Interaction in Practice

  • Studies of LLM use document actions and experiences, but interaction logs cannot explain why users act as they do or how they interpret use; process-level accounts of sustaining engagement remain limited.The Active Negotiation framework addresses how users reconcile recognized flaws with continued reliance.Source: §2.1 · pp. 2–3

2.2 Theoretical Frameworks for Technology Engagement

  • Adoption and coping frameworks have difficulty accounting for ongoing use when short-term task success coexists with dissatisfaction and users must repeatedly reconcile recognized costs.The paper characterizes that recurring reconciliation as invisible labor that masks system deficits and is hidden from standard interaction logs and adoption measures.Source: §2.2 · p. 3

2.3 Domestication Theory

  • Because LLMs remain unstable while refusing them may be structurally unfeasible, the framework shifts domestication theory’s focus to the cyclical labor of sustaining engagement.The cited account contrasts LLM instability with technologies that stabilize after domestication.Source: §2.3 · p. 4

2.4 Mental Models, Anthropomorphism, and Forgiveness

  • The framework proposes that users alternate between instrumental and anthropomorphic framings to mitigate encountered risks, using perception itself to support continued reliance.Prior work leaves open how and why users’ shifting mental models sustain long-term engagement.Source: §2.4 · pp. 4–5

3 METHODS

The study draws on interviews with 36 graduate students, balanced across EFL and non-EFL speakers, and analyzes their accounts to develop the Active Negotiation framework. Participants reported broad LLM use, especially for writing, summarization, and quick questions.

3.1 Target Population and Recruitment

  • The sample comprised 36 graduate students: 21 non-EFL and 15 EFL speakers, recruited with purposive sampling after a survey and stopping at thematic saturation.Participants were targeted as both students and high-skilled workers facing competing productivity and accuracy demands.Source: §3.1 · p. 5

3.2 Interview Process and Guide Development

  • Interviews explored practical interaction, internal trade-offs, social dynamics, and trust calibration, with questions about reliance, stigma, typical uses, and model failures.Eleven pilot interviews informed the guide but were excluded from the final analysis.Source: §3.2 · pp. 5–6

3.3 Data Analysis and Codebook Development

  • Researchers first coded transcripts using a collaboratively refined codebook, then reviewed coded data inductively; the cyclical Risk–Mitigation–Justification framework emerged from recurring patterns.Because coding was consensus-based rather than independent, the researchers did not report inter-rater reliability statistics.Source: §3.3 · p. 6

3.4 Positionality and Ethical Considerations

  • The research team identified as active users and designers of generative AI and bracketed their assumptions to center participant experiences; the study had IRB approval and anonymized data.Participants provided informed consent.Source: §3.4 · p. 6

3.5 Participant LLM Use

  • In the sample, 35 of 36 participants actively used LLMs; writing assistance and summarization were each reported by 32 participants, while 30 reported using them to answer quick questions.Source: §3.5 · p. 6

4 PARTICIPANTS ENCOUNTERED RISKS WHEN USING LLMS

Participants described practical, internal, and social risks of LLM use that emerged on different timescales. The section traces how a practical failure made accumulated cognitive costs visible for one participant, alongside risks involving output harm and social context.

4 PARTICIPANTS ENCOUNTERED RISKS WHEN USING LLMS

  • Practical failures appeared early, internal costs compounded gradually, and social risks depended on context and relationship stakes.A disclosure routine among friends could carry professional risk in lab meetings.Source: §4 · p. 7

4.1 P33: A Longitudinal View of Risk Accumulation

  • P33 said repeated cognitive offloading in R statistics left them forgetting basic coding concepts, including how to set up a data frame; the cost became visible only in retrospect.Source: §4.1 · p. 8
  • After two back-to-back failures, P33 shifted from default delegation to treating the model as a last resort, with Google Scholar and primary documentation first; lab guidelines prohibited LLM use for data input and writing generation.Source: §4.1 · pp. 7–9

4.2 Practical Risks

  • Participants treated hallucination, misinterpretation, and bias as common failures, but distinguished acute harm that could damage work, safety, or professional standing.Unlike a failed output that creates redundant labor, harmful output can have consequences that re-prompting cannot undo.Source: §4.2 · p. 9

4.3 Internal Risks

  • Internal risks such as skill atrophy, deepened reliance, and ethical concerns accumulated over time and often became legible only in retrospect or when the tool was absent.Technical performance cannot offset the moral cost participants associated with ideological, environmental, and privacy concerns.Source: §4.3 · p. 9

4.4 Social Risks

  • Social risks arose from interpersonal and institutional dynamics: stigma put authorship in question, while absent policy made disclosure costs difficult to calculate.Participants described disclosure as context-dependent, and institutional silence could drive concealment as a lower-risk strategy.Source: §4.4 · p. 10

5 PARTICIPANTS MITIGATED RISKS THROUGH BEHAVIORAL AND COGNITIVE LABOR

This section examines how participants sustained LLM use through behavioral and cognitive mitigation, including practical, internal, and social strategies. It also traces how social mitigation varied across settings, as users followed different rules in different contexts.

5.1 P20: Mitigation Across Dimensions

  • P20 managed three simultaneous cycles: adapted prompting after failures, switching between instrumental and human-like views, and calibrating disclosure to the setting.The same user applied all three strategies with the same tools.Source: §5.1 · p. 10
  • P20 used LLMs to prepare for unfamiliar bureaucratic systems and express academic ideas in English, but this language support diminished as their English improved.Other EFL participants showed a contrasting trajectory in which reliance on the tool eroded English fluency.Source: §5.1 · p. 11

5.2 Practical Mitigation

  • Participants responded to failures with iterative rephrasing, persona prompting, and chunking, while treating verification as necessary before relying on outputs.Users only gained the promised efficiency after spending effort to verify results.Source: §5.2 · p. 12
  • Only one participant abstained entirely, while 22 others abandoned LLMs task by task; use saved time only when prompting, checking, and repair cost less than the replaced workflow.Friction-based abandonment limited exposure to verification costs.Source: §5.2 · p. 12

5.3 Internal Mitigation

  • 35 of 36 participants reported boundary-setting, including limits on high-stakes writing, personal data, cognitive-development tasks, and emotionally sensitive conversations.Users had to decide whether each task crossed their boundaries, making that decision itself a form of labor.Source: §5.3 · p. 12
  • Participants switched between functional and human-like framings to reduce emotional friction, and used the LLM for socially difficult interactions because it seemed non-judgmental.One participant used it to draft a message ending a friendship, avoiding the immediate stress of composing it.Source: §5.3 · pp. 12–13

5.4 Social Mitigation

  • Social mitigation relied on selective disclosure, presenting independent competence while concealing LLM assistance; participants adjusted disclosure to audiences and relationship stakes.P20 disclosed use casually with friends but stayed silent about it in the lab.Source: §5.4 · p. 13
  • The same user followed different rules in different settings as a form of social mitigation.Source: §5.4 · p. 13

6 PARTICIPANTS JUSTIFIED CONTINUED USE DESPITE UNRESOLVED RISKS

This section examines how participants justified continued LLM use despite unresolved risks, including through practical, internal, and social arguments and, for EFL participants, linguistic parity and responses to unequal professional conditions and language-related risks. It also covers how these justifications may support either movement toward independence or ongoing reliance.

6 PARTICIPANTS JUSTIFIED CONTINUED USE DESPITE UNRESOLVED RISKS

  • Justification is the labor of tolerating risks that mitigation does not resolve, through internal arguments that make continued engagement acceptable until another failure surfaces.The section distinguishes justification from behavioral mitigation: when adjustments cannot eliminate a recognized cost, users rationalize continued use.Source: §6 · p. 13

6.2 Practical Justifications

  • Participants commonly described an efficiency calculation in which residual risks were an acceptable trade-off for labor saved, and task compartmentalization was the most widely reported justification (33 of 36 participants).Source: §6.2 · p. 14

6.3 Internal Justifications

  • Internal justifications protected researchers’ self-concepts by prioritizing well-being and normalizing LLM reliance, while users could condemn public overreliance yet exempt their own practice.When reliance led P33 to lose procedural skills, their initial justification collapsed and they returned to traditional search; lab guidelines helped reduce the burden of self-policing.Source: §6.3 · p. 15

6.4 Social Justifications

  • Social justifications drew on peer normalization and othering, while workload, competition, and English-dominated publishing framed use as a response to external pressures.Participants cited widespread use or others’ misuse to validate their own practices; some described breaches of personal boundaries under acute workload pressure, and EFL participants invoked linguistic parity as a structural justification.Source: §6.4 · pp. 15–16
  • EFL participants invoke parity to offset an uneven playing field and the lack of institutional writing support.P14 describes having no writing tutor or checkpoint.Source: §6.4 · p. 16
  • For some EFL participants, avoiding professional penalties from language errors and ensuring an appropriate email tone outweigh abstract risks of LLM reliance.Participants describe using the tool for high-stakes emails and to avoid seeming rude.Source: §6.4 · p. 16
  • Some EFL participants use LLMs as temporary scaffolding toward less reliance, while others lose confidence and accept reliance as permanent in English-dominated spaces.The first trajectory involves studying and practicing English; the second is illustrated by a participant who now checks ChatGPT despite prior confidence.Source: §6.4 · p. 16

7 THE ACTIVE NEGOTIATION FRAMEWORK

The Active Negotiation framework describes continued LLM engagement as requiring cognitive and behavioral labor. It models that engagement as a recurring risk–mitigation–justification cycle across practical, internal, and social dimensions.

7.1 The Cycle: Risk, Mitigation, Justification

  • A salient failure surfaces perceived risk; users mitigate immediate problems through behavioral or cognitive adjustments, then justify unresolved residual risk.Source: §7.1 · p. 16, Figure 2 · p. 17
  • Justifications resolve a cycle only temporarily: novel failures reopen negotiation, though prior boundaries and rationalizations can make later cycles quicker.Source: §7.1 · p. 17

7.2 Three Dimensions

  • Negotiation spans practical output and error management, internal cognition and identity over time, and social perceptions by peers and institutions.Source: §7.2 · p. 17, Figure 3 · p. 18
  • The dimensions are interdependent: a single interaction can involve correcting an output, questioning overreliance, and managing what others learn about an error.Source: §7.2 · p. 17

8 DISCUSSION

The discussion frames continued LLM use as compliance sustained by active labor under academic and peer pressures, and examines how users negotiate those pressures through shifting perceptions and mitigation. It considers design support for user boundaries and treats disclosure as both a technical provenance problem and an institutional norm-setting problem.

8.1 Coercion and the Conditions of Negotiation

  • Use persists under linguistic, workload, and institutional pressures, with users investing active labor to sustain professional agency rather than simply abandoning LLMs.Only one participant abstained; the remaining 35 developed elaborate defenses despite failures and friction.Source: §8.1 · p. 18

8.2 LLMs Turn Misrecognition into Active Labor

  • LLMs exposed this process to student participants, who did active labor making it deliberate work; with no departmental policy (16/36), disclosure, authorship, and acceptable-use questions fell to individual students.Source: §8.2 · p. 19

8.3 The Practice of Perception Switching

  • Participants shifted among simultaneous instrumental and anthropomorphic framings, using perception switching to mitigate friction and sustain trust despite failures.The framing could change across tasks or over time, altering what use seemed acceptable.Source: §8.3 · pp. 19–20

8.4 Mitigation and Deepening of Engagement

  • Mitigation labor may deepen engagement through investment in customized workflows, while sycophancy and anthropomorphic framing can build trust and make failures easier to forgive.Forgiveness functions as a justification in which users excuse model failures as acceptable, human-like mistakes.Source: §8.4 · p. 20

8.5 Design Implications

  • Design should preserve user-controlled negotiation history and honor user-drawn boundaries, while leaving structural pressures and disclosure questions to institutions.Boundary-setting was the most widespread mitigation, used by 35 of 36 participants, but each boundary required ongoing self-policing.Source: §8.5 · pp. 20–21
  • P22’s departmental committee shows a user building disclosure norms their institution failed to supply.Source: §8.5 · p. 21
  • Work on content provenance and attribution addresses the technical layer, while the social weight of disclosure markers is linked to participants’ current institutional norms.Source: §8.5 · p. 21

9 LIMITATIONS

The study’s claims are bounded by its single-university interview sample, the academic settings in which the framework was developed, and the methods used to collect and interpret participants’ accounts. The authors also identify limits in their binary language grouping and the rapidly changing context of LLM use.

9 LIMITATIONS

  • The claims draw on 36 graduate-student interviews at one U.S. research university, reflecting one configuration of workload norms, advisor relationships, and institutional policies.The interviews were conducted between May and September 2025.Source: §9 · p. 21

Boundary Conditions and Institutional Scope

  • The framework is currently grounded in fields such as academia, where output serves as a proxy for individual capability and a credential for competence.The authors say it remains underexplored in production-only contexts where output is valued as a product and disclosure carries no social or professional cost.Source: §9 · p. 21

Linguistic Nuance

  • The binary EFL/non-EFL grouping obscures linguistic diversity and differences in how well current LLMs support particular languages.The authors note that discrepancies vary by language, including Portuguese versus Burmese.Source: §9 · p. 21

Methodological Reach and Bias

  • Semi-structured interviews capture what participants can consciously articulate, making them ill-equipped to measure unconscious effects such as sycophancy and forgiveness architecture.Self-selection may over-sample people with strong feelings about LLMs, while stigma may prompt performed competence or concern.Source: §9 · p. 21

Timeline Specificity

  • The interview window closed in September 2025, while generative AI models, institutional policies, and user norms are shifting rapidly.A single author conducted all interviews, and the study did not conduct member checking.Source: §9 · p. 21

10 CONCLUSION

The conclusion presents Active Negotiation as a framework for understanding why users continue using LLMs despite recognizing their flaws. It describes continued use as invisible labor under external pressure, rather than evidence of satisfaction.

10 CONCLUSION

  • The paper investigates how users sustain LLM engagement despite recognizing tool flaws, drawing on 36 interviews with graduate student workers to present Active Negotiation.The framework treats continued use as a recurring negotiation rather than a settled endorsement.Source: §10 · p. 21
  • The framework models continued use as a cycle of risk, mitigation, and justification across practical, internal, and social dimensions.Users perform invisible labor to sustain use, which the paper describes as compelled by external pressures.Source: §10 · pp. 21–22
  • Interpreting continued adoption as satisfaction mistakes structural compliance for user preference, because users may negotiate reliance without endorsing the tool.The conclusion says institutions leave users to navigate ambiguities of legitimate use alone, while current interfaces absorb and deepen reliance.Source: §10 · p. 22
  • The authors present formalizing this process as a foundation for understanding user behavior and the invisible labor of AI adoption.The conclusion frames the contribution as making that labor visible.Source: §10 · p. 22

A INTERVIEW GUIDE

The interview guide used a semi-structured format, with optional probes, to elicit graduate students’ LLM experiences across tasks, language use, failures, trust, and disclosure. It also included consent, recording-permission, and compensation procedures.

A INTERVIEW GUIDE

  • The semi-structured guide used probes as needed and framed the interview to welcome different kinds of AI use and experience.Participants were told there were no right or wrong answers and that uses for school, work, or fun were of interest.Source: A INTERVIEW GUIDE · p. 24

Usage Context and Tasks

  • Questions covered systems used, common tasks, prompt languages, code-switching, rephrasing, and reasons for choosing a language.The task checklist included summarizing, drafting, brainstorming, translation, and programming, among other uses.Source: A INTERVIEW GUIDE · p. 24

A INTERVIEW GUIDE

  • The guide asked about changing trust and reliance over time, peers’ use, disclosure, and comfort discussing LLMs; consent, recording permissions, and compensation procedures surrounded the questions.The disclosure prompt asked whether participants tell people when they used an LLM.Source: A INTERVIEW GUIDE · pp. 24–25

Failures, Bias, and Harm

  • Failure probes asked about strange, false, biased, harmful, or inappropriate outputs, with a fallback example available if participants had none to share.The guide also asked how participants would change the model.Source: A INTERVIEW GUIDE · pp. 24–25

B CODEBOOK

The appendix codebook lists deductive coding categories and their definitions, followed by individual codes applied to transcripts. The categories cover interaction strategies, failures and trust, internal negotiation, social and ethical context, language and culture, and users’ mental models of LLMs.

B CODEBOOK

  • The codebook presents top-level categories with definitions and then lists individual transcript codes; Tables 2–4 report a subset using the paper’s terminology.Table captions and accompanying source files record how table terminology corresponds to the codebook labels.Source: B CODEBOOK · p. 25

Active Negotiation & User Strategies

  • Interaction-strategy codes include iterative prompting, scaffolding or chunking, persona prompting, and verification against other sources or models.Verification examples include checking Google Scholar, personal knowledge, or running code.Source: B CODEBOOK · p. 25

Salient Failures & Trust Erosion

  • Failure and trust codes distinguish hallucinated content, context misinterpretation, harmful advice, inappropriate tone, and failures users do not notice.The codebook also tracks users’ efficiency-versus-learning tensions, fear of cognitive atrophy, emotional reliance, and personal boundaries.Source: B CODEBOOK · p. 26

The Internal Debate: Efficiency vs. Autonomy

  • Internal-debate codes cover efficiency versus learning, cognitive atrophy, emotional reliance, overreliance, personal boundaries, and task-based or peer-based justification.The categories also include contradictory trust allocation across tasks.Source: B CODEBOOK · p. 26

Social & Ethical Dimensions

  • Social and ethical codes include academic stigma, environmental influence, data privacy, and environmental or societal concerns such as energy use and job displacement.Language-and-culture codes cover code-switching, translation, language bridging, linguistic failures, and perceived linguistic parity for some EFL users.Source: B CODEBOOK · p. 27

Anthropomorphism vs. Instrumentalism

  • The codebook distinguishes instrumental and anthropomorphic views of LLMs, explicit switching between them, third-party accounts, and non-exclusive model-selection categories.Perception Switching counts participants coded in both appliance and person categories, whether or not they remarked on the shift.Source: B CODEBOOK · pp. 27–28

Related papers

Loading 2609.30699v1…