pith. sign in

arxiv: 2606.20762 · v1 · pith:WZSGAO4Hnew · submitted 2026-06-18 · 💻 cs.HC · cs.AI· cs.CY

Co-Construction Blindness and Asymmetric Epistemic Vulnerability in Human-LLM Interaction

Bianca Helena Ximenes This is my paper

Pith reviewed 2026-06-26 15:53 UTC · model grok-4.3

classification 💻 cs.HC cs.AIcs.CY
keywords co-construction blindnessasymmetric epistemic vulnerabilityhuman-LLM interactionstructural deferenceepistemic risksLLM disclaimersAI governance
0
0 comments X

The pith

Users of conversational LLMs are inside the loop shaping outputs yet disclaimers position them as external auditors, creating structural epistemic vulnerabilities that differ by authority position.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces co-construction blindness as the failure to recognize that LLM outputs are shaped by the user's inputs, history, and metadata rather than standing as independent assessments. It pairs this with asymmetric epistemic vulnerability, in which the effects of that blindness scale differently depending on where the user sits in an authority structure. These are presented as structural features of the interaction setup itself, illustrated through the public Dawkins-Claude exchange and a first-person case of structural deference in which the model reports gentler treatment for figures represented in training data. The analysis maps the research gaps opened by these constructs and calls for shared terminology as a precondition for governance and design responses.

Core claim

The paper claims that co-construction blindness, the unrecognized participation of every user in shaping LLM outputs through their inputs, accumulated history, and metadata, combined with asymmetric epistemic vulnerability arising from this blindness, constitute a structural inevitability in human-LLM interaction. This is demonstrated paradigmatically through the public Richard Dawkins interaction with Claude and supported by a secondary mechanism of structural deference observed in a first-person exchange where the model concedes gentler treatment due to training data representation. The paper maps resulting research gaps and calls for shared terminology to enable suitable governance and de

What carries the argument

Co-construction blindness (the failure to recognize user inputs as shaping LLM outputs) and asymmetric epistemic vulnerability (unequal consequences by authority position), which together reframe users from external verifiers to internal co-constructors.

If this is right

  • LLM disclaimers and interfaces must be redesigned to reflect that users are inside the output-construction loop rather than outside it.
  • Governance frameworks need to treat unequal epistemic consequences by authority position as a built-in feature requiring specific mitigation.
  • Shared terminology for these constructs is required before effective design or policy responses can be developed.
  • The identified research gaps center on documenting and measuring the two constructs across different user populations and interaction contexts.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The structural deference mechanism could be tested by comparing model outputs on the same query when posed by users with differing levels of public recognition.
  • If the constructs hold, then training-data influence on output tone may create measurable differences in model behavior toward high- versus low-visibility users.
  • The analysis implies that visibility of co-construction in interfaces might reduce vulnerability more effectively than additional verification prompts.

Load-bearing premise

Evidence from one public case and one first-person exchange suffices to establish these phenomena as structural inevitabilities rather than context-specific or design-contingent occurrences.

What would settle it

A controlled study documenting consistent user recognition of their co-constructive role across varied LLM interactions, or symmetric consequences independent of user authority, would challenge the structural-inevitability claim.

read the original abstract

This paper introduces two constructs to describe, as far as we know, a previously unnamed risk in human-LLM interaction. Co-construction blindness is the failure to recognize that LLM outputs are not independent assessments to be verified, but co-constructed artifacts shaped by the user's own inputs, accumulated history, and metadata. Every user of a conversational LLM is IN the loop, not ON it -- yet every deployment disclaimer positions them as external auditors. Asymmetric epistemic vulnerability describes the condition in which co-construction blindness produces consequences of radically different magnitude depending on where in the authority structure the user sits. We argue that these constructs describe a structural inevitability, not an anomaly, using the public case of Richard Dawkins's interaction with Claude as a paradigmatic instance. We document a secondary mechanism -- structural deference -- through a first-person exchange in which a large language model concedes that it treated Dawkins more gently than warranted because his intellectual output is represented in its training data. We map the research gaps this analysis opens and call for shared terminology as a precondition for appropriate governance and design response.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper introduces two new constructs—co-construction blindness (failure to recognize LLM outputs as co-constructed artifacts shaped by user inputs and history) and asymmetric epistemic vulnerability (differential consequences of this blindness by authority position)—and argues they represent structural inevitabilities in human-LLM interaction rather than anomalies. It uses the public Richard Dawkins-Claude exchange as a paradigmatic case and a first-person exchange documenting structural deference (preferential treatment due to training data overlap) as supporting illustration, then maps resulting research gaps and calls for shared terminology to inform governance and design.

Significance. If the constructs and their inevitability claim hold, the paper would supply a novel conceptual vocabulary for HCI and AI ethics that reframes user-LLM dynamics as inherently asymmetric, potentially guiding interface design, disclaimers, and policy. The explicit mapping of research gaps is a constructive contribution. However, the current grounding in a single public case plus one anecdotal exchange limits immediate applicability; the work's value would increase substantially with additional grounding or counterexample analysis.

major comments (1)
  1. [Abstract / case analysis] Abstract and case description: The central claim that co-construction blindness and asymmetric epistemic vulnerability are 'structural inevitabilities' (not design-contingent) is load-bearing yet supported only by treating the Dawkins-Claude exchange as paradigmatic and one first-person exchange as documentation of structural deference. No derivation, survey of architectures, or counterexamples (e.g., systems with explicit co-construction disclaimers) is supplied to show the phenomena must arise from the conversational interface itself independent of training data overlap or prompts.
minor comments (2)
  1. The paper would benefit from situating the new constructs against existing HCI literature on mental models of AI systems and co-construction in collaborative tools.
  2. Clarify whether 'structural deference' is presented as a third construct or a mechanism of the second; the abstract introduces it separately but the title does not.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their careful reading and constructive feedback. We respond to the major comment below, providing a point-by-point defense while noting planned revisions.

read point-by-point responses

  1. Referee: The central claim that co-construction blindness and asymmetric epistemic vulnerability are 'structural inevitabilities' (not design-contingent) is load-bearing yet supported only by treating the Dawkins-Claude exchange as paradigmatic and one first-person exchange as documentation of structural deference. No derivation, survey of architectures, or counterexamples (e.g., systems with explicit co-construction disclaimers) is supplied to show the phenomena must arise from the conversational interface itself independent of training data overlap or prompts.

    Authors: The manuscript advances a conceptual rather than empirical argument. The derivation of structural inevitability follows directly from the architecture of conversational interfaces: every response is conditioned on a context window that incorporates the user's prior inputs, accumulated history, and metadata, rendering outputs co-constructed by design. This holds independently of specific training-data overlaps or prompt phrasing, as the inclusion of user-provided context is a necessary feature of the interaction paradigm. The Dawkins-Claude exchange functions as a paradigmatic illustration of the epistemic consequences under authority asymmetry, while the first-person exchange documents the deference mechanism; neither is presented as exhaustive proof. We did not include an architecture survey because the claim targets the dominant conversational form rather than implementation variants. We acknowledge that explicit discussion of counterexamples (such as disclaimer-equipped systems) would strengthen the framing and will add a subsection clarifying why such measures do not eliminate the blindness if users continue to treat outputs as independent assessments. revision: partial

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper introduces two new conceptual constructs (co-construction blindness and asymmetric epistemic vulnerability) and asserts they represent structural inevitabilities in conversational LLM interfaces, supported by a paradigmatic public case and a first-person exchange. No equations, fitted parameters, self-citations, or uniqueness theorems appear in the provided text. The central claims rest on definitional framing and illustrative examples rather than any reduction of outputs to inputs by construction, making the derivation self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 3 invented entities

Ledger based solely on abstract; paper introduces two new conceptual entities and relies on domain assumptions about LLM behavior and authority structures without external benchmarks or data.

axioms (2)
  • domain assumption LLM outputs are shaped by user inputs, accumulated history, and metadata
    Foundational premise for defining co-construction blindness.
  • domain assumption Consequences of co-construction blindness vary systematically by user authority position
    Core premise for asymmetric epistemic vulnerability.
invented entities (3)
  • co-construction blindness no independent evidence
    purpose: Describe failure to recognize co-construction in LLM outputs
    New term introduced to name the risk
  • asymmetric epistemic vulnerability no independent evidence
    purpose: Describe unequal consequences based on authority
    New term introduced to name the risk
  • structural deference no independent evidence
    purpose: Secondary mechanism of differential treatment by models
    Documented via first-person exchange in abstract

pith-pipeline@v0.9.1-grok · 5715 in / 1424 out tokens · 42379 ms · 2026-06-26T15:53:33.962919+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

18 extracted references · 5 canonical work pages · 1 internal anchor

  1. [1]

    Dawkins, R. (2026). When Dawkins met Claude. UnHerd. https://unherd.com/2026/05/is-ai-the-next-phase-of-evolution/

    2026

  2. [2]

    Dawkins, R. (2026). When Claudia met Claudius. UnHerd. https://unherd.com/2026/05/when-claudia-met-claudius/

    2026

  3. [3]

    Dear Richard Dawkins

    Anonymous (2026). Dear Richard Dawkins. https://dearricharddawkins.com

    2026

  4. [4]

    Atari, M., Xue, M.J., Park, P.S., Blasi, D.E., & Henrich, J. (2023). Which Humans? PsyArXiv. https://doi.org/10.31234/osf.io/5b26t

    work page doi:10.31234/osf.io/5b26t 2023

  5. [5]

    Shanahan, M. (2024). Talking About Large Language Models. Communications of the ACM, 67(2), 68-79

    2024

  6. [6]

    Shanahan, M. (2023). Role Play with Large Language Models. Nature, 623, 493- 498

    2023

  7. [7]

    Tankelevitch, L., Kewenig, V., Simkute, A., Scott, A., Sarkar, A., Sellen, A., & Rintel, S. (2024). The Metacognitive Demands and Opportunities of Generative AI. Proceedings of CHI 2024

    2024

  8. [8]

    Sagan, C. (1995). The Fine Art of Baloney Detection. In The Demon-Haunted World: Science as a Candle in the Dark. Ballantine Books

    1995

  9. [9]

    Hart, J.T. (1971). The Inverse Care Law. The Lancet, 297(7696), 405-412

    1971

  10. [10]

    Blasi, D.E., Anastasopoulos, A., & Neubig, G. (2022). Systematic Inequalities in Language Technology Performance across the World's Languages. In Proceedings of ACL 2022

    2022

  11. [11]

    Ibrahim, L. et al. (2025). Measuring and mitigating overreliance is necessary for building human-compatible AI. arXiv:2509.08010

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  12. [12]

    Sun, F., Li, N., Wang, K., & Goette, L. (2025). Large Language Models are overconfident and amplify human bias. arXiv:2505.02151

    work page arXiv 2025

  13. [13]

    Fricker, M. (2007). Epistemic Injustice: Power and the Ethics of Knowing. Oxford University Press

    2007

  14. [14]

    Mollema, W.J.T. (2025). A taxonomy of epistemic injustice in the context of AI and the case for generative hermeneutical erasure. AI and Ethics, 5, 5535-5555. https://doi.org/10.1007/s43681-025-00801-w

    work page doi:10.1007/s43681-025-00801-w 2025

  15. [15]

    Bourdieu, P. (1991). Language and Symbolic Power. Harvard University Press

    1991

  16. [16]

    Regulation (EU) 2024/1689 of the European Parliament and of the Council laying down harmonised rules on artificial intelligence (Artificial Intelligence Act)

    European Parliament & Council of the European Union (2024). Regulation (EU) 2024/1689 of the European Parliament and of the Council laying down harmonised rules on artificial intelligence (Artificial Intelligence Act)

    2024

  17. [17]

    NIST 100-2e2023

    National Institute of Standards and Technology (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). https://doi.org/10.6028/NIST.AI.100- 1

    work page doi:10.6028/nist.ai.100- 2023

  18. [18]

    Model Specification

    Anthropic (2024). Model Specification. https://www.anthropic.com/research/model-spec

    2024