Cross-Disciplinary Taxonomy and Modeling of Misunderstanding Generation, Amplification, and Detection, from Pragmatics to AI Agents

Babak Abbaschian

Source-linked AI summary

Cross-Disciplinary Taxonomy and Modeling of Misunderstanding Generation, Amplification, and Detection, from Pragmatics to AI Agents

Babak Abbaschian

arXiv:2608.13604v1Aug 11, 2026cs.AIcs.CLcs.HCcs.MA

TL;DR

As AI agents increasingly mediate communication, undetected misunderstandings are becoming more consequential and harder to notice. The paper synthesizes nine fields into a formal, auditable taxonomy showing that misunderstanding is a process with distinct mechanisms that generate, amplify, or detect divergence.

Problem
Research lacks an integrated, mechanism-level map of how misunderstandings are generated, amplified, detected, or allowed to persist across analytical levels.
Method
The paper develops a cross-disciplinary conceptual synthesis that formalizes communicative transformations and makes its taxonomy, evidence, and coding criteria explicit.
Results
The taxonomy identifies eleven failure modes across eight analytical layers and three roles: generate, amplify, and detect.
Takeaways & Limitations
Misunderstanding is best understood as a process in which multiple mechanisms generate divergence, some amplify it, and repair determines whether it becomes visible and resolved.
Takeaways & Limitations
The synthesis may omit mechanisms concentrated in fields outside the nine selected for review.

Abstract

from arXiv · show

Detection of misunderstanding is an urgent problem to solve because communication has moved away from real-time, in-person interaction and is increasingly handled by AI-mediated channels. This shift cuts communicators off from the resources repair depends on faster than new means of detection are being built. In this paper we analyse misunderstanding as a layered process in which a divergence is generated, may then be amplified, and is either detected and repaired or left to persist unnoticed. Consolidating accounts from nine fields of research that do not ordinarily cite one another, we identify eleven exact failure modes and show that each operates at a specific point in a communicative process rather than anywhere within it. Those points give eight analytical layers, derived from the literature rather than adopted from an existing model. Eight of the mechanisms primarily generate a divergence, two primarily amplify one already present, and one governs whether a divergence is detected and repaired. We model the eight layers formally, extending information and communication theory from the transmission of signals to the reconstruction of meaning, and we supply a source-by-source evidence matrix that makes every rating auditable, a coding manual, and nine analysed dialogue cases. No prior classification of misunderstanding both locates mechanisms at points in the process and types them by function.

1. Introduction · 2. Cross · Disciplinary Literature Review

The paper synthesizes fragmented research on misunderstanding into a mechanism-level account of how divergences are generated, amplified, detected, and repaired, especially as AI-mediated communication removes repair resources. Across nine fields, it identifies eleven mechanisms at eight process layers, formalizes their operation, and makes the taxonomy auditable and applicable to dialogue cases.

1.1 The Problem: Undetected misunderstanding occurs when reconstructed meaning diverges from intended meaning without either party recognizing it, a condition increasingly difficult to notice with AI agents.The paper focuses on this condition because AI-mediated channels can remove resources that support detection and repair.
1.1 The Problem: The literature is deep within fields but poorly integrated, motivating a mechanism-level map that specifies how misunderstandings are generated, amplified, detected, or allowed to persist.The review groups heterogeneous literatures contextually before comparing recurring mechanisms across fields.
1.3 Contributions: The resulting framework is checkable through source-level evidence ratings, a coding manual, and analysis of nine dialogue cases at the level of specific mechanisms.It supports analysis of particular primary and contributing mechanisms rather than broad labels such as poor communication.
1.3 Contributions: The synthesis consolidates accounts from nine fields into eleven distinct mechanisms located at eight analytical layers derived from the literature rather than adopted from an existing model.The fields address signal transmission, pragmatic inference, speech acts, grounding, repair, cognition, attribution, media, and cultural conventions.
1.3 Contributions: The taxonomy formalizes misunderstanding as material divergence and models the eight layers as conditional distributions carrying intended meaning to reconstructed meaning under loss.The formalization extends information and communication theory beyond signal transmission toward semantic reconstruction and effect on conduct.
1.3 Contributions: Eight mechanisms primarily generate divergences, two amplify existing divergences, and one governs detection and repair through the generate-amplify-detect classification.The typing emerged from the mechanisms and converged with distinctions independently developed in other research traditions.
Disciplinary Literature Review: The literature review identifies mechanisms at distinct communicative points, including inference, speech-act recognition, common-ground estimation, perspective formation, attribution, and channel effects.Examples include Pragmatic Inference Failure, Illocutionary Force Mismatch, Referential Misalignment, Presupposition Mismatch, Common-Ground Overestimation, Attributional Distortion, and Channel and Medium Distortion.
1.1 The Problem: Repair begins about once every 1.4 minutes across twelve languages, but text interfaces and agent-generated words reduce participants’ willingness to initiate it.The paper distinguishes reduced willingness to repair from reduced capacity to repair.

. Defining Misunderstanding

The paper defines misunderstanding as a material divergence between intended and reconstructed meaning, which may concern reference, implication, speech act, relational stance, or requested action. It separates divergence from recoverability and detection, treating communication as a chain in which encoding, channel transformation, and decoding can produce failure.

Definition: A misunderstanding is a divergence between the meaning a speaker sought to communicate and the meaning a listener reconstructed.The divergence may concern the intended referent, an implied meaning, the speech act, relational stance, or requested action.
Formal criterion: A misunderstanding occurs when semantic distance or task loss reaches a material threshold, expressed as d(M, M̂) ≥ τ.For discrete meaning categories, the criterion reduces to M ≠ M̂; prevalence is Pmis = Pr(M ≠ M̂).
Materiality: Materiality is determined by whether the divergence would change the listener’s action, commitment, or conclusion, not by whether it is detected.This preserves the analytical separation between misunderstanding and Repair Failure.
Recoverability: Recoverability is distinct from divergence magnitude and depends on evidence of divergence, observability, feedback, and available repair actions.A larger semantic distortion need not imply lower recoverability, so recoverability remains outside the definition.
Detection and resolution: Detection and resolution are separate from divergence: a misunderstanding may remain undetected, be detected but unresolved, or be detected and resolved.Undetected cases may influence subsequent actions without recognition and are hardest to observe, making evidence strongest for repair-revealed cases.

Disciplinary Conceptual Synthesis · 4. The Taxonomy: Cross

The taxonomy consolidates overlapping accounts from nine fields into eleven distinct misunderstanding mechanisms, locating them at eight layers of a communicative episode. It derives the layers first, formalizes their sequence second, and identifies functional roles only after completing that structure.

4. The Taxonomy: Cross: Nine fields were compared through best-fit framework synthesis to combine overlapping concepts while retaining mechanisms that differed in operation or location.The synthesis follows Carroll et al. (2013).
4. The Taxonomy: Cross: The taxonomy presents eleven mechanisms, eight communicative points, formal layer descriptions, and three functional roles in that derivational order.The functional roles describe the completed taxonomy rather than an externally imposed scheme.
4. The Taxonomy: Cross: The eight points are treated as layers and formally stated as the chain a meaning passes through between participants.The formal chain is presented after the mechanisms and their locations.
4.1 The Eleven Mechanisms and Where They Operate: Consolidation of differently named accounts reduced the reviewed literature to eleven distinct failure mechanisms.The density of competing names was itself identified as part of the problem.
4.1 The Eleven Mechanisms and Where They Operate: Each mechanism operates at a particular point in a communicative episode rather than anywhere within it.The analysis grouped the eleven mechanisms across eight such points, which the paper calls layers.
4.1 The Eleven Mechanisms and Where They Operate: Formation is the first layer, where a communicative goal becomes a determinate intended meaning, with one mechanism operating there.This identifies the first layer’s process and mechanism count.

1. Intent Underspecification … 4. Encoding Ambiguity

The first four layers locate misunderstanding in underspecified intent, faulty estimates of shared knowledge and perspective, and ambiguous encoding. These mechanisms arise at distinct stages from forming meaning through expressing it, with conversational timing, cognitive load, bias, and wording shaping divergence.

1. Intent Underspecification: Intent Underspecification generates divergence when the speaker’s intended meaning is vague, unstable, inconsistent, or incomplete before expression.Speaking need not wait for a fully specified message, making early commitment common in conversation.
1. Intent Underspecification: Conversational turn timing, cognitive overload, unfamiliar tasks, complexity, and time pressure make underspecified intentions especially likely.The mechanism is often inferred retrospectively from behaviour such as substantial revision.
2. Common-Ground Overestimation: Common-Ground Overestimation generates divergence when participants assume that knowledge, beliefs, assumptions, or background information are shared when they are not.The mechanism reflects failure to establish or update common ground through grounding.
2. Common-Ground Overestimation: False consensus, the curse of knowledge, explanatory-depth illusions, and transparency illusions reinforce overestimation of shared understanding and intention clarity.These biases can lead speakers to provide less information than listeners need.
3. Perspective-Taking Failure: Perspective-Taking Failure generates divergence when participants formulate or interpret messages from their own viewpoint rather than the other person’s.Egocentric interpretation is often a fast default, while perspective incorporation requires additional processing and may fail under pressure.
3. Perspective-Taking Failure: Breakdowns in automatic alignment of words and grammatical structures can produce misunderstanding without deliberate failure to consider another perspective.The bias falls more heavily on production than comprehension, placing the mechanism on the speaker’s side as often as the listener’s.
4. Encoding Ambiguity: Encoding Ambiguity generates divergence when an expression permits multiple plausible interpretations and the listener selects one different from the speaker’s intended meaning.Sources include polysemy, unclear syntax or scope, pronoun reference, unstated thresholds, and jargon without shared definitions.
4. Encoding Ambiguity: Encoding ambiguity is empirically anchored by wording that was measurably less dense at points where grounding occurred during miscommunication.Communication theory treats encoding fidelity as a distinct possible failure point.

5. Channel and Medium Distortion · 6. Referential Misalignment

Channel and Medium Distortion amplifies an existing divergence by removing, degrading, delaying, or altering information needed for interpretation and grounding. Referential Misalignment generates divergence when interlocutors connect the same expression to different referents, often before collaborative reference is genuinely established.

5. Channel and Medium Distortion: Channel and Medium Distortion characteristically worsens a divergence already present.It operates when the medium removes, degrades, delays, or alters information needed to interpret and ground a message.
5. Channel and Medium Distortion: The distortion concerns timing, tone, facial expression, gesture, visibility, shared physical context, and immediate clarification.These resources can support interpretation, grounding, and repair during communication.
5. Channel and Medium Distortion: Lean media provide fewer ambiguity-resolving resources, and different media withdraw different grounding resources.Nonverbal meaning also depends on context rather than functioning as a fixed code.
5. Channel and Medium Distortion: Losing a nonverbal channel removes information that words may not supply.Gestural cues can improve recovery of relative position and size.
5. Channel and Medium Distortion: The fifth layer is Semantic and Pragmatic Alignment, where referents and the propositions an utterance takes for granted are resolved.Two mechanisms operate at this layer.
6. Referential Misalignment: Referential Misalignment generates divergence when speaker and listener connect the same expression to different objects, people, events, times, locations, or states.The mismatch concerns what the expression is taken to refer to.
6. Referential Misalignment: Reference is established collaboratively through proposing, accepting, refining, or repairing a referent.Misalignment arises when apparent acceptance occurs before both participants identify the same referent.
6. Referential Misalignment: Experimental, conversation-analytic, and corpus evidence supports Referential Misalignment as a mechanism whose apparent grounding can conceal different referent selections.The evidence spans collaborative reference experiments, problematic-reference analyses, and corpus research.

7. Presupposition Mismatch

Presupposition mismatch arises when an utterance assumes a proposition the listener does not share, creating divergence during the Inference and Interpretation layer. Listeners may accommodate the assumption, fail to form a complete interpretation, or incur greater processing cost when context does not support it.

7. Presupposition Mismatch: Presupposition mismatch generates divergence when an utterance assumes a proposition the listener does not share.The listener may accept the assumption without challenge or be unable to form a complete interpretation.
7. Presupposition Mismatch: Presuppositions arise through factive verbs, definite descriptions, cleft constructions, particles, and change-of-state verbs.Listeners may accommodate an unshared presupposition by adding it to the conversational context; discourse-semantic accounts model resolution anaphorically within discourse context.
7. Presupposition Mismatch: Accommodating an unestablished presupposition carries greater processing cost than interpreting one already supported by context.This experimental finding concerns the processing demands of presupposition resolution.
7. Presupposition Mismatch: The sixth layer, Inference and Interpretation, reconstructs implied meaning and the communicative act being performed.Presupposition mismatch is one of two mechanisms operating at this layer.

8. Pragmatic Inference Failure

Pragmatic Inference Failure generates misunderstanding when a listener reconstructs an implied or indirect meaning that differs from the speaker’s intended meaning. It can result from divergent background assumptions, contextual relevance judgments, or salience assigned to competing meanings.

Pragmatic Inference Failure generates a divergence when the listener reconstructs an implied or indirect meaning different from the speaker’s intended meaning.
Different background assumptions can cause participants to apply conversational expectations differently.
Divergence can also arise when participants evaluate relevance against different contexts or assign different salience to competing meanings.

9. Illocutionary Force Mismatch

Illocutionary Force Mismatch generates misunderstanding when a listener misidentifies the communicative act performed by an utterance. It is more likely with indirect or face-sensitive expression, differing cultural conventions, and misleading sequential expectations.

9. Illocutionary Force Mismatch: Illocutionary Force Mismatch arises when listeners misidentify the communicative act being performed.Examples include interpreting a request as a question, a warning as a threat, a joke as sincere assertion, or a refusal as an invitation to negotiate.
9. Illocutionary Force Mismatch: The mismatch reflects a divergence between the words uttered and the act performed through them.Indirect expression, face management, cultural conventions, and misleading sequential expectations increase its likelihood.
9. Illocutionary Force Mismatch: The mechanism operates in the seventh analytical layer, Relational Interpretation, where motive, disposition, and relational stance are inferred from the same material.One mechanism operates in this layer.

10. Attributional Distortion

Attributional Distortion worsens an existing divergence when communicative behaviour is explained through character, motive, or disposition rather than situational factors or ordinary communication difficulty. This mechanism can lead unclear messages to be interpreted as carelessness, hostility, manipulation, or disrespect.

Attributional Distortion: Attributional Distortion amplifies an existing divergence by attributing another person’s communicative behaviour to character, motive, or disposition.The mechanism underweights situational explanations and the ordinary difficulty of communicating.
Attributional Distortion: Research on dispositional attribution and the fundamental attribution error grounds this mechanism.An unclear or incomplete message may consequently be interpreted as carelessness, hostility, manipulation, or disrespect.

11. Repair Failure

Repair Failure does not generate a divergence; it determines whether an existing divergence becomes visible, detected and unresolved, or detected and resolved. The taxonomy’s three two-mechanism layers preserve distinctions from the reviewed literatures and therefore require independent testing.

Repair Failure: Repair Failure governs whether a divergence becomes visible by enabling or lacking checking, clarification, confirmation, correction, or acknowledgement.It produces no divergence of its own.
Repair Failure: Repair can fail when a divergence remains implicit, clarification is discouraged, feedback is restricted, or conversational progress resumes without resolving the meaning difference.Failure therefore leaves misunderstanding undetected, detected but unresolved, or detected and resolved.
Taxonomic structure: Five of eight layers admit one failure mechanism, while three admit two because the reviewed literatures preserve distinctions that cannot be collapsed without loss.The three two-mechanism layers distinguish reference from presupposition, implied meaning from communicative act, and mistaken from correctly applied models of the other participant.
Taxonomic structure: The three two-mechanism layers concentrate the synthesis’s taxonomic work and are the distinctions most in need of independent testing.Where mechanisms can be merged without losing a distinction drawn by the source literatures, the taxonomy retains one layer mechanism.

5. Applying the Taxonomy to Dialogue Cases

The section applies the taxonomy to nine dialogue cases by identifying each case’s misunderstood object, primary and contributing mechanisms, and detection or resolution state. The cases illustrate causal distinctions and coverage of the taxonomy but do not validate it.

Case selection: Nine cases apply the taxonomy across documented material, a clinical vignette, and constructed illustrations, including corpus, experimental, and accident-investigation cases.Three cases are drawn from documented material, one is a stylised clinical vignette, and five are constructed illustrations.
Interpretive limits: The cases demonstrate application rather than validation; validation would require independent coders using the coding manual on shared cases and reporting agreement.The section explicitly states that the case analyses do not validate the taxonomy.
Analytical procedure: The primary mechanism produced the divergence, whereas contributing mechanisms enabled, amplified, or concealed it without independently producing it.This distinction prevents every plausible mechanism from receiving equal causal importance.
Case analyses: The cases assign distinct mechanisms to concrete misunderstandings, including referential misalignment, pragmatic inference failure, illocutionary force mismatch, common-ground overestimation, and presupposition mismatch.Examples include different referents for “that,” unrecognized indirect refusals, incomplete clinical common ground, and disagreement over the presupposition triggered by “stop.”
Cross-case findings: Across the nine cases, every mechanism whose dominant role is generation appears once as a primary mechanism, while Channel and Medium Distortion and Repair Failure appear only as contributors.Their contributor-only status follows from their definitions as amplification and detection-governance mechanisms rather than from selection gaps.

. Discussion

The discussion finds broad but uneven support for the taxonomy, confirms that divergence generation and detection are distinct functions, and identifies unresolved gaps in layer coverage, evidence, and formalization. Nine cases illustrate that mechanisms and outcomes must be recorded separately, while the model remains an unfitted formal description requiring future validation.

Evidence convergence: Ten of eleven mechanisms receive High convergence ratings, while Attributional Distortion receives Moderate because evidence breadth is limited to social psychology.The Moderate rating does not indicate weak evidence within social psychology; it indicates that independent research bodies have not yet contributed evidence.
Evidence convergence: The taxonomy treats intercultural difference as a condition affecting Pragmatic Inference Failure and Illocutionary Force Mismatch, not as a separate mechanism.Both mechanisms become more likely when participants rely on different cultural conventions.
Functional roles: Prevention and detection address different layers: better formulation targets divergence generation, whereas confirmation and read-back target whether divergence is exposed and repaired.Improving one function does not necessarily improve the other, and a single AI agent may improve generation while withdrawing detection.
Case analysis: Five of nine dialogue cases remain undetected, one is detected but unresolved, and three are detected and resolved.The cases assign mechanisms using observable exclusions and record interaction outcomes independently from the mechanisms that produced divergences.
Gaps and limitations: Five of eight layers carry one mechanism and three carry two, while interactive alignment and the decomposition of Regulation and Recovery remain underrepresented or insufficiently divided.The discussion identifies distinctions involving reference and presupposition, implied meaning and illocutionary act, and models of the other participant versus correlative failure.
Gaps and limitations: The formal model separates material divergence from evidence revealing it but supplies no prior, inference procedure, or data-fitted estimates.Future work should instantiate and validate the distance function for particular misunderstood objects and develop formal treatments of individual mechanisms.

7. Conclusion

The paper presents misunderstanding as a process involving generation, amplification, and detection across eight analytical layers and eleven mechanisms. It offers an auditable conceptual taxonomy while identifying independent validation and undetected misunderstandings as priorities for future work.

Conclusion: The synthesis integrates nine fields into eleven mechanisms organized across eight analytical layers and three functional roles: generate, amplify, and detect.The roles emerged from the mechanisms rather than being imported from an existing scheme.
Conclusion: 10 mechanisms receive High convergence ratings, while Attributional Distortion receives Moderate support.These ratings summarize the evidence convergence assessment.
Conclusion: Misunderstanding arises through failures in meaning formation, reference and assumption alignment, inference, and perspective-taking; channel and attributional effects may amplify divergence, while repair governs visibility and resolution.The conclusion therefore treats misunderstanding as a process rather than a single-cause event.
Conclusion: The taxonomy supplies shared vocabulary, an auditable evidence matrix, a coding manual, and nine dialogue cases demonstrating its application.The cases exercise whether mechanisms can be applied, not merely named.
Conclusion: The framework remains conceptual, requiring independent-coder testing for reliable distinction and use across languages, cultures, media, and applied settings.Future work should especially examine undetected misunderstandings, which may be consequential yet least visible in existing evidence.

ppendix A: Evidence Matrix

The appendix makes the convergence ratings auditable through a source-by-source matrix, while documenting how sources are assigned to mechanisms and evidence types. It also records empirical support for communication-channel losses, presupposition costs, and reduced repair initiation with artificial agents.

Evidence Matrix: The matrix grounds convergence ratings in independent research bodies and the range of evidence types represented.Its ratings follow the rubric in Section 4.4.
Evidence Matrix: Source assignments remain interpretive: multi-mechanism sources appear under a primary mechanism, while independent coding is intended to test those assignments.Framing, scope, motivation, and method sources are marked as not counting toward a mechanism rating.
Evidence Matrix: Nonverbal channels carry information beyond words: gestures improved recovery of object position and size, while prosody helped resolve syntactic ambiguity before phrase completion.These effects were reported for particular semantic categories, and prosody is absent from text.
Evidence Matrix: Presuppositions are processed online and accommodation carries a measurable cost, providing experimental support that raises M7 to High.The matrix also links repair detection to evidence that repair initiation falls when an interlocutor is known to be an artificial agent.

Appendix B: Coding Manual … 10. ATTRIBUTIONAL DISTORTION

The appendix presents an initial coding manual for applying the taxonomy to observed misunderstanding episodes, not a validation study. It defines a general procedure and mechanism-specific tests, including attributional distortion’s conditions for generation versus amplification.

Appendix B: Coding Manual: The manual is an instrument for future validation, combining a general coding procedure with mechanism-specific indicators, exclusion rules, and nearest-neighbor tests.It is explicitly not itself a validation.
B.1 General Procedure: Coders analyze observable interaction episodes where reconstructed meaning differs from intended meaning, excluding mishearing without semantic divergence, unrecovered ambiguity, disagreement, and successful deception.The unit of analysis is defined by divergence between intended and reconstructed meaning, not assumed inner states.
B.1 General Procedure: Coders identify the diverged object—content, referent, speech act, implication, intent or relational stance, or future action—to determine the likely primary mechanism and its role.This object-based step guides subsequent mechanism assignment.
B.1 General Procedure: The primary mechanism is normally one mechanism whose removal would have prevented the divergence, while contributing mechanisms may enable, worsen, or conceal it.Contributors can be several or absent.
B.1 General Procedure: Each mechanism is coded by episode role—generation, amplification, or detection and repair—and the episode is classified as undetected or detected, then resolved or unresolved.Channel and Medium Distortion and Attributional Distortion can generate divergences in some episodes despite differing dominant roles.
MECHANISM DO NOT CODE WHEN TEST AGAINST THE NEAREST NEIGHBOR: Nearest-neighbor tests distinguish mechanisms by excluding literal wording, triggered assumptions, act type, or later-revealed mismatches when those alternatives better explain the episode.The manual also distinguishes missed implicatures and indirect meanings from cases where literal content was understood.
8. PRAGMATIC INFERENCE FAILURE: Pragmatic Inference Failure concerns unrecovered implied meaning, whereas Attributional Distortion concerns explaining communicative behavior through character or motive, potentially discouraging repair.This distinction separates missed implicatures from relational judgments about behavior.
10. ATTRIBUTIONAL DISTORTION: Attributional Distortion is coded as generating when a dispositional attribution creates a relational-stance or intent misunderstanding; otherwise it is amplification, with repair effects recorded separately.It is not primary when another mechanism produces message-content divergence, unless a dispositional attribution additionally forms.