Speaker-Conditioned Meaning, Iterative Transmission, and the Representational Limits of Large Language Models
An interdisciplinary technical essay on computational semantics, information theory, linguistic pragmatics, and artificial intelligence
Abstract
Large language models (LLMs) have achieved remarkable proficiency in producing linguistically coherent and contextually appropriate text. Yet fluency should not be conflated with semantic fidelity. Communication may appear successful even when the meaning reconstructed by a recipient diverges substantially from the meaning originally intended by the speaker. Such divergence becomes particularly consequential when information undergoes repeated linguistic transformations, as in machine translation, summarization, retrieval-augmented generation, and multi-agent reasoning systems.
This essay examines a central problem in language-mediated artificial intelligence: the possibility that semantic information is progressively altered without the participants in a communication chain recognizing that any alteration has occurred. Drawing on information theory, prototype semantics, probabilistic pragmatics, cultural transmission research, and contemporary studies of LLM reliability, it distinguishes lexical equivalence from referential fidelity, surface-level consistency from semantic correctness, and representational capacity from observational access.
The analysis proposes that reliable language-mediated intelligence requires more than probabilistically plausible generation. It requires mechanisms for preserving speaker-dependent meaning, identifying unsupported semantic substitutions, and maintaining access to original evidence. A formal framework and experimental design are developed to investigate how semantic distortions arise, propagate, and become difficult to detect in iterative LLM systems.
Keywords: Large language models; semantic fidelity; semantic drift; lexical non-equivalence; pragmatics; information theory; speaker modeling; symbol grounding; iterated learning; representational capacity.
1. Introduction: The Illusion of Successful Communication
Human communication is frequently evaluated according to its observable success: whether a statement is grammatically intelligible, whether a recipient produces a contextually appropriate response, and whether an interaction proceeds without interruption. Such criteria, however, do not necessarily establish that the intended meaning has been preserved.
Consider a Korean speaker describing a soft, magenta-colored, heart-shaped plush toy. Having learned that the Korean word inhyeong (인형) corresponds to the English word doll, the speaker produces the sentence:
“I bought a doll.”
The statement is linguistically acceptable. Yet an English-speaking recipient might imagine a conventional human-shaped doll rather than a heart-shaped plush object. The two participants could continue the conversation without difficulty, neither recognizing the difference between their respective mental representations.
Paradoxically, a grammatically unconventional expression such as “a fluffy heart-shaped thing” might preserve more of the object's perceptually relevant properties than the conventionally sanctioned translation.
This does not mean that doll excludes plush toys, or that the improvised expression uniquely identifies the object. Rather, it demonstrates that lexical correctness, communicative intelligibility, and referential accuracy are distinct properties.
The problem becomes more pronounced when the recipient subsequently communicates their reconstructed representation to another individual. Each participant may reproduce an interpretation that is reasonable relative to the information received, while the resulting representation gradually diverges from the original referent.
The distinctive risk is not merely semantic loss. It is undetected semantic loss: the progressive modification of information under conditions in which linguistic fluency and apparent conversational success conceal the absence of semantic agreement.
This essay argues that the phenomenon has direct implications for LLM architectures and AI-mediated knowledge systems. In particular, it suggests that systems optimized primarily for plausible linguistic continuation may remain vulnerable to failures of meaning preservation even when their outputs are coherent, internally consistent, and expressed with high confidence.
2. Lexical Equivalence Is Not Semantic Equivalence
2.1. From lexical labels to conceptual categories
The relationship between a word and its meaning is not adequately represented as a simple one-to-one mapping between linguistic labels and independently fixed concepts.
Prototype theory provides an important foundation for understanding this distinction. Rosch (1975) demonstrated that members of a semantic category differ in their perceived typicality, challenging the assumption that categories are psychologically represented exclusively through uniform sets of necessary and sufficient conditions [2].
A category such as doll may include multiple kinds of objects, but different speakers may associate the category with different prototypical examples.
Consequently, two individuals can agree upon a word's conventional classification while disagreeing about the specific object that the word most naturally evokes.
Cross-linguistic lexical correspondence introduces another source of variability. A translation pair can exhibit substantial overlap without preserving identical category boundaries, conventional associations, or pragmatic implications.
The expressions pathetic and hansimhada (한심하다), for instance, overlap in negative evaluation but differ in their typical associations with inadequacy, contempt, disappointment, and pity. Their correspondence is context-dependent rather than universally interchangeable.
The important distinction is between a word's capacity to denote an object and the likelihood that it will induce the intended representation in a particular listener.
2.2. Speaker meaning and conventional meaning
Grice's analysis of meaning (1957) established a foundational distinction between linguistic meaning and communicative intention [1].
A speaker does not necessarily intend only the most conventional interpretation of the expression selected. Instead, the utterance participates in an intentional communicative act whose interpretation depends on context, background assumptions, and the listener's inferences.
For a second-language learner, vocabulary limitations introduce an additional complication: the speaker may select a word not because it most precisely expresses the intended meaning, but because it is the most accessible expression within their available vocabulary.
The distinction can be formulated as follows:
Available expression ≠ Intended meaning ≠ Reconstructed meaning
A speaker may possess a sufficiently detailed mental representation of an object while lacking the lexical resources necessary to communicate that representation precisely.
The resulting utterance is not necessarily evidence of conceptual deficiency. It may instead reflect an encoding limitation.
Conversely, a recipient's ability to decode the utterance does not establish that the original concept has been recovered.
2.3. The problem of false semantic alignment
Clark and Brennan (1991) described communication as a process requiring the establishment and maintenance of common ground [3]. Successful grounding depends on evidence that participants have understood one another sufficiently for their purposes.
However, conversational acceptance is not equivalent to representational identity.
A listener's affirmative response may establish that an utterance was interpretable, without establishing that the intended referent was reconstructed correctly.
The resulting discrepancy can remain latent until an external action exposes it—for instance, when the listener is asked to purchase an object corresponding to the description.
This phenomenon motivates a distinction between perceived semantic alignment, in which participants believe they share an interpretation, and actual semantic alignment, in which their interpretations preserve the task-relevant properties of the same referent.
The divergence between these states is especially important for AI systems that treat conversational acceptance as sufficient evidence of successful understanding.
3. Semantic Transmission as Lossy Reconstruction
3.1. The information-theoretic perspective
Shannon's (1948) mathematical theory of communication provides a rigorous framework for analyzing information transmission, although its original formulation deliberately separates the engineering problem of transmitting signals from questions of semantic interpretation [5].
The present problem requires an additional semantic layer.
Let denote an original intended meaning and the utterance produced to communicate it. A recipient reconstructs an interpretation , which may subsequently be encoded into a new utterance .
An iterative transmission process can be represented as:
At each stage, the reconstructed meaning is conditioned on the information available to the recipient and on the recipient's interpretive assumptions.
Importantly, the mapping from meaning to language need not be injective. Distinct meanings may produce identical utterances.
For example, two visually different objects may both be described as “a pink toy.”
Once the original distinction has been omitted, a subsequent recipient cannot uniquely infer which object was intended from that description alone.
This is not simply a limitation of linguistic sophistication. It is a problem of underdetermination: multiple original states are compatible with the same observed message.
3.2. The data-processing constraint
Suppose an original variable is encoded into an intermediate representation , from which an output is generated.
If the process forms the Markov chain
the data-processing inequality establishes that
where denotes mutual information.
Under these assumptions, processing the intermediate representation cannot increase the amount of information it contains about the original variable.
The implication for semantic transmission is conditional but important: if a particular distinction has been eliminated from an intermediate representation, subsequent processing cannot reliably recover that distinction without additional information.
A language model may nevertheless generate a highly plausible reconstruction. Yet plausibility does not establish recovery of the original state.
Indeed, a plausible reconstruction may be more dangerous than an explicitly uncertain one because it can replace missing information with an apparently authoritative interpretation.
The distinction between recovering information and inferring a likely substitute is fundamental to assessing the reliability of generative systems.
3.3. Semantic drift does not imply monotonically increasing error
It would be incorrect to assume that every additional transmission necessarily produces a larger error.
A transformation may preserve meaning, recover a previously omitted distinction through supplementary evidence, or even coincidentally produce an output closer to the original.
Nevertheless, repeated reconstruction can alter the distribution of possible interpretations.
In particular, ambiguous or unfamiliar details may be replaced by more conventional, memorable, or statistically probable alternatives.
A system can consequently converge toward a stable representation that is internally coherent but systematically different from the original.
The resulting stability should not be mistaken for semantic accuracy.
4. Speaker-Conditioned Meaning and Probabilistic Pragmatics
4.1. The speaker is part of the semantic problem
Conventional language-processing models are frequently evaluated according to the relationship between utterances and interpretations. Yet an utterance alone may be insufficient to determine its intended meaning.
The speaker's linguistic competence, conceptual knowledge, cultural background, professional expertise, and history of word usage can all influence the relationship between an intended meaning and the expression selected to communicate it.
A conceptual model of speaker-conditioned interpretation is:
where denotes intended meaning, the observed utterance, the speaker's relevant background, and the communicative context.
This formulation does not describe a probability distribution explicitly computed by every contemporary LLM. Rather, it formalizes the inferential dependencies that a speaker-sensitive interpretation system would ideally consider.
Crucially, the background variable should represent evidence about an individual speaker, not simply demographic stereotypes.
4.2. Rational Speech Acts and communicative inference
Frank and Goodman (2012) demonstrated that probabilistic pragmatic models can account for human judgments in referential communication tasks [4].
The Rational Speech Acts framework provides a principled account of how speakers choose expressions with an anticipated listener in mind, and how listeners infer intentions from those choices.
However, a speaker's available vocabulary changes the interpretation of their choices.
If a fluent speaker selects doll over plush toy, a listener may reasonably investigate whether the selection carries a meaningful distinction.
If a beginning learner does not know the expression plush toy, the absence of that expression offers little evidence concerning the intended object's physical properties.
Therefore, a more complete pragmatic model must consider not only what a speaker said, but what the speaker was capable of saying.
A missing lexical alternative may reflect communicative choice, limited competence, or incomplete access to a concept.
These possibilities should not be treated as equivalent.
4.3. Implications for personalized LLM interpretation
For language models, speaker-conditioned interpretation suggests that meaning reconstruction should be informed by relevant conversational history and explicit user-defined terminology.
If a user has previously established a particular definition for a term, a later response should not automatically replace that definition with the population-level interpretation learned during pretraining.
The central requirement is not exhaustive personalization. It is evidence-sensitive semantic interpretation: the ability to distinguish the user's documented meaning from an inferred default.
A robust system should also preserve uncertainty where multiple interpretations remain compatible with the available evidence.
Prematurely choosing a single interpretation may create downstream errors that become increasingly difficult to identify after subsequent transformations.
5. Prompting as Probabilistic Conditioning, Not Semantic Guarantee
5.1. The conditional distribution of language generation
Autoregressive language models generate tokens according to a conditional probability distribution:
where represents learned model parameters, the previously generated tokens, and the available conditioning context.
The Transformer architecture introduced by Vaswani et al. (2017) provides the architectural foundation for many modern LLMs [10].
System instructions, user prompts, conversational history, and language choice influence the probability distribution from which responses are produced.
One might describe these conditions as narrowing a model's semantic possibilities. Yet such a description is metaphorical rather than literal.
Prompts generally modify the relative probabilities of outputs rather than establishing impermeable boundaries around valid meanings.
Even a highly constrained prompt may assign some probability to unintended interpretations.
5.2. Output consistency and semantic correctness
A critical distinction follows between uncertainty over possible outputs and fidelity to an external referent.
A model may repeatedly produce the same answer because its conditional probability distribution strongly favors that answer.
However, repeated agreement does not establish that the answer corresponds to the speaker's original intention.
Farquhar et al. (2024) introduced semantic entropy as a method for detecting certain forms of LLM hallucination by grouping sampled responses according to their meanings rather than their surface forms [11].
This approach recognizes that linguistically different answers may be semantically equivalent.
Nevertheless, semantic consistency and external correctness remain conceptually separate.
If a model consistently interprets a plush object as a conventional doll, its outputs may exhibit low semantic variation while retaining a systematic referential error.
Thus:
A model can be confidently, consistently, and intelligibly wrong.
5.3. Fluency as a source of epistemic risk
The difficulty is intensified by the rhetorical quality of LLM-generated language.
A semantically distorted response can remain grammatical, cohesive, and professionally formulated. Such characteristics may encourage subsequent recipients to accept the output without independently verifying its relationship to the original source.
This essay proposes the term semantic provenance debt to describe the growing difficulty of reconstructing the evidential origins of claims after repeated transformations have removed, modified, or obscured their relationship to source material.
The term is proposed here as an analytical construct rather than presented as an established measure in the literature.
Semantic provenance debt would be expected to increase when successive agents rely primarily on predecessor summaries rather than retaining access to the original evidence.
6. Are LLMs Confined to a Countable Space of Meanings?
6.1. Discrete language and continuous representations
A common intuition is that an LLM must operate within a fixed inventory of meanings learned during training.
The computational picture is more complex.
A model's vocabulary is finite, and the set of finite token sequences over that vocabulary is countably infinite when arbitrary sequence lengths are permitted.
By contrast, neural hidden states are conventionally modeled as vectors in continuous spaces such as
The mathematical space of such vectors is uncountable, although physical implementations employ finite-precision numerical representations.
For a fixed architecture, bounded execution, and finite numerical precision, the number of distinguishable machine states is finite.
Yet this fact does not imply that the model operates through a simple finite dictionary of independent semantic entries.
Representation is distributed, contextual, and compositional.
A system may generate descriptions of combinations it has never previously encountered by recombining learned attributes and relations.
Consequently, finite computational resources do not entail a one-to-one correspondence between internal states and a predetermined list of human-interpretable concepts.
6.2. Representational collisions
The more important question is whether the model can preserve distinctions relevant to a particular task.
Let represent a model's encoding process, and let and denote two distinguishable external states.
If
and the downstream system has access only to that representation, it cannot distinguish the original states through further processing of that representation alone.
The same practical difficulty may arise when two states are represented differently but insufficiently separated for reliable decoding.
However, three distinct explanations must be considered.
First, the model may be structurally capable of representing a distinction but may not have learned it adequately.
Second, the distinction may be representable, while the available input fails to specify which state occurred.
Third, the representation may collapse the distinction under the relevant computational conditions.
These possibilities imply different remedies: additional training, improved observation, or changes to the representational architecture.
A failure to produce the correct distinction does not, by itself, establish that the distinction is absent from the model's representational capacity.
6.3. The grounding question
Harnad's (1990) symbol grounding problem asks how formal symbols can acquire meaning that is not dependent exclusively on interpretations supplied by external observers [8].
Bender and Koller (2020) further argued that success in modeling linguistic form should not automatically be equated with acquiring meaning in the fuller sense of language understanding [9].
These arguments are significant for LLM evaluation, although they do not establish that every text-trained system is incapable of useful semantic inference.
A model can infer relationships, make context-sensitive distinctions, and generalize compositionally while still lacking access to the external referent necessary to verify a particular interpretation.
The practical limitation is therefore not merely whether a concept can be represented, but whether its intended referent can be identified and validated.
Multimodal observation, external data, and direct interaction can provide additional constraints. They do not eliminate model error, but they can reduce reliance on purely linguistic reconstruction.
7. Iterative Transmission and the Empirical Study of Meaning Drift
7.1. Cultural transmission and reconstructive processes
The propagation of altered information across successive communicators has a substantial research history.
Griffiths and Kalish (2007) formally analyzed iterated learning with Bayesian agents, demonstrating how transmission dynamics interact with learners' prior assumptions [6].
Kirby, Cornish, and Smith (2008) experimentally investigated the cultural transmission of artificial languages, showing that repeated learning can produce increasingly structured and learnable linguistic systems [7].
These findings introduce an important qualification: repeated transmission is not inherently destructive.
Information may be regularized, simplified, or reorganized in ways that improve learnability.
However, improved transmissibility is not identical to improved fidelity to an original observation.
A representation can become easier to reproduce while losing details important to the initial task.
7.2. Direct evidence from contemporary LLM research
A particularly relevant empirical contribution is Mohamed et al.'s (2025) LLM as a Broken Telephone: Iterative Generation Distorts Information, published at the Annual Meeting of the Association for Computational Linguistics [12].
The researchers investigated iterative information transformation through translation-based experiments and reported that distortion accumulated across transmission chains, with outcomes influenced by language choice and chain complexity.
They also examined prompting strategies capable of mitigating degradation.
This study offers direct empirical support for treating iterative LLM communication as a distinct reliability problem.
Its conclusions should nevertheless be interpreted within the experimental conditions investigated. They do not imply that every additional generation necessarily reduces semantic accuracy, nor that all types of meaning deteriorate at equivalent rates.
Related translation research has also identified persistent problems associated with overly literal output. Li et al. (2025) investigated how supervised training contributes to translationese in LLM translation [15].
Together, these studies demonstrate that strong linguistic competence and sophisticated training do not eliminate the distinction between formally acceptable output and contextually appropriate meaning preservation.
7.3. The overlooked dimension: error awareness
Most analyses of semantic degradation naturally emphasize differences between the original and final outputs.
Yet another dimension deserves explicit consideration: the relationship between actual semantic deviation and the system's expressed confidence in its own reconstruction.
Consider two communication chains ending with the same incorrect interpretation.
In the first, the final agent identifies uncertainty and requests clarification.
In the second, the agent presents the interpretation as an established fact.
The semantic deviation may be identical, but the downstream consequences differ considerably.
The second system is more likely to propagate the error into subsequent decisions because it conceals the need for verification.
This suggests that evaluations of iterative LLM systems should measure semantic fidelity and error detectability separately.
8. Toward a Testable Framework for Semantic Fidelity
The preceding analysis motivates a research program focused on meaning preservation across heterogeneous linguistic and computational transformations.
8.1. Experimental design
A controlled experiment could begin with objects or situations represented through independently verifiable attributes.
For example, a visual stimulus might contain a magenta, heart-shaped plush toy, with color, geometry, material, and object type recorded as separate reference attributes.
The stimulus would be described by an initial agent. Its output would then become the sole input to a subsequent agent, creating a transmission chain of configurable length.
Experimental conditions could vary the transmission language, model architecture, decoding settings, speaker-background information, and availability of original evidence.
Four principal conditions would be informative:
- Lexical translation: Agents receive conventional word-level translations with no additional semantic annotations.
- Attribute-preserving transmission: Agents are required to distinguish observed properties from inferred properties.
- Speaker-conditioned transmission: Agents receive relevant information about the original speaker's vocabulary and intended terminology.
- Source-grounded transmission: Agents retain access to the original stimulus or an authenticated representation of it.
The first condition would provide a baseline. The remaining conditions would test whether explicit semantic structure, speaker modeling, and external grounding reduce distortion.
8.2. Measuring semantic preservation
Let the original referent be represented by a set of independently verified attributes:
A reconstructed representation at transmission step produces estimated attributes .
A task-specific semantic preservation score may be defined as
where denotes the importance assigned to attribute .
For applications where approximate matches matter, the binary indicator could be replaced with a graded similarity function.
However, a high attribute-recall score alone would be insufficient. Systems should also be penalized for unsupported additions and contradictions.
For example, converting a generic pink toy into a plastic fashion doll introduces details that were not present in the source.
Such additions may be coherent and plausible, but they are not evidence-preserving reconstructions.
8.3. Measuring undetected error
A second metric should evaluate whether confidence remains calibrated as semantic fidelity changes.
Let denote an agent's estimated probability that its reconstructed representation is sufficiently faithful to the original, and let denote measured fidelity.
A simplified confidence–fidelity discrepancy can be written as
This quantity is illustrative rather than a complete calibration metric. In practice, calibrated event probabilities, proper scoring rules, and repeated trials would be required for statistically meaningful evaluation.
A particularly revealing outcome would be a decline in semantic fidelity accompanied by stable or increasing reported confidence.
Such a pattern would indicate that the system is not merely losing information; it is becoming less capable of recognizing its own information loss.
8.4. Distinguishing three failure mechanisms
Experimental analysis should separate at least three mechanisms.
Transmission loss occurs when the source contains a distinction that disappears during encoding or reconstruction.
Semantic substitution occurs when a missing distinction is replaced with a plausible but unsupported alternative.
Confidence miscalibration occurs when the system's expressed certainty does not adequately reflect the reliability of its interpretation.
These mechanisms may interact, but treating them as a single undifferentiated error category would obscure their causal structure.
A model that preserves attributes but communicates uncertainty poorly requires a different intervention from a model that systematically fails to represent the attributes themselves.
9. Engineering Implications: From Fluent Generation to Traceable Meaning
9.1. Preserve source evidence
Retrieval-augmented generation provides one established architectural approach to supplementing parametric language-model knowledge with external information.
Lewis et al. (2020) introduced a framework combining a neural generator with retrieved non-parametric memory for knowledge-intensive tasks [13].
Retrieval alone does not guarantee accurate interpretation. However, retaining access to original evidence creates opportunities to compare generated claims against their sources.
For iterative systems, this suggests a general design principle: intermediate summaries should not automatically replace authoritative source material.
An agent should be able to recover the evidence supporting a claim rather than relying exclusively on the preceding agent's linguistic reconstruction.
9.2. Separate observations from interpretations
A reliable semantic interface should distinguish between information directly observed in a source, information supplied explicitly by a speaker, and information inferred by the model.
For example, “magenta-colored” may be an observed or explicitly supplied attribute, while “probably a decorative gift” may be an inference.
These statements should not acquire equivalent evidential status merely because they appear in the same fluent paragraph.
A structured representation could attach provenance and uncertainty to individual claims, enabling later agents to distinguish source-derived attributes from generated interpretations.
Such provenance mechanisms would reduce the risk of inference being silently promoted into fact.
9.3. Retain ambiguity when evidence is insufficient
A model should not be forced to select a single referent whenever the input supports multiple interpretations.
Where uncertainty materially affects downstream outcomes, the system should either request clarification or maintain explicit alternatives.
For example:
“The speaker used the term ‘doll,’ but the available context does not establish whether the referent is a plush toy or a human-shaped doll.”
This answer is less superficially decisive than an unqualified interpretation, but it is epistemically more faithful to the available evidence.
The optimization target should therefore not be maximal linguistic specificity under every condition.
It should be specificity proportional to evidential support.
9.4. Evaluate information at the attribute level
Embedding similarity and natural-language inference can provide useful signals concerning the relationship between source and generated text.
Nevertheless, semantic proximity should not be treated as sufficient evidence that every consequential attribute has been preserved.
A system may produce a semantically similar description while changing a numerical value, negation, material property, temporal relationship, or causal condition.
Evaluation should therefore combine broader semantic measures with structured checks of task-critical properties.
For systems handling technical specifications, scientific claims, medical information, or software requirements, these distinctions may be operationally decisive.
10. Discussion: Does Language Constrain Intelligence or Merely Its Expression?
The preceding framework also clarifies a broader question concerning the relationship between language and thought.
If two agents possess different lexical repertoires, they may incur different costs when describing equivalent concepts. Yet this does not establish that one agent's language imposes an absolute limit on the concepts it can entertain.
Language can facilitate conceptual discrimination without necessarily constituting the entirety of conceptual capacity.
Likewise, a model's inability to produce a particular description may arise from inadequate observation, insufficient training, failed retrieval, decoding behavior, or genuine representational limitations.
These explanations must be distinguished empirically.
The broader hypothesis that lexical differences between natural languages determine the long-term intellectual development of entire societies is not established by the evidence considered here.
Indeed, Coupé et al. (2019) found broadly comparable average information rates across 17 spoken languages despite substantial differences in syllabic information density and speech rate [14].
That finding does not prove that all languages are identical in conceptual organization or task-specific expressive efficiency. It does, however, caution against inferring global cognitive superiority from vocabulary size or linguistic form.
A more defensible claim is that the availability of shared conceptual distinctions, explicit notation, verification practices, and reliable records can affect the cost of accumulating and transmitting knowledge.
Scientific notation, technical vocabularies, and standardized measurement systems illustrate how communities develop supplementary representational resources to preserve distinctions that ordinary discourse may leave ambiguous.
For AI systems, the corresponding objective is not necessarily to maximize the number of accessible lexical categories.
It is to preserve the relationships among observations, concepts, utterances, and evidence across successive transformations.
In this sense, the reliability of language-mediated intelligence depends less on the apparent completeness of any particular sentence than on the system's capacity to maintain a recoverable connection between its claims and the information from which those claims originated.
11. Conclusion
The central limitation of language-mediated artificial intelligence is not that linguistic representations are inevitably inaccurate, nor that all knowledge must be reducible to a finite inventory of predefined meanings.
Rather, the difficulty lies in the possibility that plausible linguistic reconstruction can conceal a divergence between intended meaning and received meaning.
This divergence emerges when lexical categories fail to preserve relevant distinctions, when listeners infer speaker intentions from incomplete evidence, when repeated transformations substitute conventional interpretations for source-specific attributes, and when models mistake consistency for correctness.
Contemporary research provides several complementary approaches to these problems. Information theory characterizes constraints on recovering lost information; probabilistic pragmatics models the inference of speaker intentions; iterated learning investigates the effects of repeated cultural transmission; and recent LLM studies demonstrate that iterative generation can produce measurable information distortion.
Nevertheless, the relationship between a model's internal semantic representation, its confidence in an interpretation, and the fidelity of that interpretation to an external referent remains a critical area for further investigation.
The most productive research direction is therefore not simply to make language models produce more precise sentences. It is to make their semantic transformations traceable, verifiable, and appropriately uncertain.
A fluent model can generate a convincing reconstruction of something it has never adequately identified. A reliable model must additionally recognize when the available evidence does not justify such reconstruction.
The distinction is fundamental:
Linguistic coherence establishes that an interpretation is expressible. It does not establish that the interpretation is true to its source.
For future AI systems that increasingly mediate communication, reasoning, and institutional knowledge, preserving this distinction may be as important as improving raw generative capability.
References
The following references are published works with verified bibliographic identifiers. Conceptual frameworks and proposed metrics developed specifically in this essay should not be attributed to these researchers as established findings.
[1] Grice, H. P. (1957). Meaning. The Philosophical Review, 66(3), 377–388. https://doi.org/10.2307/2182440
[2] Rosch, E. (1975). Cognitive representations of semantic categories. Journal of Experimental Psychology: General, 104(3), 192–233. https://doi.org/10.1037/0096-3445.104.3.192
[3] Clark, H. H., & Brennan, S. E. (1991). Grounding in communication. In L. B. Resnick, J. M. Levine, & S. D. Teasley (Eds.), Perspectives on Socially Shared Cognition (pp. 127–149). American Psychological Association. https://doi.org/10.1037/10096-006
[4] Frank, M. C., & Goodman, N. D. (2012). Predicting pragmatic reasoning in language games. Science, 336(6084), 998. https://doi.org/10.1126/science.1218633
[5] Shannon, C. E. (1948). A mathematical theory of communication. The Bell System Technical Journal, 27(3), 379–423; 27(4), 623–656. https://doi.org/10.1002/j.1538-7305.1948.tb01338.x and https://doi.org/10.1002/j.1538-7305.1948.tb00917.x
[6] Griffiths, T. L., & Kalish, M. L. (2007). Language evolution by iterated learning with Bayesian agents. Cognitive Science, 31(3), 441–480. https://doi.org/10.1080/15326900701326576
[7] Kirby, S., Cornish, H., & Smith, K. (2008). Cumulative cultural evolution in the laboratory: An experimental approach to the origins of structure in human language. Proceedings of the National Academy of Sciences, 105(31), 10681–10686. https://doi.org/10.1073/pnas.0707835105
[8] Harnad, S. (1990). The symbol grounding problem. Physica D: Nonlinear Phenomena, 42(1–3), 335–346. https://doi.org/10.1016/0167-2789(90)90087-6
[9] Bender, E. M., & Koller, A. (2020). Climbing towards NLU: On meaning, form, and understanding in the age of data. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 5185–5198. https://doi.org/10.18653/v1/2020.acl-main.463
[10] Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30. https://proceedings.neurips.cc/paper/2017
[11] Farquhar, S., Kossen, J., Kuhn, L., & Gal, Y. (2024). Detecting hallucinations in large language models using semantic entropy. Nature, 630, 625–630. https://doi.org/10.1038/s41586-024-07421-0
[12] Mohamed, A., Geng, M., Vazirgiannis, M., & Shang, G. (2025). LLM as a broken telephone: Iterative generation distorts information. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 7493–7509. https://doi.org/10.18653/v1/2025.acl-long.371
[13] Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W., Rocktäschel, T., Riedel, S., & Kiela, D. (2020). Retrieval-augmented generation for knowledge-intensive NLP tasks. Advances in Neural Information Processing Systems, 33. https://proceedings.neurips.cc/paper/2020
[14] Coupé, C., Oh, Y. M., Dediu, D., & Pellegrino, F. (2019). Different languages, similar encoding efficiency: Comparable information rates across the human communicative niche. Science Advances, 5(9), eaaw2594. https://doi.org/10.1126/sciadv.aaw2594
[15] Li, Y., Zhang, R., Wang, Z., Zhang, H., Cui, L., Yin, Y., Xiao, T., & Zhang, Y. (2025). Lost in literalism: How supervised training shapes translationese in LLMs. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 12875–12894. https://doi.org/10.18653/v1/2025.acl-long.630
Scholarly Status
This essay is a theoretical synthesis and research proposal, not a report of newly conducted experiments. Its mathematical formulations serve different purposes: the data-processing inequality is an established information-theoretic result, while the speaker-conditioned model and proposed semantic-fidelity measures are analytical formulations intended to guide further testing.
The essay's principal contribution is the identification of a research distinction between the degradation of semantic information and the failure to recognize that degradation. Establishing the magnitude, prevalence, and operational consequences of this distinction will require controlled empirical investigation.