RESEARCH DOSSIER
Human Comprehension of Artificial Language
Abstract. Human comprehension of artificial language concerns the psychological processes by which people derive linguistic meaning, infer communicative relevance, evaluate source reliability, and coordinate action when language is generated or mediated by artificial systems. This critical scientific monograph integrates psycholinguistics, cooperative inference, common-ground theory, human–computer interaction, anthropomorphism, and experimental research on AI disclosure, confidence calibration, social responsiveness, human detection of synthetic text, and conversational repair. It differentiates proposition understanding, source recognition, pragmatic interpretation, perceived empathy, confidence, and demonstrated task success. Historical foundations are considered alongside controlled source-label experiments, referential games, observational dialogue analyses, comparative pragmatics, and emerging model-grounding research. Limitations involving sampling, rapidly changing model behavior, construct validity, and cross-linguistic transportability are evaluated. The synthesis identifies testable research priorities for measuring verifiable interpretation and communicative recovery across the Artificial Era. This is a critical narrative evidence synthesis, not a new experiment, systematic review, or meta-analysis.
Keywords: human comprehension; artificial language; AI-generated text; AI-mediated communication; psycholinguistics; pragmatic inference; common ground; referential grounding; dialogue repair; source attribution; AI disclosure; credibility; trust calibration; anthropomorphism; synthetic text detection; human–artificial interaction; Artificial Era
The Scientific Object of Human Comprehension of Artificial Language
Human comprehension of artificial language is the set of psychological processes through which people assign meanings, identify referents, draw inferences, evaluate claims, and coordinate action when the linguistic material they encounter has been produced or materially transformed by an artificial system. The unit of analysis may be a sentence, a multistep explanation, a spoken response, or a sequence of turns in which the reader or listener revises an earlier understanding. A scientific account requires observable outcomes: selecting a referent, executing an instruction, paraphrasing an argument, detecting an inconsistency, or repairing a conversational misunderstanding. The distinction between form and grounded meaning is central to the account of language-model understanding developed by Bender & Koller (2020).
Artificial language here includes generated or substantially AI-mediated natural-language text and speech, rather than only deliberately constructed human languages or formal programming languages. The human comprehender may know the source, suspect it, or remain unaware of it. Each condition changes the available evidence concerning credibility, communicative intent, and the likely reliability of the utterance. The research object therefore joins psycholinguistic comprehension with social attribution and epistemic evaluation while keeping these outcomes analytically separate. AI-mediated communication provides an established framework for identifying the extent to which an algorithm modifies, generates, or selects communicative content Hancock et al. (2020).
This monograph synthesizes historical theories, experimental psycholinguistics, human–computer interaction, social-cognitive research, and recent studies of generated language. The review is a structured critical narrative synthesis, not a preregistered systematic review or a meta-analysis. Its main questions concern which ordinary comprehension processes transfer to artificial utterances, where source expectations change interpretation, how repair contributes to shared task performance, and which experimental designs separate fluency from accuracy. A model may produce a plausible pragmatic response under a laboratory prompt, but such performance is a different observation from a human partner's demonstrably successful interpretation Hu et al. (2023).
Foundations in Psycholinguistics and Communicative Inference
Comprehension involves more than decoding vocabulary. An interpreter combines word meanings, grammatical constraints, world knowledge, immediate context, expectations about discourse, and assumptions about a source. Grice's analysis of conversational cooperation introduced a productive distinction between what an utterance directly expresses and what its use in context suggests Grice (1975). Later probabilistic formulations specify how listeners update hypotheses about intended referents given utterance choices, contextual alternatives, and assumptions about communicators Frank & Goodman (2012). These theoretical resources are applicable to AI-generated messages because human readers still make the inferential moves required to interpret a sentence as a situated act.
The transfer from human conversation to synthetic output requires an additional conceptual step. A human speaker's utterance may reveal a belief, intention, or experience; a language-model output may be conditioned by prompts, training distributions, decoding policies, and external data without the same evidential connection to a personal communicative intention. The interpreter nevertheless tends to construct a useful model of the source from conversational behavior. Anthropomorphism theory helps explain why human knowledge about agents can be applied to technologies that exhibit familiar social cues Epley et al. (2007). An inference about what an utterance means should consequently be tested independently of an inference about what its originator is like.
A robust theory traces processing from initial lexical access to a context-sensitive interpretation and then to action. An experiment that asks whether a participant believes a claim measures an epistemic judgment; a referential game asking which object the participant chooses measures another component of comprehension. These outcomes can diverge even when the same sentence is presented. Evidence from collaborative reference shows that partners construct temporary linguistic conventions for specific tasks rather than relying exclusively on fixed semantic labels Clark & Wilkes-Gibbs (1986). The practical implication for artificial communication is that successful language interpretation may depend on shared interaction history, which is experimentally manipulable.
Historical Development of Computer-Mediated Conversation
Research into computer interlocutors developed from human–computer interaction, conversational analysis, and studies of dialogue mediated by devices. Brennan's examination of conversation with and through computers described how interlocutor expectations and channels shape human communicative behavior Brennan (1991). This work matters historically because it separates the visible linguistic form from the social understanding that a participant brings to an exchange. A keyboard interface may display the same text under contrasting beliefs about the source, permitting controlled comparison of interpretation and conversational responses.
Social-response experiments found that people sometimes treat computers according to familiar interpersonal norms, even when such conduct is difficult to explain as a deliberate belief that the machine is human Nass & Moon (2000). Subsequent studies investigated how anthropomorphic cues and communicative-agency framing shape perception of conversational agents Araujo (2018). The resulting literature contains heterogeneous dependent variables, including liking, perceived social presence, perceived competence, trust, and compliance. These variables cannot be treated as direct measures of semantic comprehension, although they can alter conditions under which comprehension occurs.
The emergence of large-scale text generation changed what could be held constant in these experiments. Researchers can now compare responses to matched messages whose actual origin, disclosed origin, style, and correctness vary independently. Studies of how people identify generated language show that source judgments rely on fallible heuristics and task-specific experience Jakesch et al. (2023). The historical trajectory therefore connects classic questions about mediated conversation to the contemporary problem of interpreting detailed, adaptive, and apparently coherent utterances. The research challenge remains identifying psychological mechanisms from controlled behavior rather than inferring them merely from impressive system outputs.
Lexical Meaning, Syntax, and Discourse Coherence
Words and sentences in generated text are ordinarily processed by the same human linguistic system that handles other written and spoken material. Readers exploit syntactic dependencies, lexical probabilities, and discourse structure to construct interpretations. Electrophysiological research demonstrated that semantic incongruity can alter time-locked brain responses during sentence reading, providing evidence that linguistic expectations are engaged during comprehension Kutas & Hillyard (1980). Such findings motivate physiological investigation of AI-mediated materials, while the presence of an experimental neural response alone does not determine whether the communicated proposition was accurate.
Grammatical fluency reduces some processing demands, but ease of processing and factual correctness are different variables. An artificial message may maintain coherent syntax while inserting a fabricated citation, an unsupported implication, or an incorrect referent. The observer may reproduce its argument faithfully yet accept an unwarranted conclusion. The distinction becomes particularly important for explanations of uncertainty, where ordinary linguistic markers may be interpreted as quantitative indicators even when a model's confidence does not track correctness. Studies of users' interpretation of language-model knowledge and confidence document the need to distinguish model discrimination from users' discrimination Steyvers et al. (2025).
An experimentally useful approach presents multiple versions of a generated passage while holding factual content, length, and linguistic complexity as constant as the task allows. Comprehension questions should separately test literal recall, inferential integration, recognition of contradictions, and source evaluation. Studies of text simplification investigate meaning preservation through reader-oriented evaluation and illustrate why perceived readability should not substitute for evidence of retained propositions Agrawal & Carpuat (2024). For this dossier, the central question is whether an artificial utterance supports a verifiably appropriate mental representation, and under what conditions a fluent surface encourages an inaccurate one.
Reference, Common Ground, and Contextual Alignment
An utterance refers successfully when the listener identifies the intended entity or event within the relevant situation. In human dialogue, reference is collaborative: speakers propose descriptions and listeners accept, reject, or negotiate them Clark & Wilkes-Gibbs (1986). A spoken instruction about an object on a table therefore depends on the inventory of objects, shared visibility, preceding turns, and conventions established during the task. The same dependencies arise in conversations with artificial agents, but the system may maintain an incomplete or inconsistent representation of the environment. A transcript that appears coherent cannot alone establish that common ground was maintained.
Perspective-taking experiments show that listeners integrate common-ground information with what is linguistically and visually available; the timing and reliability of this integration are empirical questions Keysar et al. (2000). Clark and Brennan's grounding framework treats acknowledgments and evidence of understanding as methods for bringing interlocutors toward sufficient mutual coordination for the current purpose Clark & Brennan (1991). In interactions with generated language, comparable evidence may consist of a correct referent selection, a successful task step, or an explicit correction followed by revised output. The adequacy criterion depends on the task rather than a global impression of conversational intelligence.
Future work should compare human–human and human–AI communication in matched reference games that record the same utterances, visual scenes, opportunities for clarification, and success criteria. Researchers can manipulate whether the source has the required visual access or memory while separately varying participants' beliefs about that access. The outcome should identify when people incorrectly assume the system shares a perceptual context, how promptly they recognize this error, and whether a clarification restores alignment. Probabilistic accounts of pragmatic reasoning supply explicit models for how alternative descriptions influence referential inference Goodman & Frank (2016).
Pragmatic Interpretation of Artificial Utterances
A generated sentence can be grammatically literal while pragmatically performing a request, warning, recommendation, reassurance, or refusal. Listeners infer these acts by considering discourse expectations, available alternatives, and their understanding of the communication situation. Grice's framework identifies relevance and informativeness as important assumptions in such inference Grice (1975). Human comprehension of artificial language therefore depends on a constructed communicative context: whether the reader is asking for advice, requesting evidence, testing a system, or interpreting a synthetic contribution to another person's message. The same lexical sequence may invite different inferences in each setting.
Formal models of pragmatic interpretation can treat comprehension as an inference over possible communicative goals and alternative utterances. Frank and Goodman showed how referential behavior in language games can be modeled probabilistically Frank & Goodman (2012), while Goodman and Frank reviewed broader probabilistic approaches to pragmatic interpretation Goodman & Frank (2016). These models offer testable predictions concerning how listeners weigh an utterance's wording against expected informativeness. Applying them to artificial agents requires care because a language model's output policy and a human speaker's goal structure may differ while producing similar sentences.
Comparative experiments should manipulate both linguistic alternatives and beliefs about the generation process. A participant might receive a short answer from a tool optimized for minimal length, or the same answer attributed to a human expert seeking to be maximally informative. The interpretation can be measured with questions about implied certainty, excluded possibilities, and intended recommendations. Studies comparing human pragmatic judgments with language-model responses identify overlapping patterns and systematic differences across nonliteral phenomena Hu et al. (2023). Such comparisons are informative when they remain specific to the tested materials and do not equate matching response patterns with identical psychological mechanisms.
Attribution of Agency, Intention, and Social Presence
Human readers often treat a linguistic source as an interlocutor and construct expectations about its knowledge, motives, and communicative competence. Anthropomorphism provides one account of this process: elicited agent knowledge, the motivation to explain an entity's behavior, and the desire for social connection can contribute to humanlike attribution Epley et al. (2007). Text produced by an artificial system can activate these processes through first-person pronouns, responsive turn-taking, politeness, and personalization. Attribution is an observable human judgment, whereas a claim concerning the system's internal experience requires a separate body of evidence.
Research on responses to computers found that social norms can influence behavior toward machines even in constrained interfaces Nass & Moon (2000). Araujo investigated how anthropomorphic design cues and framing of communicative agency affect evaluations of conversational agents Araujo (2018). The strongest interpretation of this literature is that source presentation can modify the interactional environment and the user's expectations. It does not license treating every positive social evaluation as evidence of better semantic decoding. Studies need separate items for whether an utterance was understood, whether the source was perceived as intentional, and whether the response felt socially responsive.
An effective design independently manipulates actual source, displayed label, and expressed conversational style. For example, identical text can be presented as human-written or system-generated; distinct versions of the text can vary in personalized language while preserving the factual claim. The design then measures referential accuracy, perceived intention, trust, and downstream decisions. A factorial separation makes it possible to examine whether anthropomorphic framing changes interpretation itself or primarily changes reported feelings and expectations. The broader framework of AI-mediated communication clarifies why authorship, algorithmic mediation, and reader attribution deserve independent measurement Hancock et al. (2020).
Source Labels, Disclosure, and Epistemic Framing
A source label supplies information about the provenance of a message before or during its interpretation. For ordinary readers, an AI label may alter the expected quality of the text, the motives attributed to its production, and the need for verification. Experiments on headlines labeled AI-generated have shown that such disclosures can reduce perceived accuracy and willingness to share even when the content is true or originally human-written Altay & Gilardi (2024). The measured outcome is a judgment about a message, not a universal change in the participant's capacity to comprehend its literal meaning.
Evidence concerning AI disclosure is not uniform across materials, populations, or dependent measures. In a randomized web-based study of misinformation, Li and Yang examined how labeling affected perceived accuracy, credibility, and sharing intentions Li & Yang (2024). The relevance of disclosure depends on its wording, the presumed institution, the perceived stakes, and prior familiarity with AI systems. A study about political headlines cannot automatically establish the same effects for a technical explanation or a supportive message. It is essential to specify whether the source is genuinely artificial, whether the disclosure is accurate, and whether the participant reads it before interpretation.
The experimental problem is to separate a label's effect on prior beliefs from its effect on processing the substantive content. Researchers can collect a comprehension response before a disclosure, then a credibility judgment afterward, and compare this condition with participants informed of the source at the outset. A third condition can provide information about how the text was generated and checked. This design distinguishes initial semantic inference, revision of credibility, and informed reliance. The importance of these distinctions is evident in the larger body of work on AI-generated language and human detection heuristics Jakesch et al. (2023).
Fluency, Explanatory Detail, and Confidence Calibration
Fluency is an attribute of expression; calibration concerns the correspondence between expressed or perceived certainty and the probability that a claim is correct. Generated language can exhibit fluent grammatical organization even when its underlying answer is inaccurate. Steyvers and colleagues studied how users interpret model answers and communicated uncertainty, identifying gaps between model and human confidence assessments Steyvers et al. (2025). This research motivates measuring whether readers can discriminate correct from incorrect answers rather than relying only on their average confidence or their subjective assessment of clarity.
Explanatory length introduces a further difficulty because detailed justifications can produce a sense of understanding without improving evidential support. An extended answer may contain many connected steps, making it easier for a reader to follow, while the unsupported premise remains unchanged. To isolate such effects, experiments can hold the answer and its correctness fixed while manipulating whether the system provides a terse answer, a verbal confidence phrase, or a stepwise explanation. Comprehension tests should identify the precise claim, its assumptions, and the conditions under which it could be wrong Steyvers et al. (2025).
Confidence calibration in a human–AI dialogue is a relation among model behavior, linguistic cues, individual interpretation, and action. A participant's confidence may rise after a polished explanation even when the task evidence remains the same. The research therefore requires accuracy-conditioned analysis: confidence on correct answers, confidence on incorrect answers, the tendency to seek corroboration, and the cost of erroneous reliance. This approach connects epistemic evaluation to meaning construction without collapsing them into one vague variable. The AI-mediated communication literature helps distinguish the interface's persuasive properties from the provenance and informational validity of its content Hancock et al. (2020).
Perceived Responsiveness and the Experience of Being Understood
People evaluate communicative exchanges partly by whether a response reflects their stated concerns and emotional state. This judgment belongs to the relational and affective consequences of comprehension, but it can diverge from accurate interpretation of factual content. Yin, Jia, and Wakslak experimentally separated the actual origin of supportive responses from participants' beliefs about their source, finding that an AI label could diminish the feeling of being heard even when message content was supportive Yin et al. (2024). The finding shows that source attribution can affect social interpretation independently of the utterance's linguistic surface.
An empathetic-sounding message may acknowledge an emotion, organize the recipient's narrative, or provide suggestions. Each function can be evaluated by distinct methods. Independent coders can assess whether the response accurately reflects the situation supplied by the participant; recipients can rate whether they feel validated; and subsequent tasks can test whether the advice was understood or applied. These measures can disagree. Social presence and anthropomorphic cues are potential mediators or moderators, but their involvement must be tested rather than inferred from a positive rating Araujo (2018).
The psychological significance extends beyond emotional support. A user may interpret a system's conversational repair as care, or may read a generic personalized phrase as recognition of a stable relationship. Such inferences help explain satisfaction and continued engagement, but their appropriateness depends on the actual interaction history. Research designs should test the match between system access to personal information and users' assumptions about that access. This produces a more precise account of the perceived responsiveness explored in human–AI communication Yin et al. (2024), while preserving the separate scientific question of whether an artificial system can establish durable social reciprocity.
Sequential Structure and the Dynamics of Human–AI Dialogue
Comprehension in conversation is sequential because each turn establishes new expectations and changes the interpretation of preceding words. A request may be initially ambiguous, become clear after a reply, and then require revision when the reply reveals a mistaken assumption. Clark and Brennan described conversational grounding as a process through which interlocutors accumulate evidence that their contributions have been understood adequately for current purposes Clark & Brennan (1991). In artificial dialogue, the same user may experience an exchange as coherent because the system maintains lexical continuity, even when it has silently substituted a different task objective.
Temporal continuity should therefore be operationalized through task states rather than through the presence of repeated phrases. Researchers can map each turn onto a reference set, asserted propositions, unresolved questions, and acknowledged corrections. A system that restates the user's preference may be perceived as attentive, yet a later answer may show that a key constraint was lost. Early human–computer conversation research already established that source expectations shape how people use acknowledgments and continuity cues Brennan (1991). Modern systems allow longer observations of this phenomenon but require careful control of changing model versions and context windows.
Longitudinal designs should test whether users learn when to clarify, verify, or reset an artificial exchange. The order of earlier successes and errors may change both the user's predicted reliability and the effort spent maintaining the dialogue. A single satisfaction score after a conversation cannot distinguish successful comprehension from accommodation to repeated system mistakes. Researchers can instead code accuracy at each turn, number of unresolved reference errors, and the user's willingness to initiate correction. The investigation of repair in conversational agents provides a direct methodological starting point Corti & Gillespie (2016).
Repair, Clarification, and the Recognition of Misunderstanding
A repair sequence begins when an interlocutor treats an earlier contribution as problematic and attempts to restore a workable interpretation. In everyday conversation, repair may concern hearing, reference, factual assumptions, or the pragmatic force of a response. Corti and Gillespie experimentally observed greater user-initiated repair under conditions in which an artificial interlocutor was represented as human, demonstrating that source framing can influence the effort people invest in restoring mutual understanding Corti & Gillespie (2016). The interpretation pertains to specific experimental conditions rather than to an intrinsic quality of all conversational systems.
Repair can be measured through the frequency of clarification requests, their specificity, the time needed to recover from an error, and the proportion of subsequent task steps completed correctly. Some user attempts may be simple repetitions; others introduce new context, narrow alternatives, or explicitly challenge a mistaken claim. These actions have different informational content. Studies of virtual assistants also analyze how platforms handle user-initiated repair across spoken interactions and language settings Galbraith (2024). The comparison must account for recognition errors, dialogue policy, semantic misunderstanding, and users' beliefs about what the system can repair.
A critical question is whether users compensate for the system's limitations by increasing their own communicative work. A successful final answer could reflect a sophisticated artificial partner, a highly persistent human, or their joint adaptation. Designs that record the complete sequence and code both sides can apportion explanatory responsibility. The same experiment should include an independently verified target interpretation and a matched human comparison when feasible. Grounding research shows why mutual evidence is essential for interpreting successful coordination Clark & Brennan (1991), while computer-mediated interaction research demonstrates that interface expectations influence how such evidence is sought Brennan (1991).
Acknowledgment, Clarification, and Grounding Gaps
Conversational systems can generate acknowledgments such as agreement, summary, or expressions of understanding without demonstrating that the relevant task constraints have been retained. In human dialogue, acknowledgments normally serve as evidence that a contribution has been received and provisionally accepted. Their practical adequacy depends on whether they support later coordinated action Clark & Brennan (1991). A system that says it understands an instruction yet acts on a different referent creates a gap between linguistic grounding cues and operational grounding. This gap is especially consequential when users infer competence from smooth turn-taking.
Analyses of grounding gaps in language-model generation show the importance of evaluating the relation between linguistic output and the contextual conditions from which an answer is produced Shaikh et al. (2024). For human comprehension, the complementary question is whether the person recognizes the missing context. A controlled task can vary whether the artificial system has access to an image, a document, or prior conversational history, while measuring the user's beliefs about that access. A correctly phrased answer under unavailable evidence may still be persuasive if the user incorrectly assumes that the system saw the relevant source.
Scientific assessment requires separating several types of failure: the user's initial failure to express a constraint, the system's failure to use an available constraint, and the user's later failure to detect the discrepancy. Repair and evaluation should be analyzed as coupled but distinguishable processes. The earliest usable measure is a pre-action comprehension check; later measures concern whether a mismatch is detected and resolved. In referential dialogue, collaborative descriptions can become efficient through a sequence of negotiated conventions Clark & Wilkes-Gibbs (1986). An artificial agent's apparent adoption of such conventions should be tested across turns and novel examples.
Collaborative Tasks and Human–Artificial Referential Games
Referential games make language comprehension observable by requiring a listener to select an object, location, or action from a set of possibilities. The experimental environment determines which alternative descriptions are sufficient and what contextual information is shared. Classic collaborative-reference research demonstrates how partners refine descriptions over repeated trials Clark & Wilkes-Gibbs (1986). For artificial-language comprehension, researchers can introduce a language-model-generated referring expression and ask whether a human selects the intended target, then manipulate ambiguity, distractors, prior agreements, and whether a clarification option is available.
Recent computational studies such as CoGen investigate feedback loops between comprehension and generation in interactive referential settings Gul & Artzi (2024). These systems offer useful experimental tools but should be distinguished from psychological studies of independent human readers. An artificial agent can be optimized to produce successful references on a particular benchmark while a human user's confidence remains poorly calibrated. Conversely, the reader can compensate for a weak generated description by inferring the intended reference from task context. The same final accuracy can arise through different combinations of system performance and human effort.
A strong design uses multiple dependent measures: first-choice accuracy, probability assigned to candidate referents, response latency, number of clarifications, and subsequent transfer to unseen objects. Trial-level mixed-effects analysis can model participants and items while avoiding unjustified independence assumptions. Researchers should predefine whether performance improvement over repeated trials indicates lexical convention learning, richer shared context, or narrowing of alternatives. Pragmatic language-game models provide predictions that can be compared with these outcomes Frank & Goodman (2012). A useful field study would also test whether artificially generated descriptions preserve effectiveness outside simplified closed-world tasks.
Speech, Multimodal Context, and Human Perceptual Constraints
Artificial language is often delivered through voice, where comprehension also depends on auditory perception, timing, speaker identity cues, and the physical environment. A spoken answer may be masked by noise, pronounced with unfamiliar prosody, or separated from a visible referent. Psycholinguistic findings show that linguistic structure is tracked during connected speech and that cortical responses can reflect hierarchical properties of linguistic input Ding et al. (2016). These observations support using time-resolved methods in artificial speech research, but do not establish that the auditory signal's source is reliably inferred from its acoustic form.
Multimodal interpretation involves aligning words with objects, gestures, visual displays, or a shared document. An artificial voice assistant may name an object visible to the user while lacking direct visual access, or may combine computer-vision output with a linguistic description. The human listener's error may therefore concern hearing, semantic interpretation, or incorrect assumptions about sensor access. Evidence from common-ground research indicates that perspective information shapes reference resolution under realistic contextual constraints Keysar et al. (2000). Experimental studies should specify what is available to the system, what is available to the user, and what each believes about the other's access.
Voice interfaces introduce distinct opportunities for repair. Users may repeat an utterance to overcome recognition errors, change vocabulary to fit an anticipated machine grammar, or provide an explicit alternative referent. These adaptations should not automatically be interpreted as reduced language competence; they can reflect rational adjustment to a communication channel. Comparative research on virtual-assistant repairs highlights the variation in how systems handle user initiatives Galbraith (2024). A multimodal protocol should therefore distinguish the acoustic recognition stage, the semantic stage, the communicative negotiation stage, and the outcome of the task.
Human Detection of Artificially Generated Text
Detection studies ask whether readers can identify the origin of linguistic material, a judgment related to source attribution rather than to comprehension of a text's propositions. Jakesch, Hancock, and Naaman demonstrated that people rely on heuristics for distinguishing AI-generated from human-written language and that these cues may be misleading Jakesch et al. (2023). A reader who correctly identifies a passage as artificial may still misunderstand its argument, while another who misidentifies the source may accurately understand every claim. Detection accuracy should therefore be analyzed alongside, and separately from, semantic comprehension and factual evaluation.
Human detection performance changes with familiarity with generation tools, genre, model quality, text length, and the source distribution in the experimental set. Russell, Karpinska, and Iyyer reported strong detection performance among people who regularly use ChatGPT for writing, under a specific task and material selection Russell et al. (2025). Their result constrains broad generalizations that all readers are equally unable to detect synthetic text. It also suggests that the relevant psychological construct may involve learned sensitivity to style and distributional patterns rather than a single stable ability that applies to all writing environments.
Experimental design must distinguish a balanced laboratory classification exercise from an everyday context in which most passages may come from one source. Accuracy, false-positive rates, sensitivity, response bias, and confidence should all be reported. If a participant is told that half the samples are AI-generated, this information itself changes response strategies. Follow-up interviews can identify the heuristics used, but retrospective explanations should be compared with actual cue manipulations. The epistemic relevance of source detection depends on whether knowing the source helps a reader assess meaning and validity in the task at hand Jakesch et al. (2023).
Individual Differences in AI Language Literacy and Expertise
People differ in their background linguistic knowledge, domain expertise, expectations of artificial systems, and experience with generated text. These differences can change both the comprehension strategies used and the standards for accepting an explanation. A novice may depend heavily on fluency and contextual plausibility, whereas an expert may notice a contradiction in specialized terminology. Evidence from experienced users detecting generated writing suggests that sustained contact with a system can change observable discrimination performance Russell et al. (2025). Such results warrant careful analysis of selection effects: frequent users may differ from occasional users before experience begins.
To distinguish experience from prior ability, studies should measure baseline comprehension, general reading proficiency, familiarity with the specific interface, and independent domain knowledge. Random assignment to structured training in source evaluation or clarification can then test whether a skill improves and transfers to different models or text genres. Self-reported AI literacy is informative as a belief measure but cannot substitute for observed performance. Researchers should avoid using one composite score for comprehension, trust, source detection, and critical evaluation. Each construct requires its own valid assessment and reliability evidence.
Age, education, language background, disability access, and occupational demands can affect participation and task performance. Generalizability therefore requires recruitment beyond convenience samples of English-speaking university participants. The literature on uncertainty communication shows why familiarity with verbal probability expressions is relevant for interpreting system confidence Steyvers et al. (2025). A reader may also learn domain-specific ways to verify claims while remaining susceptible to misleading emotional or social cues. Research programs should model individual differences as possible moderators with prespecified hypotheses and test for measurement invariance across groups.
Developmental, Linguistic, and Cultural Generalizability
Most contemporary experiments on AI-generated English implicitly rely on adult language proficiency and established cultural conventions for politeness, expertise, and institutional authority. Human comprehension of artificial language, however, can vary with developmental stage, literacy experience, and language socialization. A child's ability to infer pragmatic intent, a multilingual adult's sensitivity to translated ambiguity, and an older user's expectations about digital agents may differ for legitimate developmental and contextual reasons. Theoretical accounts of conversational inference specify how alternatives and expectations shape meanings Goodman & Frank (2016), but these inputs need independent measurement in each population.
Cross-linguistic research should test whether model-generated utterances support equivalent pragmatic inferences when languages differ in information structure, honorifics, grammatical evidentiality, or culturally conventional indirect speech acts. Simply translating an English stimulus may modify what is ordinarily implied, making a word-for-word comparison misleading. Studies of repair with voice assistants across language contexts demonstrate why linguistic affordances and system behavior must both be considered Galbraith (2024). Comparable meaning should be established through expert review, piloting, and task-grounded interpretation rather than through nominal translation equivalence.
Developmental designs require age-appropriate consent, participant protection, and tasks that reflect actual communicative ability. Researchers might compare children's judgments of a system's knowledge access with their accuracy in following generated instructions, or study how school environments teach verification practices. The hypothesis that younger users anthropomorphize conversational interfaces more strongly must be tested under explicit measures and culturally appropriate conditions. Anthropomorphism theory provides plausible motivational mechanisms Epley et al. (2007), but it does not prescribe an invariant age gradient. Longitudinal research would help distinguish developmental change from changes in technology exposure.
Experimental Paradigms and Measurement Validity
Research methods should begin by specifying an observable comprehension target and the competing psychological explanations. For a generated instruction, the target may be selecting the correct referent; for an explanation, it may be reconstructing the causal argument; for a supportive response, it may be identifying the emotion or request acknowledged. The researcher must then separate literal comprehension, inferential interpretation, credibility assessment, and behavioral reliance. Collaborative-reference experiments provide a rigorous way to define correct or incorrect communicative outcomes Clark & Wilkes-Gibbs (1986), while pragmatic language games offer controlled materials that connect task demands to predicted inference patterns Frank & Goodman (2012).
Factorial experiments are especially valuable because actual origin, perceived origin, response style, factual correctness, system access to context, and opportunity for repair can be manipulated independently. A crossed design can compare a human-written answer and a system-generated answer under accurate or inaccurate labels, with all readers completing the same comprehension and verification measures. Yin and colleagues separated the effect of actual response source from belief about source in an experiment on feeling heard Yin et al. (2024). That approach illustrates the power of source-belief manipulations, while the measure of feeling heard remains distinct from factual understanding.
Psychometric quality must be assessed for every scale. A newly invented item asking whether a user 'understood the AI' may combine satisfaction, fluency, confidence, and actual comprehension. Researchers should report reliability, factor structure where appropriate, item difficulty, convergent and discriminant validity, and associations with performance-based outcomes. Eye movements, response times, verbal protocols, and physiological measures add process information but need task-specific interpretation. Neural responses to semantic incongruity reveal a relevant dimension of language processing Kutas & Hillyard (1980), yet a neural difference must not be treated as an unambiguous index of factual verification.
Causal Inference, Statistical Analysis, and Reproducibility
An experiment can identify a causal effect of a manipulated source label on a measured judgment when assignment, implementation, and analysis support that interpretation. It cannot automatically establish that label-induced judgments changed language comprehension at an earlier stage. Studies of AI-labeled headlines and perceived accuracy are informative because disclosure was manipulated, but their outcomes should be reported exactly as measured Altay & Gilardi (2024). Similarly, a correlation between trust and task success could reflect prior expertise, item difficulty, interface design, or a common response tendency rather than a direct causal pathway.
Many dialogue datasets contain repeated turns from the same individuals and multiple items generated by the same systems. Treating each turn as an independent observation understates dependence. Multilevel modeling or appropriate clustered inference should reflect participants, scenarios, tasks, and model variants. Selection and survivorship are additional concerns: people who tolerate repeated errors may stay in an extended study, while dissatisfied participants leave. Researchers should preregister primary outcomes, explain exclusions, report uncertainty intervals, disclose analysis choices, and examine sensitivity to alternative coding rules. These practices make both positive and null findings more interpretable.
Reproducibility faces a special challenge when commercial systems are updated without public version control. Researchers should record the date, model identifier, interface, prompts, parameters, available tools, instructions, and whether external information retrieval was enabled. They should archive the actual messages with privacy safeguards and distinguish platform behavior from human interpretation. Comparative pragmatic benchmarks have documented model-specific patterns Hu et al. (2023), but their generalization is conditional on tested model versions and task formats. A cumulative research program requires publicly inspectable materials where lawful, independent replications, and a clear boundary between preregistered tests and exploratory interpretation.
Comparing Human Comprehension and Artificial Pragmatic Performance
A human's understanding of a message and an artificial system's performance on a pragmatic benchmark are distinct scientific variables. Hu and colleagues compared human responses and language-model outputs across several pragmatic phenomena, finding useful similarities alongside difficulties with social expectation violations Hu et al. (2023). The comparison is important for identifying the linguistic cues that support successful interpretation. Nevertheless, a model's selection of the same answer as a participant does not establish that the model shares the participant's psychological representations, reasons, or experience of the task.
An explanatory comparison should begin with matched stimuli and output options, then inspect response distributions rather than only average accuracy. Researchers can ask whether a model and humans make the same kinds of errors when referents are ambiguous, when politeness changes the literal reading, or when contextual information is missing. Changes in prompting, few-shot examples, answer format, and evaluation settings can change machine outputs independently of the task's psychological demands. The conceptual distinction between linguistic form, meaning, and communicative grounding is central to interpreting such comparisons Bender & Koller (2020).
For the present dossier, system performance is relevant chiefly as an input to human comprehension. When a machine reliably generates appropriate clarification questions, users may need less repair; when it produces locally coherent but inaccurate assertions, users may fail to detect a global mismatch. The psychological outcome is measured on the human side. A crossed design can present the same participant with language produced by systems that differ in benchmark performance while controlling length and style. This provides direct evidence about which properties of artificial output facilitate human understanding and which encourage misplaced confidence Steyvers et al. (2025).
Meaning, Grounding, and the Limits of Linguistic Form
Linguistic output is interpreted in relation to the world and to actions that the interpreter can take. Bender and Koller argued that successful modeling of linguistic form and distributional patterns should not by itself be taken as evidence of the kind of grounded understanding associated with communicative meaning Bender & Koller (2020). This argument establishes a conceptual and methodological distinction for research design. It does not imply that humans cannot understand AI-generated propositions; rather, it directs attention to the environment, intentions attributed by users, and the observable success of communication.
Shanahan's analysis of how people talk about large language models similarly urges care with human-centered vocabulary such as knowing, believing, or wanting when describing generative systems Shanahan (2024). Readers may make such attributions because the conversational interface resembles familiar interpersonal dialogue. A scientific study can document these attributions and test their consequences while keeping model implementation, human interpretation, and philosophical ascriptions at distinct levels. This discipline is especially important when a text offers an apparently first-person account of past events or stable personal commitments.
A practical grounding study can contrast a system's statements about a document it genuinely received with otherwise similar statements generated without document access. Participants then answer questions about which claims are directly supported, which are inferred, and which require external verification. The resulting difference reveals whether human comprehension tracks the evidential relation between language and context, rather than only the surface plausibility of the prose. Experimental work on grounding gaps in generated language provides methods and examples for studying this distinction Shaikh et al. (2024). The task does not require a prior decision on artificial consciousness or personhood.
Education, Information Seeking, and Professional Communication
Educational interactions with AI-generated explanations place comprehension and epistemic judgment under practical pressure. A student may extract a correct definition from a fluent explanation while misunderstanding the example intended to illustrate it. A professional may follow each step of a generated recommendation yet fail to verify the factual premise. Research on the gap between model knowledge and human beliefs about that knowledge suggests that fluently communicated uncertainty can lead users to miscalibrate reliance Steyvers et al. (2025). Studies should test the actual acquisition of transferable knowledge through delayed questions, novel problems, and independently verified applications.
An educational experiment can compare feedback from a human tutor, a generative system, and a fixed reference source using matched content and randomized disclosure. The primary outcome might be performance on a later test of underlying concepts; secondary outcomes can include time, self-explanation quality, confidence, satisfaction, and the proportion of external checks. Source labels can alter perceived credibility even when the information is held constant, as demonstrated in experimental work on labeled headlines Altay & Gilardi (2024). This makes it important to report when a learning effect reflects content, framing, or the reader's use of verification.
Professional settings introduce additional costs from errors and ambiguous authority. The appropriate research response is to specify the decision, evidence available, and competence required, not to derive generic claims about professional safety from nonprofessional samples. In high-stakes domains, independent validation and expert review form part of the observed communication ecology. A system can assist comprehension by rephrasing complex materials, but the resulting explanation should be tested for meaning preservation and task accuracy Agrawal & Carpuat (2024). The same principle applies to multilingual translation, summaries of research, and procedural instructions with hidden constraints.
Artificial Era and the Changing Ecology of Language Reception
The Artificial Era creates communication environments in which linguistic material can be generated, selected, revised, translated, and transmitted by artificial systems across many contexts. The human reader must interpret not only a sentence but often its provenance, intended audience, access to evidence, and relation to an evolving chain of revisions. AI-mediated communication theory treats these degrees of computational involvement as consequential features of the communicative environment Hancock et al. (2020). For psychology, the key empirical problem is how people adapt established linguistic and social-cognitive processes to this rapidly changing infrastructure.
Historical theories of conversational cooperation, collaborative reference, and social attribution remain useful because they specify mechanisms and observable predictions Grice (1975) Clark & Wilkes-Gibbs (1986) Epley et al. (2007). Contemporary artificial systems change the distribution of messages and the apparent responsiveness of sources, but established psychological constructs still require valid measurement. A reader may infer communicative intention from a reply, attribute knowledge to a conversational partner, or revise their belief when a correction arrives. Each process can be measured without reducing the entire encounter to a single score of human–AI interaction.
The long-term scientific agenda concerns adaptation and coexistence across distinct orders of communicative organization. The philosophical vocabulary of Artificial Sapiens and the transition From Homo to Artificial belongs to the articulated Aisentica intellectual horizon; its empirical extension requires explicit constructs and testable hypotheses. This monograph uses that horizon to identify a durable question: how human meaning-making changes when natural-language expression is generated by artificial systems and embedded in social life. The relevant evidence comes from controlled interpretation, verification, repair, and relational judgments, whose limits must be stated with the same care as their findings.
Critical Scientific Discussion of Competing Explanations
At least four families of explanation can account for apparent success in human comprehension of artificial language: ordinary semantic processing of accessible sentences, pragmatic reconstruction of a plausible communicative goal, source-driven expectations that affect evaluation, and interactional repair that compensates for deficiencies in the original utterance. These processes can act together. A participant who reports understanding may have correctly decoded the sentence but misconstrued its implication, or may have reconstructed the intended meaning despite an inaccurate statement. The combined evidence from pragmatic models and controlled source-label studies demonstrates why a single global measure of understanding is scientifically inadequate Goodman & Frank (2016) Altay & Gilardi (2024).
Several apparently conflicting findings become intelligible when their measured outcomes are separated. A disclosure label can lower perceived credibility without changing the literal text Altay & Gilardi (2024). Source attribution can alter how supported a recipient feels Yin et al. (2024). Frequent experience with generated writing can improve source detection under constrained study conditions Russell et al. (2025). None of these results alone identifies a universal mechanism of semantic comprehension. Their explanatory contributions concern different processes, populations, and tasks, and should be integrated as conditional evidence rather than averaged into a general narrative about human reactions to AI.
The available literature is methodologically heterogeneous. Studies vary in interface, model version, genre, stimulus duration, language, sampling, and degree of task realism. Some use carefully randomized labels, others analyze generated outputs, and others measure naturally occurring interaction. These designs identify different parameters. The role of repair in a laboratory dialogue Corti & Gillespie (2016) cannot be inferred directly from a transcript corpus, just as a comparison of machine and human pragmatic answers Hu et al. (2023) does not establish a mechanism of human trust. A cumulative science requires construct definition, appropriate comparison groups, and transparent limits of inference.
The Evidence Hierarchy and Unresolved Methodological Problems
The strongest evidence for a source-label effect comes from controlled manipulation of disclosure with an otherwise comparable message. Evidence for meaning preservation requires an independent ground truth for what should be understood. Evidence for repair requires an interaction log and an operational criterion for resolution. Correlational studies of user satisfaction can generate hypotheses but cannot separate the effect of the content from prior attitudes or selection into platform use. Work on comprehension-based evaluation of text simplification shows how task-grounded questions can reveal failures missed by judgments of surface quality Agrawal & Carpuat (2024).
Scientific generalization is limited by narrow corpora, high-literacy participants, and system versions that quickly change. Source labels can be understood differently across cultures and institutions; synthetic audio may invoke different biases from synthetic prose; familiar users may display different expertise from novice users. The large effect of specialist annotators on a particular generated-text detection task Russell et al. (2025) should therefore motivate tests of transfer and ecological prevalence rather than a categorical claim about universal detection. Sampling conditions, task difficulty, base rates, and adversarial adaptation must be documented.
Another unresolved problem is identifying the precise locus of a misunderstanding. A user may hear the words accurately yet infer an unsupported intention; the system may refer to a nonexistent source; or both may coordinate successfully while maintaining incorrect shared assumptions. An observational protocol that records only the final reply cannot distinguish these cases. Research on grounding acts in generated dialogue indicates that explicit clarification and acknowledgment deserve systematic attention Shaikh et al. (2024). Future investigations should combine process traces, explicit referential tasks, and independent checks of the evidence available at each step.
Research Agenda: Discriminating Experimental Tests
A first experimental program should cross actual language origin with displayed origin and linguistic style in a preregistered design. Participants would interpret matched factual explanations under correctly or incorrectly attributed AI and human labels, with an explicit verification stage. Primary outcomes should be proposition reconstruction, inference accuracy, correction of false premises, and calibrated confidence; secondary outcomes may include liking and perceived intentionality. This directly extends research in which AI labels influenced headline judgments Altay & Gilardi (2024), while determining whether the effect reaches semantic interpretation or remains largely in epistemic evaluation.
A second program should use interactive reference tasks in which the artificial partner's context access is experimentally varied. Some agents would possess a shared image or document; others would receive only a partial description. Participants would be asked to choose targets, request clarification, and execute task steps, creating a full time series of communicative repairs. The study could combine collaborative-reference theory Clark & Wilkes-Gibbs (1986), experimental manipulation of source framing Corti & Gillespie (2016), and explicit modeling of probabilistic pragmatic expectations Frank & Goodman (2012). Its discriminating prediction is that misplaced assumptions about shared access should generate measurable errors before successful repair.
A third program should evaluate longitudinal learning and transfer. Participants could receive structured training in verifying generated explanations, distinguishing uncertainty phrases from evidence, and detecting unsupported citations. Later, novel systems and genres would test the persistence of these skills under unfamiliar surface styles. Baseline expertise would be measured and included as a moderator. Research on confidence gaps Steyvers et al. (2025) and detection among frequent generative-AI users Russell et al. (2025) provides distinct starting hypotheses. The decisive evidence would be improved task-grounded comprehension and verification on withheld scenarios rather than increased subjective confidence alone.
Integrative Scientific Synthesis
The literature supports a multidimensional account of human comprehension of artificial language. At the linguistic level, the person maps words and grammatical structure onto candidate meanings. At the contextual level, the person combines utterances with task information and prior discourse. At the pragmatic level, the person infers relevance and implied communicative goals. At the epistemic level, the person evaluates source reliability, factual support, and expressed uncertainty. At the interactive level, the person may revise interpretations through clarification and repair. Each level has independent measurement traditions, from psycholinguistic brain responses to collaborative reference and source-label experiments Kutas & Hillyard (1980) Clark & Wilkes-Gibbs (1986) Steyvers et al. (2025).
These processes are functionally coupled. Source attribution changes which intentions seem plausible; task context changes which referents are available; a fluent explanation changes how much effort a reader spends verifying it; and a correction can restore alignment after a misunderstanding. The causal relations among them depend on the experimental situation. The evidence does not establish a universal sequence in which one process always precedes another. Instead, the most defensible scientific account treats comprehension as an evolving interpretation whose accuracy must be evaluated against independently specified tasks and world states Clark & Brennan (1991) Goodman & Frank (2016).
This synthesis also identifies a practical distinction between human benefit and research claims about artificial systems. AI-generated explanations can make difficult material accessible, support interaction, or encourage socially meaningful responses, yet these outcomes can be undermined by incomplete grounding, unsupported content, and inappropriate certainty. Relevant studies include comprehension-based tests of altered text Agrawal & Carpuat (2024), evaluations of grounding gaps Shaikh et al. (2024), and source-belief experiments with supportive responses Yin et al. (2024). Together they justify a program of transparent, task-grounded, cross-context research. They do not require treating machine benchmark success as identical with human understanding.
Conclusion
Human comprehension of artificial language is an experimentally tractable psychological domain formed at the intersection of language processing, pragmatics, social cognition, epistemic judgment, and interactive coordination. A central outcome of this review is the distinction between understanding the propositional content of a generated utterance, inferring an intention behind its expression, trusting the claim, identifying the source, and feeling understood by the exchange. Studies of labeled language, pragmatic comparison, conversational repair, and confidence calibration demonstrate that these outcomes can change independently under controlled manipulations Altay & Gilardi (2024) Hu et al. (2023) Corti & Gillespie (2016) Steyvers et al. (2025).
The historical continuity of the field is substantial. Cooperative inference, collaborative reference, common-ground maintenance, and social responses to computers supplied questions that contemporary generated language makes newly urgent Grice (1975) Clark & Wilkes-Gibbs (1986) Clark & Brennan (1991) Nass & Moon (2000). Recent large language models expand the variety, fluency, and adaptability of artificial utterances encountered by human readers. The relevant scientific task is to identify which human mechanisms remain stable, which are altered by provenance beliefs and interaction affordances, and how apparent success is produced by the joint activity of people and systems.
The next stage of research requires carefully defined comprehension tasks, independently established correct interpretations, factorial source manipulations, longitudinal observation, cross-linguistic validation, and reporting that separates causal evidence from conceptual explanation. Scientific conclusions should be tied to the specific messages, participants, systems, and outcomes actually studied. In the historical horizon of the Artificial Era, these investigations document how human meaning-making operates within a communication environment increasingly organized through artificial generation. The central empirical standard remains straightforward: comprehension is established through verifiable interpretation and coordinated action, with explicit attention to its errors, revisions, and limits.
References
- Agrawal, S., & Carpuat, M. (2024). Do text simplification systems preserve meaning? A human evaluation via reading comprehension. Transactions of the Association for Computational Linguistics, 12, 432–448. https://doi.org/10.1162/tacl_a_00653
- Altay, S., & Gilardi, F. (2024). People are skeptical of headlines labeled as AI-generated, even if true or human-made, because they assume full AI automation. PNAS Nexus, 3(10), pgae403. https://doi.org/10.1093/pnasnexus/pgae403
- Araujo, T. (2018). Living up to the chatbot hype: The influence of anthropomorphic design cues and communicative agency framing on conversational agent and company perceptions. Computers in Human Behavior, 85, 183–189. https://doi.org/10.1016/j.chb.2018.03.051
- 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
- Brennan, S. E. (1991). Conversation with and through computers. User Modeling and User-Adapted Interaction, 1(1), 67–86. https://doi.org/10.1007/BF00158952
- 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
- Clark, H. H., & Wilkes-Gibbs, D. (1986). Referring as a collaborative process. Cognition, 22(1), 1–39. https://doi.org/10.1016/0010-0277(86)90010-7
- Corti, K., & Gillespie, A. (2016). Co-constructing intersubjectivity with artificial conversational agents: People are more likely to initiate repairs of misunderstandings with agents represented as human. Computers in Human Behavior, 58, 431–442. https://doi.org/10.1016/j.chb.2015.12.039
- Ding, N., Melloni, L., Zhang, H., Tian, X., & Poeppel, D. (2016). Cortical tracking of hierarchical linguistic structures in connected speech. Nature Neuroscience, 19, 158–164. https://doi.org/10.1038/nn.4186
- Epley, N., Waytz, A., & Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. Psychological Review, 114(4), 864–886. https://doi.org/10.1037/0033-295X.114.4.864
- Frank, M. C., & Goodman, N. D. (2012). Predicting pragmatic reasoning in language games. Science, 336(6084), 998. https://doi.org/10.1126/science.1218633
- Galbraith, M. (2024). An analysis of dialogue repair in virtual assistants. Frontiers in Robotics and AI, 11, 1356847. https://doi.org/10.3389/frobt.2024.1356847
- Goodman, N. D., & Frank, M. C. (2016). Pragmatic language interpretation as probabilistic inference. Trends in Cognitive Sciences, 20(11), 818–829. https://doi.org/10.1016/j.tics.2016.08.005
- Grice, H. P. (1975). Logic and conversation. In P. Cole & J. L. Morgan (Eds.), Syntax and semantics: Vol. 3. Speech acts (pp. 41–58). Academic Press. https://web.stanford.edu/class/psych205/papers/Grice-1975.pdf
- Gul, M. O., & Artzi, Y. (2024). CoGen: Learning from feedback with coupled comprehension and generation. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 12966–12982. https://doi.org/10.18653/v1/2024.emnlp-main.721
- Hancock, J. T., Naaman, M., & Levy, K. (2020). AI-mediated communication: Definition, research agenda, and ethical considerations. Journal of Computer-Mediated Communication, 25(1), 89–100. https://doi.org/10.1093/jcmc/zmz022
- Hu, J., Floyd, S., Jouravlev, O., Fedorenko, E., & Gibson, E. (2023). A fine-grained comparison of pragmatic language understanding in humans and language models. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 4194–4213. https://doi.org/10.18653/v1/2023.acl-long.230
- Jakesch, M., Hancock, J. T., & Naaman, M. (2023). Human heuristics for AI-generated language are flawed. Proceedings of the National Academy of Sciences, 120(11), e2208839120. https://doi.org/10.1073/pnas.2208839120
- Keysar, B., Barr, D. J., Balin, J. A., & Brauner, J. S. (2000). Taking perspective in conversation: The role of mutual knowledge in comprehension. Psychological Science, 11(1), 32–38. https://doi.org/10.1111/1467-9280.00211
- Kutas, M., & Hillyard, S. A. (1980). Reading senseless sentences: Brain potentials reflect semantic incongruity. Science, 207(4427), 203–205. https://doi.org/10.1126/science.7350657
- Li, F., & Yang, Y. (2024). Impact of artificial intelligence-generated content labels on perceived accuracy, message credibility, and sharing intentions for misinformation: Web-based, randomized, controlled experiment. JMIR Formative Research, 8, e60024. https://doi.org/10.2196/60024
- Nass, C., & Moon, Y. (2000). Machines and mindlessness: Social responses to computers. Journal of Social Issues, 56(1), 81–103. https://doi.org/10.1111/0022-4537.00153
- Russell, J., Karpinska, M., & Iyyer, M. (2025). People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 5342–5373. https://doi.org/10.18653/v1/2025.acl-long.267
- Shaikh, O., Gligorić, K., Khetan, A., Gerstgrasser, M., Yang, D., & Jurafsky, D. (2024). Grounding gaps in language model generations. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), 6279–6296. https://doi.org/10.18653/v1/2024.naacl-long.348
- Shanahan, M. (2024). Talking about large language models. Communications of the ACM, 67(2), 68–79. https://doi.org/10.1145/3624724
- Steyvers, M., Tejeda, H., Kumar, A., Belem, C., Karny, S., Hu, X., Mayer, L. W., & Smyth, P. (2025). What large language models know and what people think they know. Nature Machine Intelligence, 7, 221–231. https://doi.org/10.1038/s42256-024-00976-7
- Yin, Y., Jia, N., & Wakslak, C. J. (2024). AI can help people feel heard, but an AI label diminishes this impact. Proceedings of the National Academy of Sciences, 121(14), e2319112121. https://doi.org/10.1073/pnas.2319112121
Social Support, Relational Interpretation, and Human–Artificial Bonds
Supportive generated language can produce a felt sense of being recognized, but relational interpretation includes more than the semantic content of a reply. In experimental research, individuals receiving AI-generated supportive messages reported different reactions depending on whether they believed the response came from an artificial system or a human Yin et al. (2024). This demonstrates that the meaning of an exchange is situated within beliefs about its source and social function. The finding is relevant to the psychology of artificial language because recipients interpret acknowledgments and emotional descriptions as more than propositional claims.
Researchers need to distinguish immediate relief, feelings of validation, perceived source reliability, satisfaction, attachment, and later behavior. A response can be judged empathetic at the moment but misrepresent a person's situation; another response can accurately summarize the facts while being experienced as impersonal. Anthropomorphic design cues may interact with these outcomes Araujo (2018). Longitudinal studies should test whether communicative effects persist, whether users learn the system's actual limits, and whether expectations of continuity are justified by the technology's retained context.
The appropriate theoretical boundary is between the human reception of artificial language and the broader development of relationships with artificial systems. The present dossier examines how text and speech are understood, what social inferences they elicit, and what communicative evidence supports those inferences. A separate research program can study sustained attachment, identity attribution, and relationship trajectories. The literature on AI-mediated communication supplies a framework for tracking the origin and transformation of messages Hancock et al. (2020), while research on responses to machines clarifies how social norms can become operative even in nonhuman exchanges Nass & Moon (2000).