Reimagining Human-AI Communication:
A Hybrid Model for Large Language Model Interactions

1.2.1 Development and application of LLM

The launch of ChatGPT (2022) marked a communication shift comparable to the internet’s rise, reaching 100M users in 2 months (Hu, 2023). Adoption in business and daily life is accelerating (Singla et al., 2025).

In 2024, reports revealed that the performance of LLMs are continuing to be challenging competitively, that the United States led with 40 notable AI models, followed by China with 15 and France with three, while the U.S. leads in producing top AI models—but China is closing the performance gap (Maslej et al., 2025). Generative AI saw particularly strong momentum, attracting $33.9 billion globally in private investment—an 18.7% increase from 2023. AI business usage is also accelerating: 78% of organizations reported using AI in 2024, up from 55% the year before (Maslej et al., 2025).

Today, AI is increasingly embedded in everyday life, and researchers are already observing how LLMs are being applied across industries. Like LLMs’ integration to real-world financial services (Maple et al., 2024), to support academic research (AlZaabi et al., 2023), in government affairs (Wang et al., 2024b).

1.2.2 Concerns around LLMs

Behind this technological revolution lies a profound philosophical tension. The “stochastic parrot” critique, articulated by Bender, Gebru, McMillan-Major, and Mitchell (2021), characterizes LLMs as sophisticated mimicry systems that “haphazardly stitch together sequences of linguistic forms... without any reference to meaning” (Bender et al., 2021). This metaphor suggests fundamental deficiencies in understanding despite convincing outputs. However, users increasingly perceive communicative agency in these systems—67% of 300 U.S. residents attributed some degree of phenomenal consciousness to ChatGPT, with frequent users more likely to perceive consciousness (Colombatto and Fleming, 2024). This anthropomorphic tendency intensifies with exposure, suggesting prolonged interaction makes users perceive LLMs as legitimate communicative agents rather than mere tools.

LLMs’ sophisticated linguistic combination, based on transformer models predicting next tokens from vast datasets (Achiam et al., 2023), amplifies existing communication problems. Disinformation created by LLMs proves more challenging to identify (Chen and Shu, 2023), potentially threatening fundamental rights to truth. These systems also exhibit limitations distinguishing them from human interlocutors: difficulties with long-term memory, consistency, emotional understanding, and reliable cognitive pattern simulation (Wang et al., 2025). This creates a novel communication environment where users engage with entities they simultaneously know to be algorithmic yet experience as agentic—neither fully tool nor fully agent, challenging fundamental assumptions about communication itself.

a. To what extent can traditional communication theory explain the human-LLM interactions?

This question critically examines the applicability of established communication theories to LLM contexts. It investigates which components of models like Shannon-Weaver, Berlo’s SMCR, and transactional frameworks remain relevant and which encounter limited applicability. By systematically analysing where these models face limited reasonability-particularly regarding concepts of agency, intentionality, and meaning construction-this research identifies specific theoretical gaps that must be addressed. For example, how do we understand “noise” in Shannon-Weaver when applied to token prediction errors versus human misinterpretation? How does Berlo's emphasis on the “source’s” knowledge, attitudes, and culture translate when the source includes both human developers and massive training datasets?

b. What hybrid components best model human-LLM communication dynamics?

Building on identified gaps, this question explores novel theoretical constructs that more accurately capture human-LLM interactions. It examines how elements from computer science (e.g., attention mechanisms, statistical inference), linguistics (e.g., pragmatics, discourse analysis), and media studies (e.g., parasocial interaction theory) might be synthesized into a more comprehensive framework. This investigation considers various interaction contexts-from educational to customer supporting to creative-to ensure the model's robustness across diverse applications. Key considerations include how to conceptualize distributed agency, how to map feedback loops across multiple timescales, and how to account for the evolving nature of these systems through ongoing learning and updates.

c. What benefits or risks emerge from treating LLMs as communicative agents?

This question addresses the normative implications of communication framing, examining how different conceptual models influence user behaviour and system design. It investigates potential consequences of anthropomorphism, including over-trust, emotional attachment, and distorted expectations. Does framing LLMs as “agents” increase user trust, dependency, or compliance in high-stakes contexts (e.g., healthcare, finance)? By analysing cases where communication framing has led to problematic outcomes-such as users developing inappropriate dependencies or making high-stakes decisions based on LLM outputs-this research identifies specific ethical concerns requiring mitigation. It further explores how communication theories might inform design practices and regulatory approaches that promote responsible engagement with these technologies.

These questions progress from descriptive analysis to theoretical construction to ethical application, collectively addressing both the new challenges of understanding a novel communication paradigm and the practical challenge of guiding its development in socially beneficial directions.

2.1.1 Linear Transmission Models

The Shannon–Weaver “mathematical theory of communication” conceptualises messages as signals that travel unidirectionally from source to destination, degraded only by channel noise (Shannon, 1948). In LLM interaction, some elements map cleanly onto this framework: user prompts function as “information sources”, model output travels through an API or graphic user interface (GUI) “channel”, and token-level corruption (e.g., truncation) is a quantifiable form of noise. However, two fundamental mismatches arise.

First, Shannon–Weaver treats meaning as irrelevant to transmission efficiency, whereas LLM users mostly evaluate responses semantically: hallucinations that are syntactically perfect but factually wrong illustrate this gap (Bender et al., 2021). Second, the model presupposes fixed, identifiable senders and receivers. An LLM reply emerges from the joint influence of user prompts, system prompts, training data, and provider mediating, distributing agency rather than locating it in a single source (Neumann et al., 2025). Consequently, while the linear model aids measurement of latency, throughput, and token error, it cannot account for iterative prompt refinement or co-constructed intent.

2.1.2 Interactive Communication Models

Schramm’s circular model adds feedback loops, casting communication as iterative encoding-decoding between partners (Schram, 1954). Such reciprocity is visible when users chain prompts to refine tone or request citations. Interactive framing also highlights how LLMs alter their outputs after receiving critique, a behaviour sometimes described as “self-correction” (Kumar et al., 2024). Nevertheless, Schramm still presumes symmetrical cognitive capabilities, whereas LLMs fundamentally differ from humans in their approach to understanding and cognition (Cuskley et al., 2024). The theory also underestimates stochastic variation: identical prompts can yield divergent completions despite identical “channel” conditions (Atil et al., 2024).

2.1.3 Transactional Communication Models

Barnlund’s transactional model treats communication as a simultaneous, co-constitutive process embedded in context, emphasising that interlocutors are always both senders and receivers (Barnlund, 2017). This stance aligns well with prompt engineering, where user and model mutually shape discourse; each new token from the model reshapes the “field of experience” that guides the next user turn. Moreover, Barnlund foregrounds environmental and psychological “noise”, paralleling socio-technical factors such as content filters or political bias in training data.

Limitations persist. Barnlund assumes human communicators with shared perceptual channels, yet an LLM lacks sensory grounding and instead operates on textual probabilities. Consequently, while transactionalism captures reciprocity and context-dependence, it cannot by itself explain phenomena such as out-of-distribution hallucination (LLMs produce incorrect, fabricated, or spurious outputs when inputs beyond training data) or long-term memory loss (LLM’s deterioration or failure to retain previously learned or stored information over extended periods or interactions).

2.2.1 Evolution of HMC Frameworks

HMC scholarship reframes machines as message sources rather than mere channels (Guzman, 2018, Lewis et al., 2019). Guzman & Lewis argue that communicative AI requires new ontological categories because intentionality is now attributed to artefacts (Guzman and Lewis, 2020). Empirical studies confirm that users adapt human-like politeness strategies after only one turn of ChatGPT interaction, indicating mental-model shifts (Schneider et al., 2024). These findings echo HMC's proposition that anthropomorphism is relational and situational rather than device-intrinsic.

2.2.2 Anthropomorphism in Human–AI Interaction

Attribution of consciousness to LLMs rises with usage frequency; 67% of U.S. participants ascribed subjective experience to ChatGPT in 2024 (Colombatto and Fleming, 2024). Trust similarly correlates with perceived humanness. Studies show a positive correlation between consciousness attributions and advice-taking behaviour (Colombatto et al., 2025). Yet it warns of a "machine penalty": users cooperate less when they recognise non-human agency in high-stakes tasks (Wang et al., 2024c). These contradictory patterns illustrate the dual-edge of anthropomorphic design—facilitating relational engagement while risking over-trust and moral confusion.

2.2.3 Intersubjective Models of AI-Mediated Communication

Recent proposals extend HMC by theorising mutual theory-of-mind between user and agent (Wang and Goel, 2022). Intersubjective frameworks treat AI not just as conversational partner but as mediator capable of reshaping human-human interaction via translation, style transfer, or bias filtering (Aoyama et al., 2025). Such models accommodate distributed agency and align well with LLM use cases in customer service or collaborative writing, yet they rely on assumptions of stable long-term memory that current LLM architectures do not meet.

2.3.1 Transformer Mechanisms and Attention Systems

The self-attention mechanism enables LLMs to model token interdependencies (Vaswani et al., 2017), approximating pragmatic inference when prompted with few-shot examples (Brown et al., 2020). However, attention operates within a finite context window; overflow truncates early turns, breaking coherence (Bergmann, 2024). Memory-augmented proposals—SCM (Wang et al., 2023a), LongMem (Wang et al., 2023b), MemoryBank (Zhong et al., 2024)—seek to externalise long-term context, that is summarise the former dialogue for replacing existing context, so token could be simplified. But evaluation datasets such as LoCoMo reveal persistent degradation beyond ~9 k tokens (Maharana et al., 2024).

2.3.2 Training Processes and Communicative Implications

Reinforcement learning from human feedback (RLHF) optimises models for perceived helpfulness, thereby embedding dominant cultural patterns (Yin et al., 2024) and potentially amplifying anthropomorphic misperceptions. Hallucination remains a persistent issue. While RLHF has reduced hallucination tendencies, it has not eliminated them entirely (Dahlgren Lindström et al., 2025), beyond the matter of how much computational resources equipped but algorithms (Xu et al., 2024), a constraint that classical communication models do not foresee. Multi-agent debate and voting frameworks reduce errors (Chan et al., 2023), yet introduce meta-conversation layers that further complicate sender–receiver identification.

2.3.3 Specialised Applications and Communicative Contexts

Domain-specific deployments—medical triage (Sorich et al., 2024), public-health voice agents (Wen et al., 2025), government policy drafting (2025)—illustrate both promise and peril. Personalised assistants benefit from vector-database memory but raise privacy risks when recalled out of context (Liu et al., 2024b). Technical architectures thus directly shape communicative affordances, underscoring the need for interdisciplinary models that integrate system design with communication-centric theory (Yan et al., 2025).

2.4.1 Recognition of Social Impact Beyond Technical Metrics

Leading AI researchers increasingly acknowledge that LLM impacts extend far beyond computational performance metrics. Critiques often focus on performance or safety evaluations, yet the deeper societal consequences—effects on social groups, institutions, norms, and practices—demand broader investigation (Wang et al., 2024a). This has motivated calls for more comprehensive evaluation tools capturing complex social dynamics. Initiatives like the Epitome platform, focused on "deep integration of artificial intelligence and social science," explicitly seek theoretical foundations from fields like communication studies to investigate "interactive impacts of AI on individuals, organizations, and society" during real-world deployment (Qu et al., 2025). This reflects a direct response to the need for frameworks that address the socio-communicative realities driving regulatory concerns.

2.4.2 The Limitations of Technical-Only Approaches

LLM developers now recognize the inadequacy of purely computational approaches for understanding communicative and social functions. Research highlights challenges like choosing models without social science guidance and the significant impact of model alignment on outputs in ways requiring humanistic interpretation (Lyman et al., 2025). Attempts to use LLMs to simulate human participants reveal fundamental epistemological limitations, such as the "correct answer effect," demonstrating that technical systems cannot capture human diversity of thought (Colombatto and Fleming, 2024, Wang et al., 2025, Park et al., 2024). These limitations underscore why purely technical metrics cannot resolve the communication-theoretic and regulatory dilemmas identified earlier.

2.4.3 Calls for Human-Centered AI Research

The rise of Human-Centered AI (HCAI) represents a direct acknowledgment within computer science that technical excellence is insufficient. However, research mapping HCAI reveals "considerable ambiguity" about its framing, design, and evaluation, highlighting the urgent need for expertise from fields like communication studies (Capel and Brereton, 2023). As researchers note, human-centric measures (fairness, trust, explainability) are subjective, context-dependent, and often uncorrelated with conventional performance metrics (Ghai, 2023). Therefore, addressing the core challenges of human-LLM interaction, including those with regulatory dimensions, necessitates interdisciplinary collaboration that incorporates communication studies' frameworks for understanding subjective, contextual, and relational phenomena.

3.1.1 Transformer Architecture and Its Communicative Implications

3.1.1.1 From Sequential to Parallel Processing

The Transformer marked a paradigm shift by abandoning sequential text processing in favour of processing entire sequences in parallel (Vaswani et al., 2017). Stacks of multi-head attention and feed-forward layers create a unified framework for tasks from generation to translation and summarization. The modular design of Transformer networks—scaling up layers, width, or heads—enables LLMs to smoothly expand with growing data, and attend to every token in an input simultaneously. This property allows them to generate highly coherent, context-aware responses and makes LLMs more robust and effective as model and corpus size increase (Kaplan et al., 2020).

3.1.1.2 Self-Attention: Dynamic Focus for Meaning

Self-attention lets the model highlight relevant parts of a message—whether a proximate phrase or a distant clause—mirroring human selective listening (Vaswani et al., 2017). Multiple attention heads (e.g., grammar, semantics, discourse) track syntax, topic shifts, and discourse cues simultaneously (Clark et al., 2019, Zhang et al., 2025). Communication researchers can view self-attention as an algorithmic analogue to contextual inference, grounding analyses of how LLMs “understand” and respond to varied communicative intents. This algorithmic contextual inference mechanism provides the first theoretical bridge to overcome Schramm’s limitation of presuming symmetrical cognition (Section 2.1.2).

3.1.1.3 Context Window: Managing Conversational Scope

Each Transformer operates within a fixed context window, allowing every token to attend to all others via self-attention and thereby maintain coherence over long passages (Bergmann, 2024). Early models handled a few thousand tokens (Wu et al., 2024); leading LLMs—such as DeepSeek V3/R1 (Liu et al., 2024a), Llama 3 (Dubey et al., 2024), and Qwen 3 (Yang et al., 2025)—now manage around 128,000 tokens, while top-tier models like Gemini 1.5 Pro extend this to 1,000,000 tokens (Team et al., 2024). Within this scope, LLMs can recall earlier points and cross-reference distant facts; and when inputs exceed this limit, content must be truncated or summarized, a process that mirrors human memory decay over extended discourse. For communication studies, these limits are crucial for modelling how LLMs plan discourse and manage conversational continuity.

These memory constraints necessitate new transactional communication models that account for non-human forgetting—resolving Barnlund’s oversight identified in Section 2.1.3.

3.1.2 Pre-training: Learning Language Patterns – the base model

The foundation of any LLM begins with pre-training, a self-supervised learning phase—distinguished from traditional unsupervised methods by its use of automatically generated prediction targets (LeCun and Misra, 2021)—during which models learn general language patterns from massive text datasets (Bergmann, 2023, Vaswani et al., 2017). During this stage, the model is exposed to diverse sources including web text, books, articles, and code repositories, learning to predict missing or next words in sequences without requiring labelled data (Bergmann, 2023). This process employs the fundamental objective of causal language modelling, where the model learns the statistical relationships between tokens in natural language - The model learns to approximate probability distributions over sequences of tokens, developing an understanding of grammar, facts, reasoning patterns, and semantic relationships through exposure to billions of text examples (Radford et al., 2019). This phase creates what researchers term the “base model”—essentially a sophisticated document generator capable of predicting the next token in any given sequence (Radford et al., 2018).

3.1.3 Fine-tuning: Task Specialization

Following pre-training, models undergo fine-tuning to adapt them for specific tasks or behaviours. This process involves training on smaller, curated datasets that demonstrate desired input-output patterns (Ouyang et al., 2022). Instruction tuning, a common form of fine-tuning, trains models to follow instructions by providing examples of human-written prompts paired with appropriate responses.

The fine-tuning process typically involves three distinct approaches: Supervised Fine-tuning (SFT) for a more directed response pattern suitable for interactive applications (Ouyang et al., 2022), Parameter-Efficient Fine-tuning (PEFT) for specific tasks (Hu et al., 2022) and Safety Fine-tuning addresses potential harmful outputs by training models on safety-focused datasets (Bengio et al., 2025) which remains an active area of research with ongoing challenges. Academic research demonstrates that even small instruction-tuned models (1.3B parameters) can outperform much larger base models (175B parameters) on human preference evaluations (Ouyang et al., 2022).

A study on multi-task fine-tuning shows that mathematical reasoning and code generation capabilities improve significantly with proper instruction datasets (Dong et al., 2023). Further research also found that, globally propagated noise – the noise would corrupt not only one but all downstream steps during CoT - during both pre-training and fine-tuning, confirming the need for aggressive filtering of “global” errors in large datasets (Havrilla and Iyer, 2024).

Accordingly, instruction tuning effectively filters out that noise (in base model) by focusing on high-quality, human-curated examples, allowing even smaller models to surpass much larger but noisier base models on human preference evaluations.

3.1.4 Reinforcement Learning from Human Feedback (RLHF)

The final training phase employs RLHF to align model outputs with human preferences. This process involves human evaluators rating model responses on criteria such as helpfulness, harmlessness, and honesty (Ouyang et al., 2022). These ratings train a reward model that approximates human preferences, which then guides further model optimization through reinforcement learning algorithms (Gao et al., 2024).

In communication-centric terms, fine-tuning and RLHF serve distinct but complementary roles in shaping an LLM’s outputs. Fine-tuning resembles a scripted classroom session (Pareja et al., 2024), ingesting explicit prompt–response pairs, whereas RLHF functions like a workshop critique, using ranked human judgments to steer the model toward audience-preferred outputs (Kaufmann et al., 2024), while RLHF’s reward model is still a statistical proxy, not a genuine evaluator.

While the Project PRISM (Kirk et al., 2024) is a landmark for its participatory scale and the granularity of its demographic linkage, it should be noted that the research—perhaps for the first time so explicitly—foregrounds social science, ethics, and science and technology studies frameworks not just in rhetoric but in its actual methodology and analysis. This move vividly illustrates that LLM development is not merely a technical challenge. The humanities offer critical perspectives on context, subjectivity, value diversity, and the social consequences of AI systems—perspectives that cannot simply be retrofitted after the fact by technical teams alone. It is important, thereby, to situate this contribution.

In Project PRISM, actual humanities participation remains indirect: the “social science” and “community” involvement consists primarily of carefully constructed and demographically-balanced rater populations, rather than direct scholarly or stakeholder representation in framing what counts as alignment, which prompts are posed, or which outcomes are prioritized (Kirk et al., 2024). In this sense, PRISM builds on a substantial tradition in interdisciplinary AI research, but still treats social context as a metric to optimize, not as a true partner in decision-making.

As Project PRISM states, “Alignment cannot be neatly bifurcated into technical and normative components. PRISM assists in navigating these complexities with more human voices adjudicating alignment norms” (Kirk et al., 2024). Accordingly, if alignment is to be just, inclusive, and safe, humanities scholars and community stakeholders must be included from the very beginning—informing dataset design, value articulation, feedback protocols, and impact evaluation—neither simply at the stage of raters or ex post critique, nor after trained fix up, the “red-team” exercises (intentionally probing systems with challenging, adversarial, or unexpected prompts to uncover biases, vulnerabilities, and failure modes) or participatory model cards (simply declare known factors).

3.2.1 Provider Mediation: Hidden Levers that Shape Outputs

LLM service providers construct a sophisticated “mediating layer” between users and models through their deployment strategies, business models, and technical interventions. This intermediary infrastructure, while often invisible to end users, profoundly influences information accessibility, introduces content bias, and constitutes a new form of agenda setting power that warrants serious examination within communication theory frameworks.

3.2.1.1 Barriers to Entry: Re-examining Accessibility, Cost, and “Democratization”

The most common cost structures for accessing LLMs vary significantly between GUI (for customers) and API (for developers) access methods, as well as between closed-source and open-source models.

Closed-source models impose direct fees on end users and developers. GUI access often includes a free tier plus paid subscriptions—e.g., ChatGPT Plus (£20/month) and ChatGPT Pro (£200/month) (OpenAI, 2025a) or Google AI Pro (£19–20) and AI Ultra (£235) (Google, 2025b) , —with higher-tier plans offering faster response times and larger usage quotas. API access shifts to pay-per-token pricing, which can vary widely: for instance, batch versus priority requests on ChatGPT 5 and lower-capacity variants like ChatGPT 5 nano or competitors like Gemini 2.5 Pro (OpenAI, 2025b, Google, 2025a) differ substantially in cost (Chart 1). Anthropic estimates an average conversation uses about 3,700 tokens (Anthropic, 2025).

Chart 1: Model Token Cost Compare

Open-source models remove licensing fees but transfer costs to infrastructure. Deploying a 70B-parameter model such as Llama 3 for 5,000 users over four years can cost nearly $1 million on self-owned servers, $2.1 million on AWS, or $2.9 million via GPT-4o API (Kaufmann, 2025). As a result, pay-as-you-go APIs often remain cheaper for low-volume or short-term use, while self-hosting becomes cost-effective only at large scale, or when privacy or customization needs justify higher upfront investment.

DeepSeek's architecture demonstrates remarkable efficiency, achieving competitive performance while requiring only 37 billion active parameters from its 671 billion total parameter base, resulting in significantly lower computational costs (Liu et al., 2024a). Meta's Llama 3.1 405B represents what Meta AI claims as “the world’s largest and most capable openly available foundation model”, providing capabilities that rival closed-source alternatives while maintaining open accessibility (2024a).

Besides the cost itself, these open-source alternatives also directly address Knowledge Gap Theory concerns by providing information processing media accessible to users regardless of economic circumstances, that is the proprietary nature of the closed-source LLMs. Tichenor, Donohue, and Olien’s original formulation suggested that information acquisition capabilities differ based on socioeconomic status (Tichenor et al., 1970), which should be extend to the control of its proprietary information processing media – the LLM today, beyond the cost concerns. Open-source LLMs potentially disrupt this pattern by enabling equal access to advanced language processing capabilities, though infrastructure requirements and technical literacy demands remain as secondary barriers. However, the dependency of LLM development - commercial sustainability should be respect, any attempts of providing free LLM accesses should be welcomed and encouraged.

The competitive cloud services market facilitates deployment of open-source models, enabling individuals and small organizations to achieve private, cost-effective deployments. This accessibility has global implications, particularly for regions where commercial LLM services may be economically prohibitive or subject to regulatory restrictions. Research also proved that, cost-effectiveness comparisons favour open-source alternatives by margins of 95% in specific cases, fundamentally altering the economics of LLM access (Zhang et al., 2024a).

Provider normative principles introduce another layer of consideration regarding accessibility and equity. Commercial providers typically implement comprehensive data privacy protections, content neutrality policies within legal boundaries, and reliability standards that ensure consistent service availability.

3.2.1.2 Content Shaping: System Prompts, RAG, and Agenda Setting

When consumers’ direct access through GUI, LLM Service Providers actively shape LLM outputs through technical interventions, most for compliance or performance purposes (Gorwa et al., 2020). However, those interventions should be considered as modern extend of gatekeeping and agenda setting, distinguished from the used pattern of media agency’s responsibility, but spread in each dialogue, operating largely outside user awareness yet significantly influencing the information and perspectives users encounter. These interventions represent a sophisticated form of content curation that extends beyond traditional media gatekeeping into real-time, personalized information filtering and framing.

System prompts operate as foundational “super-instructions” that establish LLM behaviour parameters before any user interaction occurs. These hidden instructions define task guidelines, model persona, response tone, prohibited topics, ethical boundaries, and value orientations (Zhang et al., 2024b, Saggu and Das, 2025). Unlike traditional media content where editorial decisions are visible through bylines, publication mastheads, or explicit opinion labelling, system prompts operate invisibly, making their influence from an agenda setting perspective. Recent incidents of system prompt leakage provide revealing insights into this hidden influence layer. Research on various LLM systems reveals prompts containing specific cultural assumptions, and value hierarchies that shape responses across all interactions (Kharchenko et al., 2024). Leaked prompts have revealed instructions directing models to prioritize specific types of sources, avoid certain topics, or frame controversial subjects in particular ways (0xAb1d, 2024). These revelations demonstrate that system prompts extend far beyond technical functionality into explicit content and perspective shaping.

The agenda setting implications become clear when viewed through McCombs and Shaw's framework (McCombs and Shaw, 1972): while traditional agenda setting theory focused on media’s role in determining “what to think about”, the first layer - issue salience (Cohen, 2015). For LLM interactions, the user’s input prompt effectively establishes issue salience in each conversation. That is, by asking a specific question, the user selects the focal topic — mirroring the media’s role in elevating certain issues in the public sphere. Simultaneously, attribute salience — the second level of agenda setting — emerges through the influence of system prompts and the model’s design. System prompts determine which aspects of the issue are emphasised, which perspectives are prioritised, and how complex topics are contextualised. While much of an LLM’s attribute salience is shaped during the training process (Section 3.1), the operational touchpoint for shaping it in a specific conversation lies in the system prompt, due to its invisible yet causal role in steering the model’s responses. Thus, in LLM–user exchanges, user prompts can be conceptualised as setting first-level agenda (issue salience), while system prompts function as a dynamic, situational mechanism for second-level agenda setting (attribute salience).

Temperature: a parameter modifies the probability distribution over possible next tokens during text generation, commonly referred as “creativity parameter”. Low temperature (<1) prefer the highest probability token and high temperature (>1) for lower-probability options (Peeperkorn et al., 2024).

Retrieval-Augmented Generation (RAG) introduces external knowledge sources that significantly influence LLM responses, yet these sources are typically selected and curated by providers rather than users. RAG implementations determine which databases, news sources, academic publications, or websites the model can access when formulating responses to factual questions (Lewis et al., 2020). This selection process inherently introduces editorial bias, as the choice of sources fundamentally shapes available information, which is one-time-off information fetched in each conversation, thus differed from issues with dataset in Section 3.1. Research reveals concerning patterns in RAG implementations where source selection can create information cocoons similar to social media filter bubbles. Studies demonstrate that RAG systems can be vulnerable to poisoning attacks where malicious actors insert biased or false information into retrieval databases, highlighting the critical importance of source curation decisions (Zhao et al., 2024). The technical complexity of RAG systems obscures these editorial choices from users, who may not realize their queries are being answered using specific, pre-selected information sources rather than comprehensive knowledge bases. Perplexity AI’s recent legal challenges illustrate the real-world implications of RAG source selection. The BBC’s cease and desist action against Perplexity highlights concerns about “illegally scrape publishers’ content” and potentially “injuring the BBC’s reputation” through inaccurate representation of source material. Research indicates that 17% of Perplexity’s responses “output fell short of BBC Editorial Guidelines around the provision of impartial and accurate news”, including factual inaccuracies and missing context (McMahon, 2025). These findings underscore how RAG implementations can distort source material in ways that may mislead users about the original content's meaning or accuracy.

A new form of agenda setting implications of the co-constructed issue salience and attribute salience extend beyond each individual interactions to broader information ecosystem effects. When millions of users receive LLM responses shaped by specific system prompts and RAG sources, these systems effectively function as agenda setting mechanisms but in macro level indirectly, a new form of influence in each private conversation may turn to public influence. Unlike traditional media where audiences can choose between different editorial perspectives, most users interact with single LLM providers and remain unaware of the editorial choices embedded in system design. Educational applications like Khanmigo demonstrate particularly significant agenda setting potential, as these systems shape not just information access but learning processes and knowledge formation for students (Lake, 2023, Academy, 2025). Research revealing bias in educational AI systems shows how these platforms can reinforce harmful stereotypes, with some models portraying Latino male students as over 1,300 times more likely to be shown as “struggling” rather than as “star” students (Shieh et al., 2024). Such biases in educational AI represent a form of agenda setting that directly influences how groups of users (students) are perceived and how they may perceive themselves and their capabilities.

Agentic Mediating Layer: An advanced form of provider mediation that extends beyond basic content filtering and RAG to include autonomous task decomposition, multi-step planning, and orchestrated tool use (Rawat et al., 2025). This layer enables LLM systems to function as agentic systems capable of: (1) decomposing complex user requests into sequential sub-tasks; (2) dynamically selecting and coordinating appropriate tools or sub-agents; (3) executing multi-step workflows with iterative refinement; and (4) providing real-time status monitoring throughout task execution (Sukharevsky et al., 2025). Unlike traditional mediation that primarily shapes content, agentic mediation fundamentally transforms the interaction paradigm from conversational to task-oriented automation. For example, with permissions of direct interaction abilities to web browser, a request of “book a flight ticket from London to Beijing at [date, time]” could be assigned in a sequence of tasks on webpage – (1) check available flights from responsible provider; (2) select suitable flight; (3) confirm the details and turn to users for final authorization of purchase. Basic text response kept outputting during the tasks carrying on for status monitor. Technical definition of Agentic Mediating Layer emphasises its properties beyond RAG, to interact with certain service providers or access permissions (Li et al., 2024). However, this definition focuses on the distinctions in technical mechanisms against traditional LLMs which may risk confusion regarding attribution and accountability. With authorized access to personal data and platform permissions to obtain information or issue commands (e.g. takeout, banking, shopping, reservations), the layer can perform consequential actions in the real world. However, informational attribution remains with supplemented information (which could interact with tools), while the layer coordinates execution under policy and user consent. In communication studies perspectives, agentic mediating layer enabled task-oriented action orchestration with observable step-by-step status, but not beyond the role of external information that supplemented by the meditating layer.

Through this analysis, we see how LLM interactions create a novel form of triple-layer agenda setting: user prompts establish what topics are discussed (first-level), while hidden system prompts and RAG selections determine how these topics are framed and contextualized (second-level), and the third layer - the way system prompts and RAG (or Agentic Mediating) shapes conversations among different users in each specific cases in macro level. This represents a significant departure from traditional media agenda setting, as the “editorial decisions” are now embedded in algorithmic processes and largely invisible to users.

3.2.1.3 Attribution of Responsibility: Legal Regulation and Provider Compliance Shaping

Current regulatory approaches demonstrate varied strategies for addressing AI responsibility attribution across jurisdictions. The European Union’s AI Act establishes a comprehensive risk-based framework that categorizes AI systems according to their potential impact, with high-risk systems requiring extensive compliance measures including conformity assessments and CE marking (2024b).

The United Kingdom (Charlesworth et al., 2023) and United States (Lancieri et al., 2025) have pursued more fragmented approaches, with various agencies and departments issuing guidance and requirements for specific AI applications rather than comprehensive legislation. This patchwork approach creates uncertainty for providers operating across jurisdictions while potentially creating compliance gaps that may not adequately address emerging risks.

China’s approach through the Interim Measures for the Management of Generative AI Services focuses specifically on public-facing AI services, requiring providers to ensure content compliance with existing regulations while implementing security assessments and algorithm filing requirements (2023). The China’s framework demonstrates explicit recognition that generative AI regulation requires balancing innovation promotion with content security concerns, leading to regulatory scope limitations that exclude research and enterprise applications from direct oversight.

Compliance filtering mechanisms implemented by providers in response to these regulatory pressures represent direct content interventions that shape user interactions in real time. These systems typically operate through multiple layers including keyword filtering, topic avoidance, content classification, and response modification (Sartor et al., 2020, Marsoof et al., 2023).

Responsibility attribution frameworks emerging from regulatory approaches create complex accountability systems where multiple parties may bear responsibility for AI system outcomes. The EU AI Act’s approach to high-risk systems requires providers to maintain detailed documentation, conduct risk assessments, and ensure human oversight, effectively creating audit trails for accountability purposes (2024b). Legal scholars note that current regulatory approaches often fail to address the distributed nature of AI system responsibility, where training data providers, model developers, deployment platforms, and end users all contribute to final system behaviour (Custers et al., 2025). This distributed responsibility creates challenges for users seeking redress when AI systems cause harm, while also creating uncertainty for providers attempting to ensure compliance across complex technical systems.

The emergence of “regulatory sandboxes” in various jurisdictions represents an attempt to balance innovation with responsibility attribution by allowing controlled testing of AI systems under relaxed regulatory requirements (OECD, 2023). These frameworks acknowledge that traditional regulatory approaches may not adequately address rapidly evolving AI capabilities while providing mechanisms for gathering evidence about real-world AI impacts.

3.2.2 User Intent and Meaning Construction: A Four-Category Input Model

Users approach LLM interactions with diverse intentions, varying levels of AI literacy, and deeply embedded cognitive biases that influence how they construct meaning from AI responses. A single prompt might include social greetings (phatic), background context (informational), emotional state (expressive), and task requests (command), such as: “Hi there! I’ve been unemployed for six months and this feels like my last chance, could you help me practice answering common interview questions?”

3.2.2.1 A Taxonomy of User Inputs: A Non-Exclusive Four-Part Framework

User inputs to LLMs represent complex, multi-intentional communications that resist simple categorization yet require systematic analysis for understanding human-computer interaction dynamics. Drawing from Speech Act Theory’s focus on illocutionary force (Austin, 1975), we can establish a framework that identifies distinct communicative intentions while recognizing that most user inputs combine multiple categories simultaneously.

Commands/Instructions constitute perhaps the most recognizable category of user input, where users explicitly request LLM execution of specific tasks. Examples include “Write a poem about autumn” or “Generate Python code to sort a list”. These speech acts carry clear performative intentions and typically elicit task-oriented responses where LLMs attempt to fulfil explicit user requests. Professional training in prompt engineering can significantly improve user expertise in crafting effective commands, with studies showing perceived expertise increases of approximately 10% following structured training programs (Bashardoust et al., 2024).

Informational Content represents a second category where users provide factual context, background data, or source material for LLM processing. These inputs function as contextual grounding rather than explicit task requests, like “I’m using [Brand] [Model] phone and how should I ……” Such inputs require LLMs to integrate provided information with subsequent queries of RAG or tasks. Users who provide excessively lengthy contextual information (e.g., detailed financial transaction records, inventory sales data) may inadvertently reduce model performance as attention mechanisms become diluted across extensive input text. Conversely, users who provide insufficient context may receive responses that lack necessary specificity or accuracy for their particular use cases.

Emotional Expressions constitute a third category encompassing personal feelings, emotional support seeking, and subjective experience sharing. Examples include “I feel overwhelmed by this assignment” or “This situation is really frustrating me”. These inputs challenge LLMs to provide appropriate empathetic responses while maintaining boundaries around their capabilities and avoiding therapeutic overreach. Studies indicate that LLMs consistently exhibit “emotional bonds” including empathy and validation when responding to emotional expressions, with these behaviours becoming more pronounced in multi-turn conversations (Zhou et al., 2020). While such responses may provide users with immediate emotional support, they also contribute to anthropomorphic perceptions that may lead to unpredictable dependency on AI systems for emotional needs.

Phatic/Social Information encompasses conversational maintenance elements including greetings, courtesy expressions, and social acknowledgments like “Hello”, “Thank you” or “How are you?”. Foundational Social Response Theory states “Individuals’ interactions with computers are fundamentally social” (Reeves and Nass, 1996). Research indicates that Users carry cultural politeness norms and expectations into AI interactions; misalignment leads to pragmatic failures and dissatisfaction, highlighting human-to-AI transfer of social conventions (Luo, 2025). This tendency reveals the persistence of human social schemas in novel technological interactions and suggests that users may be applying inappropriate mental models to AI interaction, because of LLMs’ distributed agency.

Hybridity Analysis reveals that most real-world user inputs combine multiple categories in visible or invisible ways of expression simultaneously, creating complex communicative acts that require sophisticated interpretation. This hybridized nature of user inputs connects directly to RLHF training methodologies discussed in Section 3.1.4, as human evaluators assess model responses that address multiple communicative intentions simultaneously.

3.2.2.2 The Negotiation of Meaning: Anthropomorphism, Trust, and Cognitive Biases

When interacting with LLMs, users unconsciously deploy cognitive patterns developed for human social interaction, leading to anthropomorphic perceptions and trust relationships that facilitate communication while simultaneously introducing significant risks.

Anthropomorphism trends in LLM interaction demonstrate remarkable consistency across user populations and interaction contexts. Research indicates that 57% of participants attributed some possibility for consciousness with ChatGPT, while “consciousness attributions were positively related to the frequency of using social AI apps (e.g., Replika, Character AI), not to general chatbots or digital assistants” (Colombatto et al., 2025). Users consistently describe interactions with LLMs using human relational language, referring to AI “experience” and “intelligence” (Colombatto et al., 2025) despite technical understanding that these systems lack consciousness or intentionality. When users perceive LLMs as possessing human-like qualities, they may apply inappropriate expectations regarding consistency, intentionality, and responsibility that can lead to disappointment or harmful over-dependence on AI advice or emotional support. In addition, while anthropomorphism enables intuitive interaction, it obscures infrastructural power — users blame “unhelpful AI” not provider-designed constraints, while it could also act as a necessary user heuristic.

Trust building processes in human-LLM interaction demonstrate patterns similar to human social relationships while exhibiting unique characteristics related to AI system properties. Research using trust frameworks based on Castelfranchi and Falcone’s delegation-based model (Castelfranchi and Falcone, 1998) reveals that users assess LLM trustworthiness across dimensions including competence (ability to perform tasks effectively), willingness (perceived motivation to help users), and safety (assurance that the system poses no harm). User trust in LLMs appears to develop through perceived reliability in task completion, consistency in response quality, and appropriate boundary maintenance around system capabilities (Afroogh et al., 2024). However, this trust building occurs within a context where users cannot directly observe LLM internal states or decision-making processes – the explainability concern, requiring trust assessment based solely on output quality and consistency, but also indicated the demand of an operational new approach of addressing explainability, that include the methodology of intervening training dataset in Section 3.1. Experimental research examining trust dynamics between humans and different AI system types reveals that users exhibit higher trust restoration tolerance for systems identified as human rather than LLM agents, suggesting that anthropomorphic framing directly influences trust dynamics (Afroogh et al., 2024). These findings indicate that trust relationships with AI systems may be fundamentally different from human trust relationships, requiring new frameworks for understanding appropriate trust calibration in human-AI interaction.

Cognitive bias Cognitive bias manifestations in LLM interaction create systematic patterns of over-reliance, under-scrutiny, and inappropriate confidence in AI outputs. Automation bias, originally identified in aviation and nuclear power contexts, manifests in LLM interaction through both commission bias (accepting incorrect AI recommendations despite contradictory evidence) and omission bias (failing to act when AI systems don't alert users to problems). The sophisticated language capabilities of modern LLMs may exacerbate the commission bias by producing responses that appear well-reasoned and confident even when containing significant errors. Omission bias is particularly concerning in high-stakes applications where users may not pursue necessary verification or consultation with human experts because AI responses appear sufficiently complete (Parasuraman and Manzey, 2010). The availability heuristic influences user assessment of LLM accuracy based on memorable instances of correct responses, potentially leading to systematic overestimation of AI capabilities. Research indicates that those cognitive shortcuts developed for managing information overload and decision complexity in human environments may be poorly suited to AI interaction contexts where traditional social cues for reliability assessment are absent (Malberg et al., 2024).

The psychological processes underlying these shortcuts stem from cognitive schemas built through human social experience, which assume that confident, articulate communication indicates competence and reliability (Tiedens and Fragale, 2003). LLMs exploit these schemas inadvertently by producing linguistically sophisticated responses that trigger human trust mechanisms despite the absence of actual understanding or intentionality behind the responses.

Stress and cognitive load appear to increase reliance on these shortcuts, suggesting that users may be most vulnerable to cognitive biases precisely when they most need careful evaluation of AI outputs (Yu, 2016). This dynamic creates particular challenges for high-pressure applications where users may be simultaneously most dependent on AI assistance and least capable of critically evaluating AI recommendations.

4.2.1 Technical Understanding

Learners must be aware of how training, tokenization, context windows, and mediating layer shape what an LLM can and cannot express. Knowing that the causal relationship of training, meditation layer, semantic/token noise, and how to arrange appropriate context for achieving better performance.

4.2.2 Critical Evaluation

Users need structured heuristics—source triangulation, chain-of-thought solicitation, self-consistency checks—to audit outputs. Another crucial dimension is that users must differentiate Facts from LLM-performed inferences, to seek corroboration for factual claims via primary sources or reputable databases and treat analytic summaries or explanations generated by the model as hypotheses requiring verification, not as definitive truths. Besides basic critical thinking approaches, new reflect methodologies for LLM interaction like seeking responses from multiple LLMs for cross verification, or starting a new conversation with former response for reverse verification should also be considered and developed.

4.2.3 Ethical Awareness

Bias, representation, and privacy risks follow users across queries. Frameworks such as Digital Promise’s “Core Values” of Digital Equity Framework (Jackson et al., 2024) emphasizes justice and accountability, yet LLM-specific bias manifests at both training and operating stages. Recognize that model inferences may embed social biases or overgeneralizations absent in the underlying data; question the assumptions underlying these analogies, and apply social‐impact assessment to inferred insights. Further, users should trace the content back to the original source and obtain permission or provide citation in line with fair-use and licensing norms, while acknowledge LLMs’ contribution—e.g., by stating “Drafted with assistance from [Model Name]”—while making clear which portions reflect human-authored interpretation and which are AI-generated.

4.2.4 Interaction Design

Prompt engineering is now a civic skill (Walter, 2024). Refine the prompt - the question - by including or reducing context, clarifying intent, and iterating through feedback loops could achieve better response. Multimodal extensions—e.g., image plus text—follow the same principles of contextual grounding and iterative refinement.

4.2.5 Systemic Awareness

Finally, users must recognize economic and regulatory forces shaping AI outputs, understand access pricing and compliance filters explains why certain answers are poor, truncated or refused, be cautious about permissions authorized to LLM-based agent and know how to break agent’s task if unexpected act occurs. This systemic literacy prevents misplaced trust in LLM and clears the responsibility across developers, providers, and users.

5.2.1 Key Entities in the Interaction Workflow

We identify five primary entities that collectively constitute the human-LLM communication workflow:

1. Human User: The initiator of the communication, formulating prompts, interpreting responses, and often engaging in iterative refinement. The user brings their unique cognitive biases, prior knowledge, communicative intent, and emotional state to the interaction. As discussed in Chapter 4, the user’s LLM Literacy significantly influences their ability to effectively interact with and critically evaluate LLM outputs.

2. LLM: The core algorithmic entity responsible for processing input and generating responses. The LLM’s capabilities are shaped by its architecture (e.g., Transformer), its training data, and its internal parameters. It operates based on probabilistic predictions, generating sequences of tokens that form coherent linguistic outputs (Chapter 3). The LLM itself is not an intentional agent in the human sense, but its outputs can evoke perceptions of agency in users.

3. LLM Developer: The entity responsible for designing, training, and refining the LLM. This includes curating massive datasets, determining model architecture, and implementing alignment techniques like Reinforcement Learning from Human Feedback (RLHF). Developers embed certain values, biases, and capabilities into the model through their design choices, as highlighted in Section 3.1.5. Their decisions profoundly influence the LLM's inherent communicative tendencies.

4. LLM Service Provider: The entity that deploys and provides access to the LLM, often through App, web interfaces, or integrated applications. Service providers implement various technical interventions that mediate the interaction between the user and the raw LLM output. These interventions, as discussed in Section 3.2.1.2, include system prompts, Retrieval-Augmented Generation (RAG), and compliance filters, which collectively exert a significant influence on the final output and its framing. This constitutes a form of triple-layer agenda setting, where the provider shapes not only what information is presented but also how it is framed. Generally, API access faces less intervene because it requires a second develop and is meant to be integrated into an existing system who is the final service provider to users, who is responsible for content security. Under this circumstances, the LLM API provider should not be considered as the only LLM service provider (Williams et al., 2025).

*External Knowledge Bases/Tools: Usually provided by the Service Provider or an Agentic Mediating Layer, while the service provider can only access but cannot intervene, which should be mentioned independently. These include databases, APIs (to access external systems like web search engine) and other tools that the LLM can access to retrieve information or perform actions. As discussed in Section 3.2.1.2, RAG obtains these external sources, introducing another layer of mediation and potential bias through source selection and curation. The Agentic Mediating Layer, as described in Chapter 3, further extends this by enabling LLMs to decompose complex tasks and orchestrate tool use, leading to real-world consequences.

This conceptualization of distributed agency aligns with Rammert’s (2012) notion of “hybrid constellations of inter-agency” in human-machine interaction, where agency emerges from the interaction between multiple actors rather than residing within individual entities (Rammert, 2012).

5.2.2 The Communication Flow and Interdependencies

The interaction workflow is a dynamic, iterative process, moving beyond the simplistic linear or even circular models of traditional communication theory (Chapter 2). It can be conceptualized as follows:

1. User Prompt (Input): The human user initiates communication by a prompt. This prompt is not merely a request for information but an expression of communicative intent, shaped by the user's cognitive frame, prior knowledge, and expectations. The effectiveness of this prompt is directly tied to the user’s LLM Literacy, particularly their prompt engineering skills (Section 4.2.4).

2. Service Provider Mediation (Input Processing): User’s prompt first passes through the Service Provider’s infrastructure. Here, system prompts, compliance filters, and potentially RAG mechanisms (if activated, while the decision of whether initiate a RAG request could also be made by LLM) preprocess the input. This stage represents a critical point of mediation, shaping the input before it reaches the core LLM. The Service Provider also manages access to external knowledge bases and tools, which can be invoked based on the prompt's content or system configurations.

3. LLM Processing and Response Generation: The pre-processed input is then submitted to the LLM. The LLM, leveraging its Transformer architecture, self-attention mechanisms, and vast training data, generates a response. This process is inherently probabilistic, predicting the most likely sequence of tokens based on the input and its learned patterns (Section 3.1.2). The LLM’s context window (Section 3.1.1.3) dictates its ability to maintain coherence over extended interactions.

4. Service Provider Mediation (Output Processing): The LLM-generated response then passes back through the Service Provider. At this stage, further intervention occurs, including additional compliance filtering, reformatting, generating charts according to LLM’s instructions, or the integration of information retrieved via RAG.

5. User Interpretation and Feedback: The user actively interprets the response based on their LLM Literacy, critical evaluation skills (Section 4.2.2), and ethical awareness (Section 4.2.3). The user may then provide feedback for further discussion or refine their prompt then initiate a new turn in the conversation, creating a continuous feedback loop. This iterative process highlights the co-constructive nature of human-LLM communication, where both parties (human and LLM, mediated by developers and providers) mutually shape the discourse.

This Interaction Workflow Layer emphasizes the distributed nature of agency in human-LLM communication. While the user initiates and interprets, the LLM Developer embeds foundational capabilities and biases, and the LLM Service Provider actively mediates the flow of information, shaping both input and output. This complex interplay moves beyond simplistic sender-receiver models, acknowledging the multi-faceted influences on communicative outcomes.

5.3.1 Meaning as Probabilistic Linguistic Combination

Key technical mechanisms influencing this probabilistic meaning construction include:

1. Tokenization and Transformer Architecture: The fundamental unit of meaning for an LLM is the token (Section 1.1). The Transformer architecture, with its self-attention mechanisms (Section 3.1.1.2), allows the LLM to identify statistical relationships between tokens across a given context window (Section 3.1.1.3). This enables the LLM to generate linguistically coherent and contextually relevant responses, even if it lacks human-like semantic understanding. The “meaning” embedded at this stage is purely statistical – the probability of certain tokens co-occurring or following others.

2. Training Data and Fine-tuning: The massive datasets used in pre-training (Section 3.1.2) and the curated datasets for fine-tuning (Section 3.1.3) inherently shape the LLM’s probabilistic understanding of language. Biases, cultural norms, and factual inaccuracies present in the training data can be amplified in the LLM’s outputs, influencing the “meaning” it presents. RLHF (Section 3.1.4) further refines this by aligning outputs with human preferences, but as noted, the reward model is a statistical proxy, not a genuine evaluator of meaning or truth.

3. System Prompts and RAG: As discussed in Section 3.2.1.2, system prompts and RAG mechanisms introduce external influences that shape the LLM’s probabilistic output. System prompts can subtly steer the model’s tone, persona, and even the framing of information, while RAG integrates external knowledge. These interventions directly affect the linguistic combinations presented to the user, thereby influencing the raw material for meaning construction.

5.3.2 Meaning as Human Interpretation and Co-Construction

Key humanistic values and competencies from LLM Literacy (Chapter 4) that are crucial for this meaning construction include:

1. Technical Understanding (Section 4.2.1): A user’s awareness of how LLMs generate responses (e.g., tokenization, context windows, probabilistic nature) is fundamental. Understanding these technical constraints allows users to critically assess the LLM’s output, recognizing that it is a statistical product. This technical understanding helps mitigate the anthropomorphic bias (Section 1.2.2) that can lead to over-attribution of consciousness or intentionality to the LLM.

2. Critical Evaluation (Section 4.2.2): This is paramount for meaning construction. Users must employ structured heuristics—such as source triangulation, chain-of-thought solicitation, and self-consistency checks—to audit LLM outputs. Differentiating facts from LLM-performed inferences and seeking corroboration for factual claims are essential steps in validating the “meaning” presented by the LLM. This critical lens allows users to move beyond surface-level linguistic coherence to deeper semantic accuracy.

3. Ethical Awareness (Section 4.2.3): Meaning construction is also an ethical act. Users must be aware of potential biases embedded in LLM outputs due to training data or algorithmic design. Recognizing that model inferences may embed social biases or overgeneralizations is crucial for responsible interpretation. Ethical awareness guides users to question the assumptions underlying LLM responses and to consider the social impact of the information they receive and disseminate.

4. Interaction Design (Section 4.2.4): The user’s ability to refine prompts and engage in iterative feedback loops directly influences the LLM’s subsequent linguistic combinations. By providing clear, context-rich prompts, users can guide the LLM towards generating more precise and relevant outputs, thereby actively shaping the raw material for meaning construction. This highlights the transactional nature of human-LLM communication, where both parties contribute to the evolving discourse.

5. Systemic Awareness (Section 4.2.5): Understanding the economic, regulatory, and provider-driven forces that shape LLM outputs is vital for accurate meaning construction. Awareness of compliance filters, access pricing, and the “dual-layer agenda setting” (Section 3.2.1.2) helps users contextualize LLM responses, preventing misplaced trust and clarifying the distributed responsibility across developers, providers, and users. This systemic understanding allows users to interpret LLM outputs not in isolation, but within the broader socio-technical ecosystem.

5.3.3 The Interplay: Bridging the Algorithmic and the Human

Meaning does not reside solely within the LLM’s output, nor is it solely a product of human cognition. Instead, it emerges from the dynamic interaction between the two. The LLM provides the potential for meaning through its statistically generated text, while the human user actualizes that meaning through interpretation, critical evaluation, and contextualization.

This bridging process is facilitated by:

1. Algorithmic Transparency (Ideal): While full transparency into LLM internal workings remains a challenge, understanding the principles of Transformer architecture, tokenization, and training methodologies (Chapter 3) allows users to develop a more informed mental model of the LLM. This reduces the “black box” effect and enables more realistic expectations of LLM capabilities and limitations.

2. Human-Centric Design: Developers and service providers have a crucial role in designing LLMs and their interfaces to facilitate responsible meaning construction. This includes implementing clear indicators of AI-generated content, providing tools for source verification, and designing systems that encourage critical engagement rather than passive consumption. The call for humanities interventions in the training pipeline (Section 3.1.5) directly supports embedding human-centered values at the algorithmic level, influencing the probabilistic outputs to be more aligned with ethical and societal norms.

3. Iterative Refinement and Feedback Loops: The continuous feedback loop between user and LLM, mediated by the service provider, allows for iterative refinement of meaning. Misinterpretations or unsatisfactory responses can be corrected through subsequent prompts, gradually aligning the co-constructed meaning. This process mirrors the transactional models of communication (Section 2.1.3) but with an explicit acknowledgment of the algorithmic partner.

In essence, the Meaning Construction Layer argues that effective human-LLM communication is a collaborative endeavor where the human user, equipped with LLM Literacy, actively engages with the LLM’s probabilistic outputs to construct meaningful interpretations. This layer moves beyond simply describing what LLMs do (generate text) to explaining how humans make sense of what LLMs do, thereby providing a robust framework for understanding the semantic and pragmatic dimensions of this novel communicative paradigm.

5.4.1 Re-evaluating Traditional Communication Theories (RQ1)

To what extent can traditional communication theory explain the human-LLM interactions?

This hybrid model demonstrates that while these theories offer partial lenses, they are fundamentally insufficient to capture the full spectrum of this novel communicative paradigm. The Interaction Workflow Layer and Meaning Construction Layer collectively highlight where traditional models fall short and how our hybrid approach provides the necessary extensions.

Limitations of Linear Transmission Models (Shannon-Weaver): As discussed in Section 2.1.1, Shannon-Weaver models quantify token-level noise but fail to account for semantic meaning or distributed agency. Our Meaning Construction Layer directly addresses the semantic meaning by emphasizing that while LLM outputs are probabilistic linguistic combinations, the human user actively constructs meaning through interpretation and critical evaluation. The Interaction Workflow Layer explicitly recognizes the distributed agency across users, LLMs, developers, and service providers, moving beyond the fixed sender-receiver dichotomy. The hybrid model thus provides a framework where semantic fidelity and distributed influence are central, not peripheral, concerns.

Limitations of Interactive Communication Models (Schramm): Schramm’s models introduce feedback loops but presume symmetrical cognitive capabilities (Section 2.1.2). Our hybrid model acknowledges the iterative feedback loops within the Interaction Workflow Layer but explicitly accounts for the cognitive asymmetry between humans and LLMs. The Meaning Construction Layer highlights how the human user’s LLM Literacy is essential for bridging this gap, enabling them to critically evaluate the LLM’s probabilistic outputs rather than assuming shared understanding.

Limitations of Transactional Communication Models (Barnlund): While transactional models recognize co-construction and context-dependence, they assume human communicators with shared perceptual channels and do not adequately address technical constraints like memory loss (Section 2.1.3). Our hybrid model embraces the co-constructive nature of communication within the Meaning Construction Layer but grounds it in the technical realities of the Interaction Workflow Layer. The model explicitly incorporates LLM-specific limitations such as finite context windows (which lead to a form of non-human forgetting) and the potential for out-of-distribution hallucinations, providing a more comprehensive account of the transactional dynamics.

In essence, our hybrid model does not discard traditional theories but rather integrates their valuable insights into a broader, more technologically informed framework. It demonstrates that while traditional theories can explain certain aspects of human-LLM interactions, a comprehensive understanding requires a hybrid approach that accounts for distributed agency, cognitive asymmetry, technical constraints, and the unique process of meaning construction in this new communicative context.

5.4.2 Constructing a Hybrid Model (RQ2)

What hybrid components best model human-LLM communication dynamics?

Our hybrid model directly answers this question by proposing a two-layered structure that synthesizes elements from computer science, communication studies, and media studies.

The Interaction Workflow Layer: This layer is a crucial component for modeling the distributed agency and multi-faceted mediation inherent in human-LLM communication. By identifying the four key entities (Human User, LLM, Developer and Service Provider) and mapping their interdependencies, this layer provides a comprehensive blueprint of the communication pipeline. This moves beyond simplistic models to capture the complex socio-technical system in which human-LLM interactions are embedded. The concept of triple-layer agenda setting (Section 3.2.1.2), where user prompts set the first-level agenda, system prompts/RAG set the second-level agenda and the indirect macro-level, the third agenda (accumulated from the second-level shaped among conversations of all users) is a prime example of a hybrid component that accurately models the subtle yet powerful influence of algorithmic mediation.

The Meaning Construction Layer: This directly addresses the epistemological challenge of meaning in human-LLM communication. By positing that meaning is a co-constructed phenomenon—emerging from the interplay between the LLM’s probabilistic linguistic combinations and the human user’s interpretive framework—it bridges the gap between technical functionality and humanistic understanding. The integration of LLM Literacy competencies (Technical Understanding, Critical Evaluation, Ethical Awareness, Interaction Design, Systemic Awareness) within this layer provides the necessary human-centric tools for navigating the complexities of algorithmic meaning-making. This layer acknowledges that while LLMs generate text, humans ultimately construct meaning, thereby reclaiming the user’s central role.

Integration of Technical and Humanistic Elements: The hybrid model’s strength lies in its seamless integration of insights from Chapter 3 (technical mechanisms) and Chapter 4 (LLM Literacy). For instance, understanding the LLM’s context window (Section 3.1.1.3) informs the Critical Evaluation competency (Section 4.2.2) by highlighting the limitations of long-term memory. Similarly, the call for humanities interventions in the training pipeline (Section 3.1.5) directly feeds into the ethical considerations within the Meaning Construction Layer, advocating for the embedding of human-centered values at the algorithmic level. This interdisciplinary synthesis is a key hybrid component that allows for a comprehensive and actionable understanding of human-LLM communication.

5.4.3 Benefits and Risks of Treating LLMs as Communicative Agents (RQ3)

What benefits or risks emerge from treating LLMs as communicative agents?

This hybrid model suggests that while users may naturally anthropomorphize LLMs (Section 1.2.2), a critical understanding of the underlying mechanisms is essential for mitigating risks and harnessing benefits.

Benefits: The perception of LLMs as communicative agents can enhance user engagement, facilitate more natural and intuitive interactions, and promote co-creative partnerships. The Meaning Construction Layer acknowledges that users actively co-construct meaning with LLMs, and a degree of perceived agency can make this process more fluid and productive. For example, in educational or therapeutic contexts, a perceived sense of partnership can foster trust and encourage deeper engagement. The model suggests that the benefits of treating LLMs as agents are realized when users are equipped with high levels of LLM Literacy, allowing them to engage critically and ethically.

Risks: The primary risk of treating LLMs as communicative agents without a critical understanding is the potential for over-trust, emotional dependency, and susceptibility to misinformation and manipulation. This hybrid model highlights these risks through several components: The Interaction Workflow Layer reveals the hidden mediation by service providers, demonstrating that the “agent” the user interacts with is not a singular entity but a product of multiple influences. This awareness helps mitigate the risk of misplaced trust in a seemingly autonomous agent. The Meaning Construction Layer emphasizes the probabilistic nature of LLM outputs, reminding users that the LLM is not an intentional, truth-seeking agent. This understanding is crucial for combating the risk of accepting hallucinations or biased information at face value. The LLM Literacy competencies, particularly Critical Evaluation and Ethical Awareness, provide the necessary tools to deconstruct the illusion of agency and to critically assess the LLM’s outputs. By understanding that the LLM is a sophisticated linguistic tool rather than a conscious entity, users can better manage their expectations and avoid the pitfalls of uncritical anthropomorphism.

In conclusion, our hybrid model suggests that the question is not whether to treat LLMs as agents, but rather how to engage with them given our natural tendency to do so. The model advocates for a form of “critical anthropomorphism”, where users can leverage the benefits of a more natural interaction style while remaining critically aware of the underlying technical mechanisms and socio-technical influences. This nuanced approach, grounded in the dual layers of this model, provides a robust framework for promoting responsible and beneficial human-LLM communication.

This hybrid model serves as a foundational theoretical framework for future research into human-LLM communication. It provides a lens through which scholars can analyse the evolving dynamics of human-AI interaction, offering insights for the responsible design, development, and deployment of LLM technologies. By emphasizing the interdisciplinary nature of this field, the model underscores the imperative for continued collaboration between computer science, communication studies, and other humanities disciplines to ensure that the advancement of AI serves human-centered values and promotes effective, ethical, and equitable communication in the age of artificial intelligence.