AI-textuality: Expanding intertextuality to theorize human-AI interaction with generative artificial intelligence

Abstract

This paper introduces the concept of AI-textuality that extends Bakhtin’s notion of intertextuality to encompass interactions involving texts produced by generative artificial intelligence (GenAI). Intertextuality provides a valuable lens for understanding how GenAI outputs are created through the assemblage of digital and multimodal texts from vast datasets. Building on this perspective, this paper presents a model to theorize the intertextual nature of AI-generated texts in educational settings. The model frames human-AI interaction as an interplay between humans’ experiences in a material space and GenAI’s computational processing in a digital space. Drawing on data from a high school classroom, the model illustrates how students engaged with GenAI to discuss key ideas in the lesson. It highlights how students’ engagement with AI-generated texts was shaped by their textualization of experiences, the intertextual probability of GenAI’s responses, and the joint construction of knowledge between students and GenAI. Through the illustration, the paper argues that AI-textuality has the potential to revise and expand traditional understandings of intertextuality to incorporate our interactions with GenAI in the digital age.

Introduction

Generative artificial intelligence (GenAI) is a type of AI that learns from human language systems to produce original texts of various semiotic modes, including verbal language, images, music, and video. Most GenAIs utilize large language models (LLMs) to identify language patterns from an extensive corpus of textual data, enabling them to generate coherent and human-like texts. Given the centrality of language in the operation and application of LLMs, linguistics theories offer powerful tools to understand the use of GenAI in various social contexts (Tang 2024a). However, research exploring human-AI interaction from a linguistic perspective remains limited, with only a few exceptions (e.g. Ahn et al. 2024; Dippold 2024). Maly (2024: 11) aptly noted that the lack of a “sociolinguistic theory on AI” hinders a theoretically grounded discussion on the investigation of GenAI in applied linguistic research. Particularly, current research on GenAI has paid little attention to how texts are accessed, interpreted, produced, and communicated through the interactions between individuals and GenAI systems.

From a language and literacy perspective, I have previously argued that human-AI interaction should be understood as mediated actions involving reading, writing, speaking, and listening through texts (Tang 2024b). This perspective underscores the role of language as a meaning-making resource in mediating human-AI interactions. In this paper, I extend this line of inquiry further by drawing on the concept of intertextuality to better understand and analyze human-AI interaction. Intertextuality denotes the idea that the comprehension, use, and creation of a text are shaped by other related texts, or “inter-texts.” This concept was a significant focus of research in applied linguistics and education during the 1980s and 1990s (Lemke 1992), culminating in a special issue in the journal Linguistics and Education (Bloome 1992), and an edited volume titled, Uses of Intertextuality in Classroom and Educational Research (Shuart-Faris and Bloome 2004). With the increasing prevalence of digital media at the turn of the century, intertextuality has evolved to encompass a complex web of texts across modes and media, thereby influencing and overlapping with the research on multimodality (e.g. Tang and Tan 2017; Gordon 2023) and transmedia (e.g. Jenkins 2008; Deumert 2014; Duret, Duret and Pons 2016).

Given that GenAI outputs are generated through the assemblage of multiple texts that are digital and multimodal in nature, intertextuality offers a valuable perspective to understand these interactions. In this paper, I introduce the concept of “AI-textuality” to theorize the intertextual nature of AI-generated texts. AI-textuality refers to the intertextual process through which meaning is co-constructed between humans and GenAI systems, encompassing the interpretation, production, and integration of AI-generated texts within a broader network of interrelated texts. With AI-textuality as a new concept, the paper aims to develop and illustrate a model of human-AI interaction in educational settings that apply chat-based GenAI tools (e.g. ChatGPT, Claude, Gemini).

I begin by reviewing the literature on intertextuality, tracing its intellectual roots from Bakhtin to more contemporary scholars in the social sciences. Following this, I introduce the model incorporating AI-textuality and illustrate its utility with data involving high school students using ChatGPT in a classroom environment. I conclude by discussing the potential of this model to revise and expand traditional notions of intertextuality that can incorporate our interactions with GenAI in the digital age.

Theoretical Roots of Intertextuality

Intertextuality refers to the juxtaposition and interconnectedness of texts—broadly defined to include written words, utterances, and other multimodal semiotic resources—where each text is shaped by and contributes to a network of other texts. The theoretical basis of intertextuality can be traced back to Mikhail Bakhtin’s (1981) seminal works on dialogism and heteroglossia in the 1920s and 1930s. Dialogism posits that all language is inherently dialogic, meaning that every utterance or text is a response to previous texts and anticipates future responses. This dialogic nature of language implies that texts are always in conversation with one another, forming a continuous chain of speech communication across time and space. Heteroglossia, meanwhile, refers to the coexistence of multiple voices or discourses within a single text. This plurality of voices reflects the social and ideological diversity of the context in which the text is produced. Heteroglossia always contributes to the dialogic nature of texts, as these various voices often reference or echo other texts, either explicitly or implicitly. Bakhtin’s work underscores an essential idea that meaning is not fixed but emerges from the interaction between different voices and texts, each bringing its own perspective and context.

Building on Bakhtin’s ideas, Julia Kristeva (1980) coined the term “intertextuality” in the 1960s. Kristeva expanded Bakhtin’s notions of dialogism and heteroglossia, proposing that “any text is constructed as a mosaic of quotations; any text is the absorption and transformation of another” (p. 37). She argued that texts do not derive meaning in isolation, but through their relationship with other texts, whether explicitly (through direct quotation or allusion) or implicitly (through shared cultural frameworks). Kristeva also highlighted the role of the author in constructing meaning, as the author brings their own intertextual knowledge to bear on the text, creating it through the lens of other texts they have encountered. Her work challenged traditional notions of authorship and originality, suggesting that the creation of a text is less about the solitary genius of an author and more about the recombination and recontextualization of existing texts.

Intertextuality in Educational Research

In educational research, intertextuality has been applied to understand how students and teachers engage with texts in various social and classroom contexts. Broadly speaking, the literature can be grouped into four perspectives, each offering a unique view of how texts shape and are shaped by the social interactions and cultural understandings of people who engage with the texts. Each perspective of intertextuality is also crucial to my theorization of AI-textuality.

Social Material Construction of Intertextuality

This perspective emphasizes how intertextual connections are formed through social and material interactions within educational settings. Rooted in the work of Bloome and Egan-Robertson (2004) and expanded by others (e.g. Ivanic 2004; Pappas et al. 2004), this social material view highlights that intertextuality is not a static feature residing within a text, but is one that must be actively constructed and negotiated through discourse and social practices in a material world.

According to Bloome and Egan-Robertson (2004), intertextuality is a linguistic process that must be proposed, recognized, and acknowledged by participants in an interactional event. In an education context, intertextuality is constructed through teacher-student interactions, where explicit connections between texts are often drawn through prompts, questions, and references embedded in the classroom discourse. Furthermore, they argue that intertextual connections are not abstract but are grounded in the material basis of language and interactional events (c.f. Pennycook 2017). This material basis includes the concrete use of sound, ink on paper, pixels on a digital screen, and other tangible resources that provide the material anchors for intertextual referencing, as well as the transformation of these materials through our social interactions (e.g. listening to writing or typing).

Pappas et al. (2004) extend this understanding by examining how students engage with texts during classroom interactions. Their study highlights how first- and second-grade children actively construct meaning from information texts during read-aloud sessions with their teachers. They identified four categories of intertextual links: textual, experiential, personal, and world knowledge links. These categories illustrate how students connect new information with prior readings, hands-on exploration, personal experiences, and general knowledge. This approach emphasizes the interactive and intertextual nature of reading in educational settings, where meaning is co-constructed as students relate new information to existing knowledge and broader discourse within the classroom. Teachers also play a key role in facilitating this process as they guide students to recognize and establish these intertextual connections.

From this perspective, a text cannot be defined a priori outside the specific situational context in which it is constructed and used. In other words, a text is not a static entity. Instead, it can only be defined in the moment of interaction when it is read, written, viewed, drawn, or made by humans. Bloome and Egan-Robertson (2004: 29) argue that “a text is the product of textualizing. People textualize experience and the world in which they live, making those phenomena part of a language system.” Thus, the experience that people have with the material world is integral to intertextuality.

In terms of AI-textuality, this perspective highlights the fluidity and situated nature of texts, emphasizing that meaning is always co-constructed within specific human interactions. The perspective is crucial to AI-textuality as it emphasizes that AI-generated texts are meaningless without their intertextual connections to human interactions in specific socio-material contexts. For instance, students’ engagement with GenAI during a classroom activity is mediated not only by the prompts they enter, but also by the surrounding materials, discussions, and tools, which all play a crucial role in co-constructing the meaning attributed to AI-generated responses.

Social Semiotic Construction of Intertextuality

The social semiotic construction of intertextuality explores how texts are interwoven through both explicit and implicit references to other texts, emphasizing the broader social patterns and ideological meanings that shape these intertextual links. This perspective is primarily informed by the works of Fairclough (1992) and Lemke (2004), both of whom were influenced by Halliday’s (1978) view of social semiotics.

In intertextuality research, Fairclough (1992) makes a useful distinction between manifest and constitutive intertextuality. Manifest intertextuality refers to direct and explicit references to other texts, such as quotations, citations, or paraphrases, which are identifiable from linguistic markers or contextualization cues that people use to create links across texts (Bloome and Egan-Robertson 2004). Constitutive intertextuality, or interdiscursivity, involves the implicit “configuration of discourse conventions that go into [a text’s] production” (Fairclough 1992: 104). Going beyond the level of text to its constitutive social practice, constitutive intertextuality reflects the heteroglossic voices interacting within a text and influencing its content, structure, style, and tone. Through critical discourse analysis, Fairclough provides examples of constitutive intertextuality, showing how political or corporate rhetoric, for instance, can be infused with underlying ideologies that are not explicitly stated but are nonetheless present in the text’s language and framing. These intertextual connections are closely tied to the concept of genre, which Fairclough views as more than a structural feature of texts. Genres are sets of conventions that influence how people produce and interpret texts, and serve as sites where implicit voices and ideologies are negotiated and maintained. This makes constitutive intertextuality crucial for understanding the ideological underpinnings of a text and how different discourses and voices interact, compete, and influence one another within a given context.

Lemke (2004) further connects constitutive intertextuality to Halliday’s (1978) systemic functional linguistics (SFL). SFL posits that language serves as a resource for creating meaning across three metafunctions: (1) ideational—for representing the external world and our experiences, (2) interpersonal—for enacting social interactions and relationships, and (3) textual—for organizing information within a text. From these metafunctions, Lemke argued that the “patterns of intertextuality” can be determined by how likely members of a particular community connect texts that share the same topic (ideational), point of view (interpersonal), or structure (textual). These patterns shape our recognition of related intertexts and how we use them to make meaning of a particular text. In some learning areas like the natural sciences, the intertexts that are recognized as discussing the same topic (e.g. energy, evolution, Periodic Table) are remarkably consistent and standardized, thus allowing members of the community, such as teachers, to easily assess whether different texts convey the same scientific concepts even if the exact wording differs. This recurring “thematic pattern,” as Lemke (1990) calls it, provides a tool to understand and describe how shared intertextual meanings are constructed and communicated within a discursive community.

The social semiotic construction of intertextuality is particularly critical for understanding how GenAI (particularly LLMs) operates by recognizing the linguistic patterns of intertextuality from a massive corpus of texts. For example, when prompted to write about “climate change,” a GenAI tool utilizes statistical probabilities derived from its training data to identify and reproduce the thematic patterns often used in scientific and media texts on this topic. The resulting outputs, which align with those dominant patterns, can be observed and traced through manifest and constitutive intertextuality. The process of how LLMs identify and reproduce thematic patterns will be further elaborated in a later section.

Readers’ Construction of Intertextuality

This perspective focuses on how individual readers make sense of a text through the lens of their accumulated reading history. According to Hartman (2004), intertextuality is not merely a feature of texts, but is also shaped by the complex interaction between the text, the reader, the author, and the broader context in which reading occurs. This view positions the reader as an active participant in the construction of intertextuality who, like a centoist, “synthesized various discourses, motifs, and images by stitching together the textual voices of others into a patchwork intertext that resembled a collage/montage” (Hartman 2004: 357). In other words, readers draw from a vast reservoir of textual resources, constructing an evolving “inner text” that reflects the interplay of various influences and embedded voices (c.f. Kamberelis and Scott 2004). As readers encounter new texts, they engage in an intertextual loop; new insights from the current text are used to revise past interpretations, which in turn inform their ongoing engagement with the current text.

The continuous process of forming textual links is facilitated by readers’ metacognitive awareness (Flavell, Miller and Miller 2002), which supports their ability to monitor and reflect on their evolving understanding (Zohar and Barzilai 2013). As readers actively engage with their inner texts, metacognition enables them to create a web of meaning to navigate a complex space where multiple perspectives, discourses, and interpretations intersect. In some ways, the notion of inner text and metacognition extends Vygotsky’s (1986) postulate of an “inner speech” that is internalized from social speech.

The reader’s construction of intertextuality is essential for understanding how readers interpret AI-generated texts, as it emphasizes the role of prior knowledge and inner texts in shaping meaning during their interactions with GenAI. For instance, a reader’s interaction with GenAI depends on their accumulated textual history to evaluate the GenAI’s response. This evaluative process may involve revising their “inner text” to accommodate new insights generated through the interaction, illustrating how meaning is co-constructed dynamically between the reader and GenAI. Moreover, readers’ prior exposure to specific genres, cultural discourses, or thematic patterns also shapes how they contextualize and integrate GenAI outputs into their broader knowledge framework.

Multimodal Construction of Intertextuality

Lastly, multimodality extends the traditional notion of intertextuality by expanding the definition of “text” to include non-verbal modes such as images, sounds, gestures, and spatial arrangements. This broader understanding acknowledges that meaning is constructed not only through words but also through multiple semiotic modes that interact to create complex webs of intertextual connections. In this sense, the development of multimodality theory builds on the foundation of intertextuality research, which emphasized the interconnectedness of texts and the dynamic processes through which meaning is generated (Lemke 2004).

Drawing from a social semiotics perspective, Lemke (1998) explores how meanings are “multiplied” when different semiotic modes are combined. He argues that meaning-making is not confined to a single semiotic system (e.g. verbal language) but emerges from the interplay of multiple modes, where the meanings produced in one mode can contextualize and reshape the meanings in another. In educational texts, this co-contextualization of meaning allows students to synthesize information from various sources to gain a more holistic understanding. Lemke’s (1998, 2004) work extends traditional logocentric boundaries of intertextuality by incorporating multimodality into the generation of new meanings and understandings. Kress and van Leeuwen (2006), in their seminal work on social semiotics, further elaborate how meaning is shaped through the orchestration of different semiotic resources based on their affordances. For example, while written language is well-suited for expressing categorical and abstract ideas, images are more suitable for conveying spatial and quantitative concepts more effectively. Kress and van Leeuwen also emphasize the agency of meaning-makers in selecting and combining these resources, which are influenced by the context and purpose of communication.

The multimodal construction of intertextuality is vital for understanding how modern communication forms a complex web of interconnected meanings across semiotic modes and technologies, evolving from analogue multimedia in the 1960s, through digital media in the 2000s, to the current use of multimodal GenAI in the present age. This perspective underscores the dynamic interplay between semiotic modes, where each mode contributes uniquely to the overall meaning. It is relevant to AI-textuality as contemporary GenAI tools can operate across multiple semiotic modes, such as generating textual explanations alongside visual or auditory outputs, and vice versa. For instance, a student might prompt GenAI to generate an explanatory text that includes a graph or diagram. The interplay between these modes requires users to navigate and integrate multimodal outputs into their understanding. This co-construction of meaning highlights how AI-textuality leverages multimodality to create richer and more nuanced interactions, where human users interpret and synthesize AI-generated content across verbal, visual, and other semiotic modes. By bridging different modes, AI-textuality demonstrates the importance of considering multimodal connections in understanding and designing educational interactions with GenAI.

Model of AI-Textuality

Informed by prior research on intertextuality, I developed a model to understand how human-AI interactions with GenAI are mediated through the production, interpretation, recognition, circulation, and juxtaposition of texts, a concept I refer to as AI-textuality. The key components and relationships within this model are visually illustrated in Fig. 1 and elaborated in this section. The arrows in Fig. 1 represent the intertextual influence from one form of text to another. A thin arrow signifies a specific and causal flow of text in the intertextual process; for instance, a user textualizes their experience in a particular moment within an interactional context into a prompt for the chatbot. The user’s text is subsequently processed by the LLM to generate a response, which is interpreted by the user, and so on. A block arrow indicates a broader influence between the components, reflecting more general intertextual relationships outside a specific place and time.

Interactional Context

The best place to start the elaboration of this model is at the point when users first enter a prompt into a chat-based GenAI system. From the social material perspective of intertextuality, the interactional context is critical in shaping the users’ prompt. Informed by Bakhtin’s (1981) dialogism, the first prompt entered into a GenAI chat is not the very first text in a continuous chain of prior texts. Instead, the users’ prompt is always mediated and shaped by the immediate texts present in the interactional context. For example, students might be reading a textbook, writing a report, or watching a video, and these interactions shape their engagement with GenAI. In this sense, students are actively juxtaposing different texts (i.e. textbook, written report, video, AI-generated texts) to create intertextual meanings across them during the interaction. The interactional context also includes the voices of other participants and the surrounding activities and physical settings, all of which may influence the intertextual process. For instance, a student might discuss with a peer how best to frame a question for GenAI, or the material objects involved (e.g. whiteboard, laptop, scientific apparatus) may also shape the resources used to make those intertextual connections.

Through these interactions with relevant texts and material objects in the interactional context, users then “textualize” these socio-material interactions into prompts for the GenAI system. The concept of textualization (Bloome and Egan-Robertson 2004) emphasizes that the users’ text in any prompt is not merely a string of written or spoken words but a product of human experience within a material world. One of the major critiques of GenAI is that it is disconnected from the physical world, as “the language processing and outputs generated by AI are not contextualized to a material reality” (Tang and Cooper 2024: 5). Therefore, the users’ textualization play a critical role in contextualizing GenAIs’ digital outputs to their experiences. Without this crucial intertextual link to the user’s texts, one can argue that GenAI outputs are contextless and thus meaningless.

Once prompted by the users’ text, GenAI then processes the words or symbols in that text and generates a responding text. When the users interpret the AI-generated text, the process is again mediated by an intertextual connection to the immediate texts in the interactional context, continuing the cycle depicted in Fig. 1. In addition to these immediate material-based texts, the users’ “inner texts” (Hartman 2004) also play a role in shaping their interpretation.

Inner Text

Drawing on Hartman’s (2004) work on intertextuality, the inner text refers to the users’ history of textual resources and engagements that shape their interpretation of the present text. When users engage with AI-generated texts, their interpretation is informed not only by the juxtaposition of the immediate texts within the interactional context but also by an internalized network of previously encountered texts evoked during the reading process. This suggests that interaction with GenAI is a dialogic process in which users actively synthesize and revise their understanding by connecting new texts to their reservoir of previously encountered texts, whether those texts are generated by humans or GenAI.

Metacognition—awareness and self-regulation of one’s thought processes—empowers users to connect AI-generated texts to their inner texts by reflecting consciously on how their prior knowledge, experiences, and assumptions shape their interpretation of the AI’s outputs. For example, users might ask themselves when interpreting AI-generated texts: “How does this information align with what I have read elsewhere, and where might it challenge or expand my understanding?” By engaging in such reflective questioning, users can critically evaluate the relevance, accuracy, and potential biases in the AI-generated content. This metacognitive process is crucial to ensure that user’s engagement with GenAI is driven by an active construction of intertextual meaning to their inner texts, while simultaneously updating and revising their understanding.

Human-AI Interface

As shown in Fig. 1, the human-AI interface involves the exchange of texts between users and GenAI. This component is the most visible aspect of the AI-textuality model as readily seen in the interface of most commercial chatbots. During a chat session, the users’ prompts and AI-generated texts are typically retained and displayed (or spoken) in a conversational manner. An example of such a conversation with ChatGPT is shown in Fig. 2. From the perspective of manifest intertextuality, this human-AI chat is developed from a series of interconnected texts that build upon one another. For example, users may refer to key phrases or quotes introduced by the GenAI in a prior exchange and build on those ideas in their prompts for further information. They may also employ metadiscourse markers (Tang 2017) to refer back to a previous exchange (e.g. revise the earlier paragraph on … ) or to direct future responses (e.g. in future, I want you to … ).

Similarly, GenAI systems exhibit manifest intertextuality by temporarily storing users’ texts during a single chat session to provide more contextually informed responses for subsequent exchanges. For instance, GenAI can explicitly reference an earlier exchange to create a web of meaning that incorporates users’ information shared in prior exchanges. Some GenAI systems like ChatGPT also offer optional memory settings that permit the GenAI to store users’ details and preferences across chat sessions, with the purpose of providing more tailored responses in future conversations. This is also a form of manifest intertextuality built into GenAI systems.

The social material view of intertextuality provides a useful framework for conceptualizing AI-generated texts as more than just strings of words, sounds, or pixels assembled by a machine based on statistical probability. As previously discussed, a text cannot be defined a priori in isolation from the situational context in which it is constructed and used. Outside of a conversation, an AI-generated text has no intrinsic meaning. Instead, the meaning of an AI-generated text is constructed at the moment when a user prompts for the text and subsequent reads and interprets it. In other words, an AI-generated text only becomes meaningful within the context of human interaction. This view of AI-generated texts expands Draper and Broomhead’s (2010) definition of a text as “any representational resources or objects that people intentionally imbue with meaning, in the way they either create or attend to the object, to achieve a particular purpose” (p. 28, italics mine for emphasis) Therefore, an AI-generated text can be considered a text, not because it contains words, symbols, or clusters of patterned pixels, but because people imbue it with meaning. The intertextual connections between human interpretations and GenAI outputs thus underscore the essential role of human interaction in attributing meaning to AI-generated texts.

Large Language and Multimodal Models

The next component in the AI-textuality model is large language models (LLMs) and large multimodal models (LMMs). Intertextuality provides a valuable lens to explain how GenAI operates through LLMs and LMMs. LLMs, like OpenAI’s GPT (Brown et al. 2020) and Google’s BERT (Devlin et al. 2019), are trained on extensive datasets drawn from an enormous corpus of texts. To understand any given word in a language system, the model first analyzes the context of the word in relation to other collating words from all the texts in its training data. Using a computing technique called “transformer architecture” (Vaswani et al. 2017), the LLM transforms the relationships among those words into a vector (list of numerical values), which represents a complex probability of potential words or word sequences that tend to co-occur. When generating text, the model uses these values to calculate the most likely word sequences in response to a user’s text input. This calculation based on large training data is what enables LLMs to understand relationships between words through their intertextual patterns. For images, the training process is similar, as LMMs analyze visual data by identifying intertextual patterns in shapes, colors, and spatial relationships among adjacent clusters of pixels, then transforming these patterns based on a probabilistic calculation into coherent outputs.

In many ways, the process used by LLMs/LMMs reflects the social semiotic construction of intertextuality. Like human meaning-making, the model does not treat an individual word or cluster of pixels as isolated entities but recognizes them as part of a larger, interconnected web of meaning “against the background of other texts” (Lemke 2004: 3). Each AI-generated text is not created in isolation but is a reconfiguration of fragments, or “mosaic of quotations” (Kristeva 1980: 37) drawn from prior texts. Thus, like human-authored texts, AI-generated texts are also shaped by implicit influences of underlying discourses and voices derived from a network of other texts. However, a key difference is that AI-generated texts rely on statistical associations in a computational process, without any awareness of the social or ideological contexts underpinning human communication. To give an example, an LLM trained on a specific dataset is more likely to use the term “English as second language learners,” reflecting an Anglocentric discourse that foregrounds deficiency in the use of English, rather than a term like “emerging language learners” that may be more commonly found in a different corpus. Similarly, an LMM like DALL-E 3 is more likely to generate an image of a male scientist, unless instructed otherwise, if its training data has historically featured male scientists (see Cooper and Tang 2024). These examples reflect how LLM and LMM rely on the statistical frequency of words or images in their training data, shaping the intertextual probabilities toward particular voices embedded in the corpus, without any awareness of the social and ideological forces behind those representations.

Lemke’s (2004) concept of thematic pattern provides further insight into how LLMs/LMMs can generate content that often seems accurate. As the LLM identifies recurring linguistic structures in its training data, it also picks up the thematic patterns of how certain keywords and phrases tend to cluster around specific topics or ideas. These patterns are key in enabling the model to predict which words or sequences are most likely to follow a given input. These repeated thematic patterns align naturally with the probabilistic calculations of human language determined by LLMs during their training. Just as Lemke (2004) argues that members of a community use thematic patterns to make connections between texts with similar topics, points of view, or structures, LLMs similarly operate by identifying these patterns across their training corpus. Similarly, LMMs extend this process by identifying recurring visual patterns across vast image datasets, allowing them to generate outputs that align with dominant visual discourses. The result is an AI-generated multimodal text that reflects the thematic patterns present in the training data. With this understanding, we can predict that the more frequently texts of a recurring topic appear in the training data, the more likely GenAI can generate the same thematic pattern with greater accuracy.

Training Data and Feedback Mechanism

From an intertextuality perspective, it is clear why the training data has an enormous role in shaping the voice of a GenAI’s output. Popular GenAI tools like ChatGPT, Gemini, and Claude are trained on large and diverse datasets. According to OpenAI (2024), the LLMs powering ChatGPT are “developed using three primary sources of information: (1) information that is publicly available on the Internet, (2) information that [they] license from third parties, and (3) information that users or human trainers provide.” All training data have a certain cutoff date, which means that the LLM is not trained by new data beyond that date. Once the training is complete and the statistical calculations are finalized, the LLM no longer accesses the training data to generate texts. For instance, the training data used for ChatGPT-3.5 and ChatGPT-4o are based on information up to September 2021 and October 2023, respectively (OpenAI Platform n.d.). However, newer GenAIs after ChatGPT-4 now have browsing capabilities that allow the model to access real-time information from the web when needed, depending on the nature of the user's query.

When users interact with GenAI, their text inputs are typically used to provide feedback and fine-tune the LLM. This does not mean that their texts are stored in a database or added to the training data. Rather, as mentioned earlier, the LLM analyzes the statistical associations between the words in the users’ texts, transforming those associations into numerical values that update the weighting of the vectors linked to those words. In this sense, the behavior of LLMs and LMMs is not entirely static but can be influenced by the collective voices of users through intertextuality. This implies that aggregated user interactions can shift the model’s parameters over time, allowing it to adapt and reflect the evolving language patterns of its user community. On the positive side, this feedback mechanism enables GenAI to become more attuned to the changing discourse of its users. On the downside, GenAI may become more proficient and biased toward the dominant languages and cultures of its users, while limiting its capability in engaging with underrepresented languages and voices of minority groups.

It is a commonly mistaken view that a GenAI system stores textual data within its LLM or LMM in order to analyze or retrieve those data when prompted. Furthermore, while most GenAI platforms must be connected to the Internet to access the servers on which they operate, this does not mean that they always retrieve real-time data from the Web for every prompt. As explained and shown in Fig. 1, a GenAI responds to most interactions based on the statistical probabilities calculated from its training data before the cutoff date. It does not retrieve any text stored within itself or from the Web, unless the GenAI requires recent or real-time data from the Web based on the users’ query.

Influence of AI Company

No model of AI-textuality would be complete without addressing the influence of the AI companies that design and build GenAI systems. It is important to recognize that GenAI technologies are developed by companies with vested socioeconomic interests. These companies therefore have significant voices in shaping how their LLMs and LMMs operate. For instance, the GPT model that is used by OpenAI’s ChatGPT prioritizes text generation and conversation in line with its original goal. By contrast, Google Gemini prioritizes information and language understanding, which is more aligned with Google’s business model. By tracing the voices of the GenAI to their corporate owners, an intertextuality perspective reminds us to critically examine how economic and political agendas can influence the diversity of voices reflected in human-AI interactions.

Methods

The AI-textuality model presented in this paper was developed based on a theoretical grounding of intertextuality in educational research (as summarized earlier) and an empirical analysis of classroom interactions involving a GenAI tool from a research project in Australia. In this section, I present an overview of the project and the data that led to the development of the AI-textuality model.

Research Project and Design

The data used in this paper to illustrate the AI-textuality model were drawn from a larger action research project situated in an independent school in an Australian capital city (Tang et al. 2024). The project involved a collaboration with four high school teachers to implement lesson activities incorporating the use of ChatGPT 3.5 for student learning and assessment. The research objective was to develop a proof of concept illustrating the use of GenAI in secondary school teaching and to evaluate student learning gains through this process. More information of the project can be found on this website: https://gainer.au/

For this paper, the data were analyzed using a case study design (Stake 2000) to support the development of the AI-textuality model. In particular, the following research question guided the analysis: What are the key intertextual processes involved in the co-construction of meaning between students and GenAI during classroom activities? The purpose of this case study is not to make generalized claims about AI-textuality in any context. Instead, the purpose is to present a “telling case” (Mitchell 1983) to illustrate the theoretical model in a specific educational context. Telling cases are particularly useful because “understanding them will lead to better understanding, perhaps better theorizing, about a still larger collection of cases” (Stake 2000: 437). Hence, the selection of the case is to enable researchers to gain a better understanding of AI-textuality and subsequently develop it for further theorization and investigation.

Research Data Sources and Analysis

The telling case shown in this paper is taken from the work of one teacher and 15 of his students in a Grade 10 English classroom. Ethnographic methods comprising student chat logs, classroom observation, video recording, and interviews were used to generate data from the classroom. The individual student chat logs comprised the primary data source as they provided direct records of student interactions in the human-AI interface. Each student had a personal account with ChatGPT 3.5. After their chat, they were instructed to share their chats with the research team via email. Subsequently, the chat logs were analyzed based on the type of questions asked by the students. There was a total of 36 student questions and each question was coded according to their discursive functions (e.g. elaborate, expand, contextualize, verify, clarify). See Tang et al. (2024) for more details on the discursive functions.

Another primary data source was classroom observation and video recording of three key lessons that occurred in the classroom. Two researchers were involved as participant-observers during those lessons. Two high-definition video cameras mounted on tripods were used to record the classroom interactions. One camera on wide-angle captured the overall classroom environment centering on the teacher, while the other camera zoomed in on a group of students to capture their interactions with their laptops and occasionally with each other. Following Erickson’s (1992) interactional ethnography approach, the classroom videos (5 h in total) were systematically viewed and divided into distinct and meaningful segments called episodes, with boundaries determined by shifts in the interaction or task. Memos and labels describing the episodes were generated to support the analysis and purposeful selection of key moments for further interpretation. The analysis is essential for understanding the interactional context that shaped the students’ prompts in ChatGPT.

Finally, semi-structured interviews with the teacher and selected students were collected as a secondary data source to support data triangulation in the case study. The teacher interview focused on understanding the pedagogical design and his reflections on the use of ChatGPT in the lessons. The student interviews explored their experiences with ChatGPT, focusing on how they framed their prompts and interpreted the AI’s responses. These interviews offered a complementary perspective on the data, enabling a deeper understanding of the interplay between the classroom environment, participants’ agency, and AI-mediated interactions.

The research project and data collection followed ethical guidelines and were approved by the Human Research Ethics Committee of Curtin University (Project number: HRE2023-0413).

Illustration of AI-Textuality

The following case is selected to illustrate AI-textuality and the key intertextual relationships involved in the meaning-making process during a lesson on poetry. The main task in the lesson required students to research a poet of their choice, focusing on how the themes in a particular poem were influenced by the sociopolitical issues experienced by the poets (e.g. racism, disillusionment). At the start of the lesson, the teacher explained and demonstrated how to ask questions using ChatGPT, stressing the importance of having a conversation with GenAI and checking its responses. He also modeled an example based on his interest in Wilfred Owen’s poems. After their research with ChatGPT, the students completed a written assignment providing a thesis statement and supporting evidence about their selected poet.

To highlight key intertextual relationships that exemplify the AI-textuality model in action, a student called Zack (pseudonym) is presented here as an illustrative case. His interactions with ChatGPT also reflect patterns observed among other students in the classroom, thus representing similar findings found in the dataset. Figure 2 shows a chat session initiated by Zack, who was researching the works of T. S. Eliot. During the lesson, Zack had three separate chat sessions with ChatGPT. The first chat session focused on the Romantic poet Percy Shelley. After a few exchanges, for some reason, Zack decided to change his focus to T. S. Eliot in the second and third sessions. In the third chat (which is shown in Fig. 2), Zack’s initial question was similar to the one he asked about Shelley in the first chat, except he replaced “the Romantic poet Percy Shelley” with “the Modernist poet T.S. Eliot.”

Textualizing From Experience to Textual Prompts

Informed by the AI-textuality model, it is essential to examine the immediate texts within the interactional context that shape how users like Zack framed their initial prompt in the chat. To analyze this, we look for evidence of other texts that students had juxtaposed in constructing their questions for ChatGPT (Bloome and Egan-Robertson 2004). One crucial text came from the teacher. When the teacher explained how to use ChatGPT at the beginning of the lesson, he wrote several steps on the whiteboard for the students to follow during their interactions with GenAI (see Fig. 3). Comparing this text with Zack’s prompt, it is evident that Zack borrowed the phrase “please provide some insight into the context” from the first step of the teacher’s instruction. In doing so, he appropriated the teacher’s voice, which was itself shaped by the voices in the curriculum focusing on poetry analysis. However, Zack did not simply copy the teacher’s prompt entirely. He also exercised agency in choosing T.S. Eliot and adding more context and details. In his prompt, Zack stated the purpose of his inquiry explicitly, breaking down “context” into five distinct categories—historical, social, economic, cultural, and political—and requested specific examples from Eliot’s poems in each category. Other students also followed a similar approach to their questions. This shows how the students’ initial texts in the human-AI conversation became a “mosaic of quotations” (Kristeva 1980: 37) that juxtaposed multiple voices from the teacher, the curriculum, and themselves.

Besides the teacher’s text, Zack and other students were also juxtaposing other texts during their interactions with ChatGPT. In the classroom, each student had a personal laptop with several applications open, including a browser-based ChatGPT 3.5, Google Chrome for Internet searches, and Microsoft OneNote for taking notes. For instance, when Zack was interacting with ChatGPT, he had another browser tab open with T.S. Eliot’s poem “The Hippopotamus.” In a previous chat session, Zack copied the text of the poem into ChatGPT and asked: “can you provide some insight into the poetic devices used in this poem?” ChatGPT responded by highlighting several prominent features, including metaphor, contrast, imagery, personification, and allusion. As he read these texts, Zack copied sentences and paragraphs from ChatGPT and other websites into OneNote, where he later recontextualized them for his assignment. Some students also copied photographs of their poets from web searches into their OneNote assignments. In this way, the students’ interactions with ChatGPT were mediated by a juxtaposition of multiple texts across several computer applications.

The students’ inner text is another important factor in mediating their interactions with GenAI. In Zack’s case, his distinction between the “Romantic poet Percy Shelley” and “the Modernist poet T.S. Eliot” is particularly telling. During the lesson, there had been no discussion of Romanticism and Modernism, nor the specific poems associated with these movements. According to the teacher during an interview, the students had previously learned about the structure and focus of poetry in different historical periods. This suggests that Zack connected his interaction with AI-generated texts to previously encountered texts beyond the immediate context of the lesson. Although it is unclear why Zack shifted his focus from Shelley to Eliot, his follow-up question about the criticism of the church suggests he may have identified more closely with the themes in Eliot’s work (e.g. disillusionment, contradiction) based on his own experiences. Zack’s inner text thus played a role in shaping his interpretation of the AI-generated texts and his subsequent actions in the interaction.

In sum, the various texts juxtaposed in the interactional context and students’ inner texts were essential in shaping the ongoing human-AI interaction. Through these interactions, the students’ experiences were textualized into their prompts for ChatGPT. This process of textualization is what contextualized the AI-generated texts to the students’ material reality. In any analysis of a human-AI chat (e.g. Fig. 2), it is important to consider the textualization process that occurs in the material space surrounding the users’ text inputs. Without this consideration, it is easy to overlook the meaning that is constructed through the interaction between the users and GenAI.

Intertextual Probability of GenAI’s Responses

Moving onto the GenAI side of the AI-textuality model, both manifest and constitutive intertextuality are useful tools for analyzing ChatGPT’s responses to Zacks’ textual inputs. In terms of manifest intertextuality, it is clear from Fig. 2 that the structure of the first AI-generated text directly mirrored Zack’s initial prompt. ChatGPT’s first paragraph on T.S. Eliot’s protest themes corresponded to Zack’s stated purpose in the first line. In addition, ChatGPT’s organization of Eliot’s work into historical, social, economic, cultural, and political contexts aligned with Zack’s request for “insight” and “detailed responses and examples” in each area.

To understand the extent of constitutive intertextuality in AI-generated texts, it is useful to compare Zack’s chat with that of another student, Steven, who was also researching modernist poetry. Steven asked ChatGPT, “What events happened during modernism poetry?” The AI-generated response is shown in Fig. 4. While the AI-generated texts in Zack’s and Steven’s chat (Figs. 2 and 4) are distinct, we can see a recurring thematic pattern (Lemke 1990) involving these common words: modernist poetry, T.S. Eliot, Great Depression, economic struggles, and disillusionment. This thematic pattern appears not only in these two instances of AI-generated text but also across thousands of texts on T.S. Eliot found in literary and educational books, media, and websites throughout the English-speaking world. This example illustrates that as LLMs learn to identify the intertextual patterns of the English language from its training data, they also replicate the thematic patterns that tend to cluster around prominent ideas, authors, or events. The recurring cluster of meanings in these patterns explains how ChatGPT generated similar discourse about T.S. Eliot and modernism in both Zack’s and Steven’s chats.

Intertextual Co-construction of Knowledge

The final component in the AI-textuality model to examine is the human-AI interface, where students’ series of questions and ChatGPT’s responses were exchanged. Tracing the intertextual linkages between Zack’s inputs and the AI-generated responses show how his prompts shaped the trajectory of the dialogue. For instance, Zack’s follow-up question, “elaborate on the context present in The Hippopotamus,” reflected a juxtaposition of both ChatGPT’s voice and his own. In the first AI-generated response, ChatGPT provided contextual information on T.S. Eliot but did not address “The Hippopotamus.” Zack’s follow-up question thus directed the focus to the poem, demonstrating his agency in the interaction. After ChatGPT’s response, Zack’s subsequent question on “criticisms of the church” juxtaposed the satirical nature of the poem (introduced by ChatGPT) with his own interest in pursuing the reasons behind the criticism.

These follow-up questions, posed by students like Zack, serve important dialogic functions in driving the interactions with GenAI. Some questions such as “elaborate” contribute to the ongoing dialogue, while questions like “clarify” help regulate or verify the information provided by GenAI (see Tang et al. 2024). This range of follow-up questions provides discursive tools for users to shape the content generated by GenAI. Thus, based on the intertextual connections between users’ questions and GenAI’s responses, we can argue that both humans and GenAI have different roles in co-constructing knowledge during these interactions.

Discussion

The AI-textuality model presented in this paper offers a novel framework for understanding human-AI interactions, particularly in educational settings. By extending intertextuality to include AI-generated texts, this model repositions GenAI not merely as a technological tool, but as a joint participant in the process of meaning-making with humans. It also frames GenAI system as a socio-technical assemblage that incorporates: (1) the interactional context that situates the user’s textual inputs in a material world; (2) the user’s inner text representing their prior knowledge and textual history; (3) the intertextual connections between LLMs/LMMs and the training data, and (4) the human-AI interface that can be analyzed as a heteroglossic exchange between the user and numerous voices amalgamated by GenAI. Additionally, the model highlights the co-construction of meaning as a dynamic process influenced by both human and GenAI contributions. These components together illustrate how AI-textuality operates as an interconnected system of text production, interpretation, and interaction between humans and GenAI.

While the AI-textuality model in this paper encompasses the multimodal capabilities of GenAI, such as the generation of images and other semiotic modes, the empirical focus of this study was limited to linguistic-based interactions. This emphasis aligns with the research context of an English classroom, where students engaged with ChatGPT primarily to explore and analyze linguistic content. Future research could extend this work by investigating how multimodal GenAI tools support meaning-making across different semiotic modes in educational settings.

The reconceptualization of human-AI interactions from the AI-textuality model has significant implications for both educational research and practice, as it opens up new ways of thinking about how students and teachers interact with, interpret, and use AI-generated texts. It also has the potential to revise the traditional concept of intertextuality to incorporate our interactions with GenAI as a non-human participant in the meaning-making process.

Utility of the AI-Textuality Model in Classroom Teaching and Learning

One of the utilities of the AI-textuality model is to broaden the analysis of student learning with GenAI in the classroom. In most classroom settings, students often engage with various texts (e.g. textbooks, digital resources, writings on whiteboards, peer contributions, scientific apparatus), and GenAI responses become another form of text in this intertextual web of meaning-making. The model highlights how students’ prompts and their subsequent interactions with GenAI are shaped by: (1) the texts they encounter in their material surroundings, (2) the role of textualization from their experiences to their textual prompts, and (3) the intertextual probability of GenAI’s responses from its textual data. For example, Zack’s interactions with ChatGPT were not isolated exchanges between him and a machine. Instead, the interactions were informed by the teacher, the curriculum, and his personal reading history, all of which were interwoven into a dialogue with an amalgamation of texts that was statistically assembled by the GenAI. The model and the analysis in this paper illustrate how GenAI operates within a complex intertextual network, whereby meaning-making is co-constructed through the interplay of multiple texts and voices in a socio-material context.

The model also highlights the pedagogical value of promoting an intertextual co-construction of knowledge through dialogic interactions between students and GenAI. This process is evident in Zack’s follow-up questions, which shaped the responses from ChatGPT and led to a collaborative process of knowledge building. Encouraging students to engage in such iterative questioning and dialogic interaction with GenAI has the potential to promote deeper learning and critical thinking. Teachers can harness this knowledge by designing assignments or tasks that emphasize the iterative nature of learning through dialogue with GenAI, thereby promoting a classroom environment that values exploration, reflection, and revision with GenAI as a dialogic partner instead of passively accepting authoritative answers from a machine.

Implications for Educational Research and Practice

The AI-textuality model also has broader implications for educational research, particularly in examining the role of GenAI in multimodal and transmedia learning environments. As the model demonstrates, GenAI operates through the assemblage of multiple texts across different modes and media. This opens up new avenues for studying how AI-generated content interacts with traditional texts and how students navigate this increasingly complex digital landscape. In another study, I have explored the potential of multimodal GenAI in generating curriculum-aligned images for educational purposes. Future research can focus on how students juxtapose AI-generated texts with other semiotic resources to create their own multimodal texts for various purposes and contexts. In the AI-textuality model, this connection is highlighted in the “inner text” component (see Fig. 1), which can include drawings, videos, music, and other popular media that children engage with outside of school.

In terms of education practice, the AI-textuality model suggests that teachers need to take on a new role as facilitators of human-AI interaction. They need to help students develop the skills to critically reflect and evaluate how they are generating content with GenAI, and not by GenAI. This requires a shift in teacher education to prepare teachers to understand how to navigate the intertextual complexities that arise in human-AI interactions. Teachers must also be prepared to address the ethical implications and risks of using GenAI, including issues of bias and the potential for misinformation. Moreover, the feedback mechanism of GenAI, as described in the model, highlights the importance of understanding how GenAI systems learn and evolve based on user interactions. By understanding these dynamics, teachers can foster a more critical and reflective learning environment where students are able to not only use GenAI effectively, but also responsibly.

Revising Intertextuality in an Age of GenAI

The AI-textuality model proposed in this paper offers a significant revision to the traditional concept of intertextuality by introducing GenAI as a non-human participant in the intertextual process. Traditional intertextuality theory has largely focused on the relationships between human-authored texts, emphasizing how texts reference, echo, or transform one another within a social and cultural framework (Kristeva 1980). However, with the increasing use of GenAI, the boundaries of this theory need to be expanded. From an intertextual perspective, AI-generated texts are not independent of human inputs. They incorporate both the users’ voices through their prompts and the community’s voices through the intertextual patterns derived from a massive corpus of human-authored texts. AI-textuality thus redefines intertextuality by acknowledging that non-human entities like GenAI are not passive repositories of information, but active agents that influence the access, interpretation, production, and communication of texts. This shift also aligns with a post-humanist and new materialism paradigm that recognizes the shared agency between humans and non-humans in language and communication (Barad 2007; Pennycook 2017).

By conceptualizing the role of GenAI as a non-human participant in an intertextual chain of texts, the concept of AI-textuality allows us to explore how GenAI contributes new forms of knowledge and discourse in research and education. This challenges the traditional notion of authorship and originality as AI-generated texts are created through the recombination of past voices and knowledge from existing texts, but without the same conscious intent or ideology that humans bring to text production. Consequently, AI-textuality also invites educational researchers and theorists to reconsider how authority, agency, and voice are distributed in meaning-making processes where human and non-human actors collaborate. This expanded view of intertextuality positions GenAI as a transformative and collaborative partner in the evolving digital landscape of knowledge creation and dissemination.

Notes on contributors

Kok-Sing Tang is a Professor and Director of Graduate Research in the School of Education at Curtin University. He holds a BA and MSc in Physics from the University of Cambridge and an MA and PhD in Education from the University of Michigan. His research examines the role of language, discourse, and multimodality in scientific literacy and generative AI. He is editor-in-chief of Research in Science Education (RISE) and author of “Discourse Strategies for Science Teaching & Learning” published by Routledge.

Acknowledgements

This paper draws on data from the research project: “Transforming Teaching, Learning and Assessment with Generative AI in a Western Australian School” funded by the Innovation & Excellence Award, School of Education, Curtin University.

Ethical approval

This study adhered to the ethical considerations and was approved by the Human Research Ethics Committee of Curtin University (Project no.: HRE2023-0413).

Conflict of interest

I certify that I have no affiliations with or involvement in any organization or entity with any financial interest or non-financial interest in the subject matter or materials discussed in this manuscript.

References