Composing with AI

What We already Know: Generative AI and the History of Writing With and Through Digital Technologies

Anthony T. Atkins University of North Carolina Wilmington
Colleen A. Reilly, University of North Carolina Wilmington

Introduction

As Johnson (2023) recently reminded writing faculty, computers and writing/digital rhetoric scholarship, pedagogy, and experience have prepared us for teaching with developing technologies—AI presents a new challenge but not one that should cause us panic. After all, as Johnson and Agbozo (2022) emphasize, faculty "have been using technology to teach, critique, and remediate the writing process for more than one hundred years" (p. 158). We have a rich scholarly tradition that demonstrates basic principles, including that all technologies are political and rhetorical and that "policing is not pedagogy" (Johnson, 2023, p. 172). In 1999, at another pivotal moment in the history of teaching with web and digital composing technologies, Selfe admonished all writing faculty to pay attention to and interrogate the use and construction of technologies because failing to do so allows the social inequities they exacerbate to go unchallenged. Many scholars embraced that perspective and examined writing and communication technologies in terms of their construction, labor implications, and ethical ramifications. Based in this extensive scholarly tradition, our chapter highlights some of the strategies scholars have already developed to work with digital technologies, however new and powerful, to enhance our writing pedagogies and prepare our students for all contexts. We argue that we have access to the most qualified scholars to address the challenges of AI because they recognize solutions to conundrums presented by previously developed digital platforms and tools.

To this end, our chapter focuses on three central themes present in the literature that we saw as most generative when grappling with AI. These themes include the challenges posed to our conceptions of writing and authorship; the access and accessibility implications of information and communication technologies; and the degree to which technologies reveal and mask their mediation of content. Our chapter draws upon only a small amount of the wealth of scholarship from the corpus related to computers and writing and digital rhetoric; we selected texts that most resonated with our focus, knowing that full coverage was not feasible. Additionally, our chapter addresses only the text generation type of AI, specifically chatbots, while acknowledging that many other types exist that create visual, audio, and multimedia compositions. Our discussion is enhanced by short video responses that we integrated into our chapter that provide four scholars' responses to this question: What are the central lessons that our field of computers and writing learned in working with past digital technologies that can best prepare us to assist our faculty and students when composing with generative AI? As these videos illustrate, our field has a wealth of relevant knowledge to draw upon and this knowledge continues to grow and inform how we as scholars and faculty navigate researching, writing, and teaching with AI.

Challenges to Writing and Authorship

During the early 21st century, scholars in computers and writing and digital composition theorized and interrogated the relationships between technologically-mediated forms of composition and traditional concepts of writing, texts, and authorship. The research in this vein highlights the challenges to traditional standards of composition and authorship posed by a range of digital content and development processes, including coding, writing markup and metadata, and authoring in multimedia. Teachers and scholars examined and proposed strategies to aid students to reconceive of their roles as writers and authors when working with a myriad of digital compositions and learn to use the technologies necessary to produce them. For example, authors of digital compositions, such as hypertexts, cede meaning to users who chose their own pathways through the linked content, acting as co-authors and emphasizing the inability of authors to control the integrity of their text and make sustained linearly structured arguments (Cripps, 2004; Hocks, 2003; Purdy & Walker, 2012). This past scholarship grappling with a range of challenges related to new composing practices and context proves relevant for composing and teaching with generative AI.

In light of the recent escalation in the development of writing and communication technologies, it is easy to forget that at the turn of the last century, arguing for importance of non-discursive texts and other types of compositions as literate practices that require equal emphasis in writing curricula was a disruptive and even radical move. For example, in a nonlinear print article, Wysocki (2001) explores how visuals assume primacy in meaning-making in digital environments, forcing a reconsideration of what it meant to author content at the start of the 21st century (see also Cripps, 2004; Hocks, 2003). Wysocki asserts that words are “always visual elements," but the argumnents made in hypertexts "cannot even be found primarily in ‘words’” (p. 232). In a similarly disruptive and multimodal text that is at once article and representation of an oral address originally supported by synchronized slides (84 in total), Yancey (2004) questions the nature of writing and the teaching of composition in response to the proliferation of the genres spurred by new information and communication technologies. Her remarks situate the communicative revolution of the early 21st century historically in relation to previous moments of change in literacies such as the serials and newspapers of the 19th century and the evolution of writing instruction in higher education during the 20th century. Yancey (2004) calls for a new composition focused on circulation of texts, genres, media, and content production across domains. This revisioned composition necessitates interactions between media, an integration of the canons of rhetoric including assembly, and processes of mediation and remediation facilitated by and represented in digital technologies that challenge reconsideration of established rhetorical strategies.

Other scholars focused on the expansion of literacy standards and instruction to include producing non-discursive content that underpins and facilitates the production of texts in digital environments. For instance, Sorapure (2006) highlights four categories of composing practices in Flash, a commercial product, that contribute to meaning making, including the text and images displayed to readers/users and the code and comments underpinning the structure of the composition, making the content possible, and enabling social interactions between developers. When Sorapure (2006) published her article, these categories posed a significant challenge to what it meant to be a text, a writer, and a user—blurring established boundaries by shifting the roles depending on the specific use and moment of composition and consumption. Similarly, Cummings (2006) emphasized the importance of viewing coding (specifically markup languages, like XML) as an act of writing albeit requiring expanded audience considerations as the “coder’s first audience is a machine” (p. 434); unless the machine can comprehend the code as written, the content will not be visible and usable by human audiences. As Eyman and Ball (2014) also detail, digital compositions require a greater range of literacies to produce successful compositions, including proficiencies in technical infrastructural support for the development and dissemination of multilayered digital compositions. Designers of digital compositions needed to integrate optimal coding, metadata, file formats, and other technical affordances to compose usable, accessible, and visually appealing webtexts and other digital designs (Eyman & Ball, 2014). To improve their products, writers of code, like writers of text, also “refine ideas” (Cummings, 2006, p. 433) to achieve their desired output. Interestingly, as Lancaster (2023) highlights, chatbots like ChatGPT also refine ideas by remembering a certain amount of content from a conversation or interaction with a user and building upon it to provide more targeted responses.

The challenges to traditional concepts of writing and authorship that scholars have been grappling with over the last two decades (and before) have been magnified exponentially by the introduction of generative AIs like ChatGPT. In his discussion of new and challenging forms of visual compositions, Kress (2005) argued that writing would remain a powerful force because “elites will continue to use writing as their preferred mode” (p. 18); however, the introduction of AIs like ChatGPT destabilizes that mode of communication making it not the province of human elites but of the machine—open and accessible to all who can use it effectively (Byrd, 2023). When generating content with an AI, writers direct the output in part by authoring appropriate prompts directing the AI to respond according to specified parameters. While collaborating with the AI, writers are also collaborating in remove with the creators and writers of the content on which the AI was trained. As Lancaster (2023) explains, AI language models such as ChatGPT respond "in a predetermined way, based on its trained model, the input data, earlier parts of the conversation and a random number, known as a seed” (p. 3). Creating the best strategic input for the AI in the form of successful prompt engineering is essential to facilitating the most effective and relevant output from the AI, recalling Cummings’ (2006) identification of the machine as the first audience for code-based compositions. As Cummings (2006) noted in relation to writing code, “The act of writing for the machine and writing for a human audience develop similar skills, and one experience can be harnessed to inform the other” (p. 442). However, in this case, the AI also participates as an author, a conversant, and an active collaborator in producing texts with human actors.

The participation of generative AI in the composing process further disrupts established norms related to writing and authorship by prompting a reassessment of the standards by which any composition is evaluated as “original writing.” This aspect of working with generative AI like ChatGPT has, as of this moment in 2023, caused the most widespread consternation in scholarly communities and the public. However, examining past scholarship in digital composition reveals that questioning notions of individual authorship and plagiarism preceded the introduction of generative AI. As noted above, proficiency in digital composition requires collaborating and coauthoring with digital tools to produce multilayered content, some of which is only readable by and may be developed through machines. The codes, metadata, and other digital constructs may be derived through using models and templates produced by human and nonhuman actors, which, as Johnson-Eilola and Selber (2007) argue, forces us as writing teachers to reconsider our standards for identifying and evaluating authorship, originality, and plagiarism. Johnson-Eilola and Selber (2007) highlighted the importance of assemblage in digital composition—using existing codes, templates, digital objects, and even texts and reconfiguring and repurposing them to solve specific problems or address contextual needs (see a more recent discussion of assemblages in Yancey & McElroy, 2017).

This reconsideration involves revisioning of authorship, originality, and even creativity (Johnson-Eilola & Selber, 2007, p. 400). Authoring with generative AI extends and magnifies these redefinitions, because AI technologies now have the potential to produce content more creatively; as such, these technologies are no longer just a tool that aids the human user to achieve their ends. A recent publication by Johnson-Eilola and Selber (2023) emphasizes this point—that agency is no longer only human—people have to think like a technology to be successful users in “complex sociotechnical systems” (p. 83). They argue for an object-oriented ontology (OOO) that focuses on the creative reuse of digital elements to make new assemblages to solve communication and design problems. Particularly relevant for working with AI is Johnson-Eilola and Selber’s (2023) admonition that “OOO asks us to think like objects, decentering ourselves or flattening the normally hierarchical ontologies that put humans on top” (p. 86). In our new writing and communication environment, humans certainly cannot presume to be paramount in creativity, production, or importance. Rather than focusing their efforts on detecting students’ use of AI with software or embedded watermarks (Lancaster, 2023), an effort that promises frustration resulting from a never-ending cycle of design, detection, and redesign of AI, instructors should focus on teaching students to productively use AI for specific sorts of writing and design tasks (Lancaster, 2023) and seek AI chatbots developed by more ethical human actors (Byrd, 2023). As Lancaster (2023) concludes, assignments that can easily be completed by AI should be rethought and potentially eliminated.

Writing with technologies in digital environments has also highlighted the need to examine the infrastructure—technical, cultural, and organizational—that do and do not enable the technologies to function, be accessible, and participate in communicating content with human users. In considering the importance of infrastructure concerns related to digital composing, past scholarship again provides useful insights. As DeVoss et al. (2005) argue, we notice infrastructure at points of breakdown and disruption. DeVoss et al. (2005) highlight the when of infrastructure—and the systems that construct it and determine its use—that comes into play at points of disruption and continues to evolve with attempts at use and intervention when composing. Infrastructure is ubiquitous and relies on standards and other policies, often textual in nature (Frith, 2020), that facilitate its function and imbue the system with values, perspectives, and ideologies. DeVoss et al. (2005) highlight the importance of examining the “often invisible issues of policy, definition, and ideology” (p. 16) that underpin infrastructures essential for digital compositions and composing practices. More recently, Frith (2020) provides an example methodology for investigating such technical standards in his study of the Tag Data Standard “which is the major standard for the Electronic Product Code and a key piece of the Internet of Things” (p. 403). Frith (2020) highlights the role of standards written by people as “discursive infrastructure” (p. 403) in making the physical and technical assets of digital spaces function but in a way that is generally invisible to humans interacting with those technologies:

Obviously, for people who create standards, these documents are a major part of their job. The standards seemingly disappear, on the other hand, when their guidelines are built into material objects and rendered invisible to end users. And related to relationality, one of my major arguments in this article is that technical standards show how writing can become an infrastructure upon which other infrastructures are built. Take the Internet as an example. The Internet is enabled by layers upon layers of material infrastructure, including cables, modems, and so forth. Those material infrastructures are built upon and shaped by international standards documents. (Frith, 2020, p. 406)

As DeVoss et al. (2005) and Frith (2020) both emphasize, numerous texts contribute to building technological infrastructures that enable digital compositions, including policies, standards, and codes.

Users often only pay attention to and consciously investigate infrastructures when they break down or prevent them from performing desired tasks; the scholarship of DeVoss et al. (2005) and Frith (2020) alert us to examine the structural texts that help to yield such disruptions and locate the strategies and ideologies behind them. Generative AI is no different—the technical standards that make it work come into question when the technology causes problems or functions in an unexpected or seemingly aberrant manner. For example, generative AI can “hallucinate,” meaning that it produces rational and real-sounding content that is actually false (Lancaster, 2023). Hatem et al. (2023) cite Chat GPT 3.5’s own definition of hallucinations in which the AI explains that such false information is generated when “a machine learning model, particularly deep learning models like generative models, tries to generate content that goes beyond what it has learned from its training data” (p. 2). The modeling done by generative AIs to produce text responses based on established paradigms without concern for content veracity could be seen as an extreme version of the need to harness identifiable genres for participatory communication as outlined by Bawarshi (2000). As Hatem et al. (2023) emphasize, AI hallucinations pose serious consequences for those relying on AI for healthcare information; such problems are structurally part of how current generative AI functions, making it crucial for humans to interrogate the veracity of the information they receive. As a side note, using the word hallucinations instead of the more accurate misinformation “is inaccurate and stigmatizing to both AI systems and individuals who experience hallucinations” (Hatem et al., 2023, p. 2).

That the technical constructs running generative AI result in the dissemination of false information highlights the need for closely examining the infrastructure, including the standards, making these technologies work. As we write this in October 2023, the Biden Administration issued an Executive Order requiring standards for AI safety and security (White House, 2023). This executive order addresses some of the issues that have been raised regarding the dangers of AI identified by those who have worked on developing and using the technologies (e.g., Hatem, et al. 2023). For instance, the Executive Order requires that “developers of the most powerful AI systems share their safety text results and other critical information with the U.S. government” and “develop standards, tools, and tests to help ensure that AI systems are safe, secure, and trustworthy” (The White House, 2023). Additionally, the National Institute of Standards and Technology (NIST) is in the process of monitoring, developing, and encouraging adherence to standards for AI. As their website reflects, these standards are in flux, and AI continues to evolve. The research in our field proves useful in providing all scholars and writing instructors with the motivation, methodologies, and theoretical perspectives to use to examine infrastructures, including those important for and in development around AIs.

The final category in this section relates to the professional development needed to engage productively with and help students to learn how to work effectively with digital composing technologies. Parallel with the calls discussed above asking for writing scholars and instructors to redefine composition practices to include coding and collaborations with technologies is the recognition of the significant learning curve that scholars and instructors face. For example, Sheppard (2009) highlights the additional efforts required of faculty in computers and writing and digital rhetoric to help students to productively learn to use new information and communication technologies in a sophisticated way requiring the “work of theory, analysis, and argumentation” (p. 123) related to these technologies. Sheppard (2009) highlights the skills needed to develop useful multimedia texts—both new literacies and technological skills. Generative AI makes even more demands through exponential developments in technological complexities coupled with a lack of transparency in terms of the design and structures that constitute these new technologies; as a result, a robust collaborative peer education environment is needed to meet this challenge (Byrd, 2023; Korzynski et al., 2023; Lancaster, 2023). As the immense scholarly interest around writing with AI demonstrates, scholars and instructors in computers and writing and digital rhetoric have embraced their responsibility to help their students navigate working successfully with this and all new information and communication technologies just as they did in previous eras (Sheppard, 2009).

Access, Accessibility, and AI

Another theme within the scholarly tradition of computers and composition is access and accessibility. Digital access, we tend to believe, is ubiquitous, that everyone has access to the Internet and to information. While the digital divide has decreased since the year 2000, it continues to affect the same demographics now as it did then. Grabill (2003) explains this divide by noting how it affects people from diverse backgrounds: “In terms of these broad demographic categories, the divide still exists and is still highly correlated to income, education, and race/ethnicity” (p. 462). Composition scholarship has addressed access many times, most notably in a special issue of Kairos: A Journal of Rhetoric, Technology, and Pedagogy called Accessibility and Multimodal Composition (Yergeau et al., 2013). In this section, we extend previous discussions of access and accessibility as they relate to Generative AI.

Access is an issue with many previous writing and communication technologies and refers to users' ability to obtain the machine, the tools, and/or the information needed to complete tasks that access claims to provide. Without the fundamental tools to access information and engage in critical thinking about information found in online environments, students' educational growth is stifled, leaving them unprepared for rapidly changing workplace and organizations contexts. At the college level, we know that some universities and departments within universities provide more access than others. For example, Atkins and Reilly (2009) note this imbalance in their local ecology:

However, although students in our writing courses are often not aware of the roadblocks imposed by university infrastructures and institutional politics to developing sustainable new media composition initiatives, they are certainly cognizant of the personal consequences of these impediments: inadequate resources, inconvenient or irregular technological access, and inconsistencies in educational experiences across the same degree program. (n.p.)

We recognize the importance of access not only within college structures but in rural communities (among other kinds of communities) where access to the Internet, for instance, remains significantly impeded.

As many scholars of rhetoric and writing have highlighted for decades, African Americans and other communities of color have been systematically excluded from full access to and representation in digital environments. To redress these exclusions, scholars, including Adam Banks (2005), have proposed concrete strategies to expand access to and inclusion in digital spaces. Banks (2005) argues that African Americans have pursued “transformative access” to digital technologies and power through using these technologies. Banks' (2005) approach parallels that of our chapter: he examines African American rhetorical strategies in a variety of historical contexts, highlighting those that can be harnessed for interventions in current digital spaces. Charlton McIlwain (2020) also draws on history to demonstrate the often-overlooked participation of African Americans in the development of computer and digital technologies over time as well as the U. S. Federal government’s use of computer technologies to surveil and disrupt the Civil Rights Movement. Likewise, Brinker (2009) uses case studies of nine digital diaspora communities to examine how they harness online access and digital spaces to forge community, galvanize around their identities, and affect change in their current environments and that of the countries that they may have left. Her approach provides a framework for exploring the work of diasporic communities in digital spaces.

While access to digital devices and the Internet continue to improve, there are other, sometimes, invisible challenges related to access. Access becomes complicated further when we, more specifically, investigate who has the access to knowledge about how interfaces are and can be constructed and who can affect algorithms that control the data delivered in digital environments. As Grabill (2003) notes when referring to similarities between multiple variables that influence digital access to information:

Understanding such a complex of connections allows for the development of a rhetoric of the everyday that has theoretical power and empirical relevance. It allows some understanding as to how culture is constructed, how identity is conceived and practiced, and how any number of public acts of persuasion are carried out and given meaning within concrete (and discursive) contexts. (p. 458)

Arola (2010) interrogates this hidden content that leaves some users disempowered when working with design templates, such as those rolled out by Web 2.0 technologies and social media platforms like MySpace and Facebook. As Arola (2010) points out repeatedly from the title to her piece, “The Rise of the Template, the Fall of Design,” to her explanations of how users are “discouraged” from attempting to make adjustments or design decisions with the platforms of Facebook and MySpace:

In spite of what seems to be pedagogical attention toward modes beyond the alphabetic, we need to acknowledge that in practice Net Generation students, as well as ourselves, are discouraged in Web 2.0 from creating designs. We are certainly posting information, but this information has become “content” placed in a “form” beyond the user’s control. (p. 6)

Thus, when we talk about access, we also mean not just access to the machine, to the mobile, or even to platforms, but rather to the ways that any of them operate below the surface. Arola (2010) explains that what is missing from social media platforms is access to creatively alter designs and make the space a place of their own. Access to algorithms, interfaces, design, and other tools that may be considered “back-end” operations are rendered invisible by labeling them as confidential and proprietary or accessible only to those with specialized knowledge. In the case of Web 2.0 and social media platforms, design is lost, is invisible, and inaccessible.

Similar issues related to a lack of access to basic aspects of how technologies function and a related inability to control their output manifests when working with generative AI. For example, most commercial chatbots do not reveal the corpus used to train their AI—thus users cannot interrogate the content drawn upon to produce the output they receive—giving them an incomplete understanding of the rhetorical context for the information they receive. This is a different but related sort of access problem to those described above. McKee and Porter (2020) acknowledge the inability of AIs to address a rhetorical context of communication between machines and between the machine and humans. Generated output from AI can be unpredictable and lacking considerations of ethics and rhetorical principles:

The ethics of human-machine writing requires of both humans and machines a deeper understanding of context and a commitment to being a good human, a good machine, and a good human-machine speaking well together. (McKee & Porter, p.111)

They highlight a key problem with both past technologies and generative AI: generative AI, like many new technologies we have encountered before, is rhetorical because of its dependence on users, speakers, or writers and has no ability to understand ethics and the rhetorical situation. McKee and Porter (2020) argue that AI like past technologies cannot and do not address rhetorical concerns, noting that Microsoft's Twitterbot, for instance, was made available to the public without any contextual knowledge “particularly, of what constitutes racism, sexism, homophobia and anti-semitism” (p.111), thus creating a communicator who lacked rhetorical knowledge or any way of considering it when generating responses initiated by a user. Users’ lack of access to the functioning principles and texts informing the output of the AI makes them lack agency and be forced to accept output informed by a repository of bigoted texts, as Byrd (2023) highlights. No training in prompt engineering (discussed below), which also presupposes access to instructional resources or informed teachers, can fully protect users and prevent encounters with content based upon the vast expanse of biased texts on which the AI was trained. As Byrd (2023), argues, the refusal of OpenAI to reveal details of their chatbot's architecture, supports the idea that “ChatGPT may not be an ethical tool for our purposes as writers and researchers” (p. 138).

In contrast, generative AI may promise some advances when it comes to providing accessibility in digital spaces for individuals who experience content differently or have issues with visual processing. In previous scholarship that, as a result, uses older types of approved terminologies, Browning (2014) expands on two models of disability: the medical model of disability and the social model of disability. She writes, “Many efforts at accommodating individuals with disabilities, though often well intentioned, coincide with a medical model of disability in that accommodations are simply added on to existing structures and systems” (p. 98–99). To provide better access and accessibility, Browning says, “Rather than simply retrofitting our universities, our classroom spaces, and our pedagogies, we must actively integrate disability, in thoughtful and critical ways, into all aspects of our teaching” (p. 99). Wood (2017) agrees, arguing that the basic conceptions of time that structure in-class writing and longer writing projects developed outside of class need to be rethought to assist all students and that the policies developed should be created with student input. Furthermore, Fox (2013) argues that focusing on disability studies in composition classrooms helps to highlight the mind-body connection embedded but often elided in the use and development of digital technologies and foreground the ways that universal design helps to increase accessibility for all users.

Henneborn (2023) notes that society has not done well when it comes to providing people with disabilities accommodations or acknowledging a disability that prohibits a user from participating in technologies of the workplace as she notes that the history of the workplace has not been thoughtful when addressing accessibility:

“We haven't done well as a society with the digital divide that exacerbates the barriers between persons with disabilities (as well as other marginalized communities) and others” (n. p.). As Kerschbaum (2013) highlights, even multimodal texts that are partially accessible by, for example, providing a transcript of a video, prove to be inaccessible overall because another part of their content, such as the images or navigation, are essential and yet not designed to be accessible and so restrict readers with particular disabilities from using the content.

In contrast to access, which is complicated by generative AI, these new technologies may have some potential to support accessibility in innovative ways. Businesses, corporations, and other institutions are ready for generative AI and the emergence of ChatGPT to provide hope for their employees who identify with disabilities. At its 2023 Ability Summit, Microsoft outlined its plans to employ AI tools in its products, such as Office 365, to alter contrast for some users and generate descriptions of images on demand for others (Cuevas, 2023). Henneborn (2023) argues that generative AI can address some accessibility challenges by creating “inclusive interfaces.” Tools like keyboard navigation, alternative text, voice-enabled interface/speech-to-text, text/image-to-speech, color contrast, dyslexia-friendly fonts, and clear language are all considered basic requirements for inclusive interfaces and appropriate accessibility and all can be enhanced by AI. Henneborn (2023) offers a few examples of such tools: “For instance, Google’s Dialogflow has built-in integration with Google Cloud Speech-to-Text API, allowing developers to create chatbots that support voice-enabled input” (n.p.). Dialogflow CX is the “advanced” edition and Dialogflow ES is the standard. This tool allows individuals to create chatbots and/or voicebots, and while this tool is not free, there appears to be a free trial (https://cloud.google.com/dialogflow). Another AI-powered tool is Be My Eyes or what may also be referred to as Be My AI (https://www.bemyeyes.com), which is visual assistance for users with low visibility, and this particular tool is used by a number of large corporations like Verizon, P&G, Google, and Microsoft, to name a few. Generative AI also seems to have the ability to aid users who experience dyslexia by supporting additional add-ons or plug-ins. For example, Dyslexie Font is a plug-in designed to address dyslexia by making reading and understanding easier (https://www.dyslexiefont.com). While many of these tools to assist with accessibility may currently be free, we know from the development of past technologies that monetization is almost inevitable. Support for accessibility will inevitably collide with issues of access when commercial enterprises seek to make money from the developing technologies. Even ChatGPT currently has a “freemium” model where users can use ChatGPT for free, but if one wishes to extend its use to other tools or to a “premium” version then one would need to pay for an upgrade.

More information about the many effects of generative AI on access and accessibility will be known as we continue to deploy these systems in more contexts. As the technology develops, we need to draw upon the lessons from our scholarly tradition that tell us to interrogate the rhetorical context in which these technologies operate and examine the access users have to their architecture and to the full range of affordances that enable creative and innovative composition and use. The next section focuses on some pedagogical approaches to these issues through examining how scholars have proposed interacting with other technologies, such as Wikipedia.

Masking Mediation

Throughout the history of computers and writing and digital rhetoric, scholars have grappled with and embraced disruptive information and communication technologies and the challenges they pose to their contemporary research and composing practices. As Bolter and Grusin (1999) and Bolter and Gromala (2003) highlight, new technologies often obscure their mediative practices, promising to provide direct, unmediated access to information, entertainment, and production. Any user’s understanding and awareness of a digital technology's process and depth of mediation—the degree to which the technology draws attention to its means of production, to how it provides access to content or delivers result&mash;is partly determined by users’ prior experience with similar technologies, proficiency with using new technologies, and curiosity about how the technologies work. Communication and information technologies gain power in part by obscuring the degree of mediation and promising users that they will provide direct access to knowledge and information without requiring users to comprehend how the content is developed and delivered. One way to accomplish this transparency is to use the structural conventions of established communication technologies to achieve transparency (Bolter & Grusin, 1999; Hocks, 2003).

The way that numerous scholars in computers and writing and digital rhetoric have approached teaching with Wikipedia since its inception in 2001 provides a productive model for navigating and composing with technologies that downplay and obscure their processes of mediation like generative AI; as a result, we will explore that model throughout this section. Like Wikipedia, generative AI can be approached as a transparent window into vast amounts of easily accessible knowledge without the user’s need to understand the technical mechanisms facilitating its production. Both Wikipedia and generative AI proffer information that appears reasonable and professional (Lancaster, 2023), persuading users, like our students, to accept the output at face value without critically examining its veracity.

To combat the seductive transparency of technologies like Wikipedia, scholars developed structured inquiry and assignments designed to transform student users from passive consumers into critical producers of content. In the case of Wikipedia, this requires individuals to work behind the surface of the encyclopedic content as displayed to understand the layers contributing to and supporting its production, including the organizational structures and policies, the debates within the Wikipedia community, and the technical knowhow needed to contribute correctly formatted content. As Reilly (2011) explains in her article about teaching Wikipedia as a “complex discourse community and multi-layered, knowledge-making experiment,” empowering student users to become critical producers of content with Wikipedia requires that they literally look behind article layer of the text to interact with and contribute to the layers of the Wiki technology that allow them to engage in conversation with other contributors (Discussion tab), edit the content of the article (Editing tab), and examine the complete history of changes to the text (History tab).

Based on their analysis of large-scale (6000 students) survey research by the Wiki Education Foundation, Vetter et al. (2019) recommend best practices for Wikipedia writing assignments that include making the assignments “extended and substantial” to allow students to “learn about values, processes, and conventions of Wikipedia” (p. 62). Vetter et al. (2019) also recommend critical analyses of Wikipedia articles before contributing in order to develop critical thinking around how Wikipedia is designed, how content is supported, and how sources are cited. To design such opportunities for analyzing and contributing to Wikipedia, instructors need professional development to learn the content and technological intricacies of the platform (Sura, 2015).

In addition to analyzing the content and contributing to it, McDowell and Vetter (2020) argue that Wikipedia’s very practices and policies, particularly those requiring citation and verification for information, serve a pedagogical purpose and can be harnessed to aid students to develop information literacies related to the legitimacy of online information. The polices prompt students to learn how to analyze information for its veracity—not rely on others (McDowell & Vetter, 2020; see also Azar, 2023). In addition, Wikipedia has the benefit of being a community governed by the collective (run by a nonprofit): new participants can learn to navigate the norms (McDowell & Vetter, 2020) and work with other contributors in asynchronous but interactive manner. As Purdy and Walker (2012) explain, contributing to wiki-based compositions foreground the importance of dialogue for knowledge production. McDowell and Vetter (2020) argue that Wikipedia's policies requiring verification, a neutral tone, collaboration, and citations educate new users and enlist them to maintain the legitimacy of content on the site and participate in removing or questioning that which is not supported by sources and verifiable knowledge (Azar, 2023)—through such policies, contributors to Wikipedia develop critical digital and information literacies that they can employ in other contexts. Finally, Wikipedia uses community-based policies “to reconstruct more traditional models of authority” that support the legitimacy and veracity of the content; and Wikipedia is transparent about its purpose and intentions unlike most other (commercial) sites and apps online (McDowell and Vetter, 2020).

Many of the lessons outlined above related to working with Wikipedia and exposing its processes of content mediation to conscious examination and interrogation can be adapted to help our students work with and critically analyze the output of generative AI. This process can begin by examining how generative AIs produce content. Byrd (2023), Lancaster (2023), and many others help to demystify how AI technologies like ChatGPT work from a technical perspective for instructors and students. As was the case when teaching with Wikipedia and other new technologies, highlighting the technology’s processes of mediation entails gaining a basic understanding of the technical specifications that power it. Students need to learn that ChatGPT, for example, is a large language model (LLMs) meaning that it has been trained to produce new language modeled on the texts it has processed and was asked to generate (Byrd, 2023; Lancaster, 2023). As Byrd (2023) clearly explains, “[LLMs] have really created mathematical formulas to predict the next token in a string of words from analyzing patters of human language. They learn a form of language, but do not understand the implicit meaning behind it” (p. 136). As a result, AI can produce false information when the corpus it uses does not contain the accurate content (Byrd, 2023; Cardon, et al., 2023; Lancaster, 2023; Hatem, 2023). Understanding these technical processes can help students to approach the output from AIs more critically and skeptically, just as they are taught to do in relation to Wikipedia. Such instruction provides inoculation, demystifying the output and opening it to interrogation.

Students also need to be taught the protocols for productive use of generative AI as they do with Wikipedia. Prompt engineering is the process of iteratively developing instructions and queries to submit to the AI to garner superior output (Korzynski et al., 2023; Lo, 2023). As Korzynski et al. (2023) emphasize, prompt engineering is a human language process requiring collaboration with AI. Just as students must learn how to structure and tag their Wikipedia articles to meet the genre specifications approved of by the community, students also have to structure queries to gain the best results from the AI. They also need to dialogue with the AI as they did with other contributors in the Talk tab in Wikipedia to participate fully in a successful collaboration. The obvious difference is that when writing for Wikipedia, students collaborate with other users, not with AI. A number of scholars have developed frameworks to guide prompt engineering. For example, Lo (2023) outlines the CLEAR Framework: prompts to AIs should be concise, logical, explicit, adaptive, and reflective. Importantly, this framework emphasizes that success is produced iteratively and contextually in response to the output of the AI and purpose of use (Lo, 2023). Korzynski et al. (2023) review a range of other similar approaches to prompt engineering; they outline the essential elements of useful prompts, including the context or role, the instruction or task, the relevant data, and the genre or form for output (p. 30). Such discussions of prompt engineering emphasize that scholars, instructors, and students can learn to productively collaborate with generative AI, as they do with Wikipedia and its corresponding community, and overcome hurdles to engaging with it productively.

Just as scholars recommend critically analyzing Wikipedia articles prior to using content from and contributing to Wikipedia, so must students learn to do the same with the output of generative AIs. Lancaster (2023) recommends finding sources to corroborate and support the veracity of content generated by AIs. Scholars already report developing assignments that ask students to investigate the veracity and usefulness of text produced by an AI (Byrd, 2023). As noted in the previous section, once students understand that the quality of output is driven in part by the quality of the input, they may gain agency and confidence to critique the resulting information produced by that AI.

Some of our lessons from teaching with other technologies like Wikipedia do not apply to generative AIs. For example, as discussed above, the mechanisms of mediation by generative AI are often proprietary, making it impossible to fully comprehend how the technology delivers content as we can with Wikipedia. AI companies are commercial enterprises unlike Wikipedia, which is a nonprofit. In response, Byrd (2023) recommends using open-access LLMs instead to produce more ethically and fully transparent content. Finally, generative AI’s rapid evolution may eventually make detecting its output as the work of machine-generated mediated content less possible by human readers. Lancaster (2023) proposes a process of adding watermarks to AI-generated content but acknowledges the potential futility of that approach. Additionally, this approach would require coordination and cooperation with corporate entities that, as noted above, maintain control over their technologies and standards and have little to gain from revealing their proprietary information and demystifying the power of their chatbots to magically anticipate users' needs and surprise them with content they can use as their own.

As the above discussion reflects, the work of scholars and instructors with previous technologies like Wikipedia can provide insights about what questions to ask and how to advocate for restrictions and guidelines to protect students and the public in their work with generative AIs. Developing educational policies and best practices around writing with and using digital content, like Wikipedia, was necessary, and invested scholars in our profession need to do the same for AI.

Conclusion

Our chapter highlights several of the major themes of scholarship related to computers and writing and digital composing and connected that legacy of scholarship to current issues related to writing with generative AI. Numerous additional themes are relevant to this discussion, including issues of privacy and surveillance, which Johnson also highlights in his video response (above) to our question. Our scholarly tradition of interrogating privacy and surveillance in digital spaces is robust and growing and includes numerous recent articles, special issues, and books including a publication in Kairos based on a Town Hall at Computers and Writing in 2015 (Beck et al., 2016), a collection of essays edited by Beck and Hutchinson Campos (2020), and a special issue of Computers and Composition edited by Hutchinson Campos and Novotny (2021). The work in these publications highlights the largely invisible intrusions into privacy and ubiquitous surveillance that individuals encounter when learning, working, and playing in digital environments. They also emphasize the lengths to which the corporations and developers of the technologies will go to hide those risks and keep the infrastructure and algorithms powering the technologies invisible to users and unavailable for scrutiny as proprietary information. This line of scholarship provides guidance for examining the potential threats to privacy posed by generative AI, which, given its newness, are somewhat murky and speculative. Our scholarly tradition warns us to be skeptical and wary, but we, like government researchers (Busch, 2023) and data privacy enterprises (Securiti Research Team, 2023), must extrapolate the potential harms based on what is known about how generative AIs function and how privacy has been compromised by similar technologies. For example, a review of basic privacy scholarships highlights that any corpus of data is vulnerable to hacking, exploitation, or accidental leakage, and AI is no different. The vast stores of data collected to train a Chatbot like ChatGPT have already been compromised and will continue to make users and those whose data has been secretly collected to train AIs vulnerable to the release of personal medical, financial, social, and other information (Securiti Research Team, 2023). As Morris (2020) explains, users of chatbots who have rare disabilities may be at even greater risk of privacy violations when using the AI to seek information to learn about their conditions and seek treatment options. As she notes, “past incidents of re-identification of individuals from anonymized datasets…indicate the difficulty of truly anonymizing data” (Morris, 2020, p, 36). In 2023, not only do individuals with disabilities or illness risk exposure, but so do women seeking reproductive healthcare options and individuals from other marginalized communities, such as youth who identify as trans. Our scholarly tradition helps us to identify risks related to privacy and alert our students, colleagues, and others, but, unfortunately, provides few tangible solutions to these and other significant problems outlined above related to generative AI and future technologies to follow.

That the problems are intractable and yet unknown cannot cause us to give up. As Johnson’s video highlights, as teachers we cannot surrender because we must help our students to resist the apathy that can result from the enormity of the task of negotiating the rapid and risky technological changes that must be faced pedagogically, organizationally, and socially. The scholarship explored above reflects that our field has faced such upheaval before and found ways to work with and through the technologies, whatever form they take. We take inspiration and instruction from that history.

References