Composing with AI

Interface Heuristic Development

In “Articulating Methodology as Praxis,” Sullivan and Porter advocate for a heuristic rather than rules-based approach to writing studies methodology. Connecting with feminist critical approaches, they described heuristic approaches to research as “dynamic and relational focus rather than a static one” (p. 59). To scaffold enduring AI critical literacies, heuristic approaches to teaching and learning with GPT will attune to the “dynamic and relational aspects” of our classrooms, including the vulnerabilities of our students to the “violence of literacy” (Selfe & Selfe, 1994). This would entail knowing our students more deeply than their requests for accommodations and any visible marginalities. This heuristic attunement to GPT and other technologies is part of doing and paying attention to “everyday rhetoric” (Grabill, 458). Grabill defined this as “a commitment to the mundane, the obvious, and the unstudied as intellectually and epistemologically equivalent to political and ethical commitments to ‘difference,’ to the ‘oppressed,’ to the technopoor’ and other issues of social justice” (p. 465). Our field has advocated for heuristic interface development in community, corporate, and even in AI Generative LLM tech development circles. Heuristic development and evaluation should be adaptable to changes in public knowledge of such technology, the settings in which its used, the positionalities of communities, and the evolving iterations of interface design and its underlying infrastructure.

Hocutt, Ranade, and Verhulsdonck (2022) examined Meena, Google’s (now retired) conversational chatbot. Unlike ChatGPT’s black-boxed infrastructure and processes, Meena’s have been made publicly available on GitHub (Hocutt et al, p. 119). Ultimately, Hocutt et al. determined that these contextual chatbots could respond with more personalized responses, or “localized microcontent that can help users in contextual ways,” but not without humans getting in the loop (p. 125). They identified four “insertion points” for technical communicators to influence chatbot development and performance and concluded that future technical communicators should be supported in their educational programs by more explicit and rigorous training in data and AI. They remind us that a better future with AI is not guaranteed: “Without intervention by technical communicators, ML [machine learning] will continue to train on biased discourse, continuing to marginalize those already marginalized by those discourses, or foster black-boxed processes that perpetuate discrimination or lack of explanation or transparency of decision-making to users” (p. 128). While our field should directly influence technological designs and functions, we’d have to find a route to overcoming powerfully conflicting values that drive surveillance capitalism and education, respectively. Pragmatically, we could consider our classrooms as the foundational space for AI critical literacy development. Although technologies have always been implicated in systems of power, ChatGPT ushered in a new era of super-human circulatory world-making power of AI, or at least the perception of that power. As I recently observed through activities I’ll describe in the next sections, our college writers are especially vulnerable to the productive power (Stanfill 2015) of GPT’s interface, and their generations will bear the heaviest burden of its consequences.

As Corinne Jones (2021) and others have observed, human agency is especially difficult to claim since interface design alludes our conscious attention—so much so, that “they often seem neutral and become invisible” (p. 2). That invisibility allows users to look past interface before we’re aware we’re doing so, perpetuating their power dynamics and dominant logics in our actions and the uptake of its information. Jones enumerated three ways interfaces produce power and circulate dominant values. Interfaces have an: (1) “affect on how people circulate information, (2) affect what people know by shaping what information or content they see through circulated materials, and (3) affect who circulates materials as they shape how people know themselves as circulatory actors” (p. 4). Jones adopted and modified Stanfill’s (2015) Discourse Interface Analysis framework to seek the “unconcealment” identified by Selfe and Selfe (1994). Stanfill observed interface heuristics can reveal “how social forces that produce technology manifest in web deployment” (p. 1060). Applying Foucualt’s theory of productive power, Stanfill examined interface affordances as producing (rather than preventing or oppressing) “correct uses” and norms of user behavior. The correct uses and norms are not wholly fixed and determined within its interface, but they’re not easy to resist since these productive affordances are the designed protocol for interactions between humans and computers. Stanfill (2015) categorized interface design features by functional, cognitive, and sensory/aesthetic affordances. These affordances appear as interactive elements like menu items, dropdowns, and text boxes (Functional); messages to the user about what’s can and should be done (Cognitive) and visual/tactile features that attract attention, increase pleasure, and create brand image (Sensory/Aesthetic). These interface affordances are “sites of Foucauldian productive power because they encourage certain practices while hindering others” (Jones, p. 3). While affordances seek to determine users’ appropriate interactions, those interactions are also the primary gateway for the user to access and potentially benefit from the infrastructure and information behind them. These affordances “create norms and make certain practices and positions seem commonplace” (Jones, p. 3). Jones distinguished between industry-oriented heuristics like Nielsen’s as being focused on site designers’ perceptions of users’ needs while the Stanfill (2015) DIA heuristic “is focused on the effects of those affordances in terms of productive power (p. 1062) and it does not assume that site users know what they want to do on a site” (Jones, p. 4). For Stanfill and Jones, interfaces are particularly problematic for the ways in which they position users to the underlying computer infrastructure, processes, and information and “reflect hegemonic cultural norms” (Jones, p. 3). Jones examined U.S. Chamber of Commerce website interfaces related to government COVID-19 communication and public tweets about interacting with these website interfaces, and her conclusions are particularly resonant with advanced conversational chat interfaces like ChatGPT 3.5/4: “[B]oth interfaces produced a normative practice of circulating pre-existing content. However, they also functionally normalized and produced prosumers in a surveillance economy, thereby alienating people from their labor and perpetuating power” (p. 5). Jones’s conclusions could’ve been describing generative writing technologies—they circulate pre-existing content, they normalize users into accepting and consuming their functionality and products, and they take information and other types of data from their users to improve a product or its profit-making potential without any credit, financial or intellectual, to the original knowledge producers. The following section will briefly adapt principles from interface theory and heuristic development by Selfe and Selfe (1994), Grabill (2003), Stanfill (2014), Jones (2021) to a brief analysis of ChatGPT 3.5/4 interface.

Interfacing ChatGPT Heuristically

Like Stanfill’s DIA heuristic, building Critical AI literacies directs our attention to formal structures of interface design and advance a materialist take on AI technologies as they show up in particular instantiations like ChatGPT’s web-based chat interface. Heuristic analysis of formal features goes beyond the functionalities of such designs to ask, “what is foregrounded, how it is explained, and how technically possible uses become more or less normative through productive constraint” (p. 1062, emphasis in orig.). These productive constraints—the designed elements of the interface—are the means by which users, our students, are “normed” to ChatGPT’s norms and values.

In the following sections, I’ll provide a ChatGPT interface analysis influenced by students in my Fall 2023 section of Introduction to Document Design at East Carolina University. This course fulfills a writing-intensive requirement for all majors, and it tends to combine majors from English and Art and Design with majors in health sciences, computer science, business, and other non-liberal arts majors looking to pick up the writing credit. Since I’m writing about heuristic analysis in writing classrooms, I wanted to include their perspectives and, to the extent possible, provide a supportive, participatory environment for interface heuristic development and analysis with GPT.