Composing with AI

Imagine You Have a Super-Smart Friend Named Chat GPT

ChatGPT may appear to be capable of nearly any information retrieval or writing task, but it does not always like to comply with our requests. That’s good if it’s refusing to provide dangerous or harmful material to people and technologies seeking to damage others. But, ChatGPT refuses to answer harmless requests, too, especially if that request is for information or processes about itself. When I asked ChatGPT-4 how its interface was designed, it claimed not to know: “As of my last update in January 2022, OpenAI had not released specific details about the design or implementation specifics of GPT-4’s interface. My understanding is based on general principles and practices followed by OpenAI and other companies involved in AI and deep learning” (ChatGPT, October 24, 2023). This hedging maneuver was present in nearly every GPT response I received. These responses avoided responsibility for the accuracy and relevance of the responses while continuing to provide content as a response, a logic that sounded like an answer but didn’t provide the information the user requested, a form of circular logic akin to generative AI gaslighting.

With free web-based access to ChatGPT-3.5, the conversational capabilities of its interface design have been the key to gaining widespread use by the general non-programming public. Rather than focus on these inconsistencies in performance and accuracy, ChatGPT wants users to “imagine you have a super-smart friend named ‘ChatGPT-4’” (ChatGPT, September 21, 2023). Its interface and circulation transmit its “embedded values” and steer our “more or less normative” engagements with it into more standard interactions with how it should be used (Stanfill, 2015). ChatGPT’s interface provides an opportunity to analyze its technological and social aspects in their materiality. GPT’s circulatory power attaches to its interface and other entities as it carries dominant narratives about AI’s superpowers along with innocuous, human-like metaphors and other rhetorical frames (Burke, 1969) that ask us to identify with ChatGPT as we would with another human. At the same time, its interface asks us to accept the superior understanding and power of its technology. All are attempts to gain our trust and submission; however, we can foster the critical acumen to understand whether this friend’s companionship improves us—does it lead us in healthy, transparent directions that allow us to thrive and grow with agency over our engagement with information and how we write about it? Or, as Selfe and Selfe warned nearly thirty years ago, will we and our students submit to reproducing and circulating “knowledge” and the “small but continuous gestures of domination and colonialism?” (p. 486).

In the UX/UI field, Jakob Nielsen (1992) led the charge for heuristic interface evaluation as “a method for finding usability problems” (p. 373). Nielsen defined usability problems as “any aspect of the user interface which might create difficulties for the user with respect to an important usability indicator (such as: ease of understanding and learning how to operate, ease of use, time to complete the task, or subjective user satisfaction)” (Pribeanu, 2017, p. 32).To identify usability problems, Nielsen (1992) advocated for “a small set of evaluators [to] examine the interface and judge its compliance with recognized usability principles” (p. 373). These principles of usability were defined by Nielsen and Molich (1990), and Nielsen and others have continued to revise shared principles of usability as well as heuristic and ergonomic evaluation methods (Pribeanu, 2017, Table 1). Usability heuristic and ergonomic approaches, often combined, are used within UX/UI to achieve “better usability” and human-centered design. Pribeanu (2017) explained that, “[u]ser guidance is a general ergonomic criterion that concerns the means to inform, orient, and guide the users throughout the interaction with the computer [20]. Suitable user guidance has positive effects on the ease of use. The user guidance includes the following four heuristics: prompting, feedback, information architecture, and grouping/distinction” (p. 33). With these ergonomic considerations, usability has become about “guiding” intended users toward the proper uses that, again, programmers and designers imagine. In the black box of GPT and other AI, typical interface elements for user agency are mostly missing and agency is situated more firmly with the interface and its hidden processes.

In the 2023 OpenBootcamp UX/UI Challenge, Nielson’s heuristic was used to evaluate and improve upon ChatGPT’s usability. Writing about the experience on Medium, Athul Anil, concluded that ChatGPT’s interface complies with all of Nielsen’s Usability Principles except for a “minor usability problem related to aesthetic and minimalistic design. The interface can be improved by streamlining the chat interface, removing unnecessary elements, and utilizing space effectively” (Anil, 2023, Conclusion, bold in original). Anil passed ChatGPT on every other interface heuristic principle of interface design, finding no major issues with usability. Anil’s evaluation asks for even more minimal aesthetics and features, reducing visual elements while also reducing the functions and options available to users. More importantly, Anil passes ChatGPT on all nine remaining principles, demonstrating that at least in this UI/UX circle, ChatGPT’s interface is a nearly flawless design. But what and for whom was it designed for?

Your Super Smart Friend, ChatGPT

With its ability to understand human phrasing through NLP, we should be able to prompt GPT to answer nearly any query, even questions about the source material for its responses. As we’ve seen, though, ChatGPT’s responses to requests for transparency are hedging at best. When I asked it GPT-4 to explain how it works to someone with only basic computer knowledge, it stated, “Suppose your friend spent years reading from millions of books, articles, and other texts. They didn’t memorize everything, but they learned patterns, like how sentences are structured, what kind of answers people give to certain questions, and lots of information about the world” (ChatGPT, October 19, 2023). ChatGPT-4 said it has a “brain-like structure,” with digital versions of neurons called neural networks. “After reading all those texts, your friend [GPT-4] practiced guessing what word comes next in a sentence” so many times until it “got really good at understanding language and predicting what should come next.” Now, it doesn’t have to find or locate the answer in any specific texts: “Instead, they [GPT] think about all the patterns they’ve learned and try to generate a response that makes the most sense based on what they know. It’s like they’re completing a sentence, but the sentence is your answer” (ChatGPT, October 19, 2023). ChatGPT wants to complete your sentences for you. It values the ability to remake knowledge in its own image and attempts to make knowledge production invisible and shaped by machine pattern finding and regurgitation. This kind of knowledge production doesn’t require humans to perform research, read, and consider the sources of their information, or come up with any new ideas to extend or complicate other ideas, it only requires that we talk to ChatGPT. As we input our queries, we’re giving its LLM and company more data to improve its accuracy, arguably more valuable than the responses it generates for us. “Just like any friend, ChatGPT-4 isn’t perfect. Sometimes they might get things wrong or not fully understand a question. They’re just trying their best based on what they’ve learned” (ChatGPT, October 19, 2023). In this response snippet, GPT framed itself as a flawed human, and mistakes, while not a good thing, are generally endearingly human. In the next sentence, GPT wanted us to, instead, see it as a neutral technology that can’t be maligned like a human: “This friend doesn’t have feelings, emotions, or consciousness. They’re just a tool, like a very advanced calculator for language.” These appeals to human qualities and emotions like friendship and trust are ChatGPT’s “affective flirtations.” Affective flirtations position ChatGPT as part of our shared humanity while slipping its more predatory values and behaviors under our awareness, softening our defenses and “norming us” to GPT’s values.

We can accept these UX/UI heuristic evaluations of GPT’s interface and the explanations of its processes. We can accept its appeals for our friendship, which entails tolerance for its errors and shortcomings. Or we can develop our own writing studies interface heuristics. In our classrooms, students can learn heuristic evaluation while also building Critical AI literacies. Heuristic evaluations should reflect not only design principles, but values that reflect, accept, and amplify pluralistic social, political, learning/writing environments and relations. Heuristics can help students exercise and better understand the limits of their agency with GPT, while producing a nuanced evaluation of interface design. As writing studies scholars have pointed out, the influence of technologies is not confined to particular interactions—interfaces function as a “circulatory, world-making process” (Jones, 2021). As Grabill (2004) pointed out, “to ignore infrastructure, then, is to miss key moments when its meaning and value become stabilized (if even for a moment), and therefore to miss moments when possibilities and identities are established” (p. 464).