Agents of Change: Decoding Human Voices in Large Language Models

Unlike online articles and journals, which attribute bibliographic information where necessary, ChatGPT does not reliably display its sources, effectively framing itself as the sole content creator of the outputs it generates. Even the language we use when talking about AI-generated content, such as “ChatGPT said” and “ChatGPT generated,” strengthens a default perception of its position as an author. This erroneous attribution erases the work of the human authors who provide the chatbot with the necessary inputs to formulate a response—not only the users who interact with the platform, but also the engineers who code and train the algorithm in preparation for its release to the public. Thinking of engineers as authors in the creation of algorithms is crucial in closing the accountability gap, and understanding the relationship between many engineering tasks and the field of rhetoric is one way to encourage this perspective.

Although engineers may not be the sole owners of a language model, their involvement in the chain of operation requires a sense of collaborative accountability.

The field of engineering and the field of rhetoric are typically categorized and perceived as mutually exclusive, with one having to do with “observation and logic” and the other with “symbols and emotions” [5]. In reality, the two are deeply interlinked as engineering, specifically the act of coding, is highly rhetorical in nature. Writing code that supports a particular design goal is analogous to writing arguments that support a thesis. Rhetorical theories that argue that language can offer a window into reality “greatly underestimate the role of human interaction” [6]. In actuality, language and datasets shape the world we live in, creating what we see as reality [6]. Consider how the very process of data collection involves selecting a data pool which may, either consciously or unconsciously, exclude certain members of the population. Engaging with datasets as though they constitute objective reflections of reality can be incredibly dangerous, leading engineers to fail in recognizing the implicit biases within their algorithms “such as the underrepresentation of minoritized groups and stereotype-aligned correlations” [7]. If engineers can shift their perspective on their own status and position within workflows—and view themselves not as technicians engaging with “transparent and neutral vehicles for information”—but rather as rhetoricians who are engaging with highly rhetorical tools of knowledge, they can better identify and eliminate harmful biases from their work [8].

That being said, from a recognition perspective, authorship in the field of literature is vastly different from engineering: while writers are clearly attributed for their work on the front covers of books, engineering contributions are nowhere near as public-facing. With multiple authors often working on a single algorithm, “the number of operations in a causal chain may be so extensive or convoluted that it is practically impossible to disentangle individual contributions to final outcomes” [9]. As authorship becomes distributed among multiple agents, some human and others nonhuman, agency becomes fragmented as well. This division of labor, known as the problem of “many hands,” distances engineers from a sense of full ownership over the products of their work, making it easier to relieve themselves of ethical responsibilities [9]. After all, if models themselves are perceived as authors upon deployment—and behind-the-scenes, there are an undefinable number of contributors involved in the design process—then engineers are at high-risk for feeling more like ghostwriters than authors.

This is where data work recognition processes, such as documentation, become integral in supporting accountability efforts. By “outlining the nature of an individual or team’s contribution to an overall system” [7], engineers can begin to write themselves back into their work as authors with the capacity to make substantial edits and alterations. Beyond documentation, following a human-in-the-loop (HITL) process of writing, “in which human collaborators retain the majority of the rhetorical load” [10], allows engineers to see themselves as active contributors—even if some of this writing is also done by machines. Although engineers may not be the sole owners of a language model, their involvement in the chain of operation requires a sense of collaborative accountability. As writers, as well as rhetoricians, engineers have the power and duty to rewrite unethical arguments in code.

This movement toward collective accountability is particularly important when it comes to voicing ethical concerns since, under constant pressure to remain silent and complicit, individual engineers are often limited in their ability to advocate. Even when workers choose to “write, speak, or act” in response to wrongdoing, they are not always “recognized or heeded by others” [3]. This becomes especially clear when observing the treatment of whistleblowers [11]. Rooted in parrhesia, meaning free and truthful speech [12], whistleblowing is a rhetorical practice where speakers aim to move companies and executives to pursue corrective actions that address internal wrongdoings. From the beginning of their education, engineers are taught the importance of whistleblowing; objects like the iron ring—which graduating Canadians receive at the Calling of an Engineer ceremony—symbolize and remind students of their lifelong commitment to responsible innovation. However, maintaining a strong commitment to responsible innovation can be a challenge in corporate settings, where employees “frequently [find] themselves trying to reconcile contradictions within and among multiple domains of accountability” [13]. As professionals, engineers have a duty to protect public health, safety, and welfare since their activities have a “direct and vital impact on the quality of life for all people” [4]. Yet, as employees, they also have a duty to their employer. These potentially competing responsibilities become increasingly difficult to navigate when considering how organizations mistreat whistleblowers, often forcing employees to weigh their own needs for economic security alongside threats to public safety.

The rapid advancement of AI technologies has only exacerbated whistleblower mistreatment.

The rapid advancement of AI technologies has only exacerbated whistleblower mistreatment. For example, OpenAI, the company responsible for creating ChatGPT, was recently accused of stripping whistleblower rights with illegally restrictive NDAs. As Ha reports, the agreements in the NDAs “prohibited and discouraged both employees and investors from communicating with the [U.S. Securities and Exchange Commission] (SEC) concerning security violations, forced employees to waive their rights to whistleblower incentives and compensation, and required employees to notify the company of communication with government regulators” [14]. Ironically, these “agreements” encapsulate the very reason why whistleblowing is necessary in holding big tech accountable. By prioritizing financial growth and algorithms above human workers, corporate systems unevenly distribute agency across organizations, such that the whistleblower “is always less powerful” than those toward whom the whistle is directed [15]. However, with the support and pressure of external actors, organizations that would typically silence or ignore employee voices are forced to listen.

Pushing back against corporations’ efforts to avoid responsibility becomes a matter of finding and building out a community. While individual efforts to whistleblow may be challenging, adopting an ethos of collective accountability—one that is grounded by a shared identity as rhetoricians—can unite workers together and encourage social movement organizing. According to the framing that C. Tilly provides, social movements are a “sustained challenge to powerholders in the name of a population living under the jurisdiction of those powerholders by means of repeated displays of that population’s numbers, commitment, unity, and worthiness” [16]. Made up of employees “living under the jurisdiction” of company policies, labor unions are a form of powerful social movement that use rhetorical strategies to displace power out of the hands of executives and into the hands of workers. For example, following the wrongful termination of Timnit Gebru, more than 200 workers at Alphabet, Google’s parent company, unionized to fight a wide range of issues at the company, such as Project Maven, a Pentagon partnership for drone footage-analyzing AI [17]. This type of mobilizing is only possible when workers see themselves as rhetorical authors with the power and public responsibility to influence their conditions of work.

In reality, engineering and ChatGPT are highly rhetorical in nature, as both are subject to human error and bias.

Union members have since staged employee walkouts, coordinated matching Zoom backgrounds [19], distributed petitions, planned public rallies, and even resigned from their positions, all in the name of protest [20]. While the Alphabet Workers Union, pictured in Figure 1, was originally organized by software engineers, it is now a wall-to-wall union, meaning that it is open to all employees, including temporary workers and contractors. This inclusive method of social movement organizing ensures that every employee— not just white-collar engineers—has access to a rhetorical megaphone. For workers who are typically silenced and invisibilized, access to community support is a lifeline since, for many workers, their livelihoods rest on their job security. An intersectional, wall-to-wall approach in unionizing is therefore necessary to avoid reinscribing the same harms committed by corporations themselves.

As these examples illustrate, just as accountability and authorship are collective responsibilities, so is the movement toward responsible innovation. Unions demonstrate to corporations that individualism is not a permanent or sustainable state of living; they remind corporations that agency is not a finite resource—when in community with others, it can grow exponentially.

The rapid development of AI has altered the realm of current realities and future possibilities in the field of technology. However, as with any innovation, there are risks involved when proper systems and safeguards are not in place, especially during the early stages of design. For one, as objects of corporate monetization and productivity, LLMs such as ChatGPT have restricted our sense of human contributors’ agency. Since ChatGPT is not designed to adequately disclose its sources of information, it is positioned as an independent content creator. As a content creator with decades of online memory, ChatGPT appears to operate within the realm of fact and logic. By proxy, so do the engineers working behind the scenes to code and train the chatbot. In reality, engineering and ChatGPT are highly rhetorical in nature, as both are subject to human error and bias.

Moving engineering outside of the realm of ’objectivity’ and into the realm of rhetoric allows workers to view themselves as writers and take on a more active role in the coding process. This shift in perception, coupled with tangible tools for co-authorship such as HITL writing and documentation processes, can help engineers see themselves in the work they produce, thereby bridging the gap between engineers and algorithms. If engineers have a deep sense of accountability for their work, they may be more inclined to become advocates for responsible innovation.

That being said, even if engineers recognize that they have a duty to innovate responsibly for public safety, the conditions of their work environment are not always conducive to these ethical responsibilities; corporations repeatedly silence employees who attempt to broadcast internal wrongdoing. In such cases, whistleblowers, framed as rhetorical truth-tellers, must turn toward community-building practices to effectively pressure executives in pursuing corrective actions. After all, rhetoric, like engineering, has never been an individual pursuit; it is rooted in community and collaboration with others. As AI systems continue to develop and advance, human contributors must not be erased from the design process. When emphasis is placed on the writers behind AI, people—engineers and nonengineers alike—can regain agency and ownership of their work.