Extractive Artificial Intelligence and Its Challenge to Technical Communication

Abstract

Mainstream artificial intelligence (AI) is an extractive industry that exploits both humans and nonhumans. The extractive underpinning of mainstream AI systems means that technical communicators must be careful when advocating for accessibility and inclusivity in AI because those efforts may expose marginalized groups to further exploitation. Extractive AI also necessitates that technical communicators reconsider how their own discipline may be complicit in the damaging logics and practices of extraction.

Artificial intelligence (AI) presents an urgent challenge to a discipline such as technical and professional communication (TPC) that is still coming to grips with its own complicated relationship with power. The social justice turn in TPC has played a vital role in uncovering and resisting the politics of domination that underpin some of the field's most deeply rooted ideologies (see, e.g., Jones, 2016; Jones et al., 2016; McKoy et al., 2022; Shelton & Warren-Riley, 2023). Through this social justice lens, scholars in TPC and neighboring fields have already critiqued generative AI systems for perpetuating “linguistic injustice” and “unequal access” (MLA-CCCC, 2023, p. 7; see, e.g., Bjork, 2023a, 2023b; Byrd, 2023; Carradini, 2024; Johnson, 2023; Morrison, 2023; Owusu-Ansah, 2023; Vee, 2023; Wise et al., 2024). Typically, activists and educators respond to such inequities by advocating for inclusion and accessibility. Yet as I will show, when these AI systems are seen through the lens of extraction, the practices of accessibility and inclusion risk exposing marginalized communities to further exploitation.

In this commentary, I define “extractive AI” and use the example of Otter.ai to illustrate how viewing mainstream AI systems through the lens of extraction undercuts social justice efforts in TPC. I then provide an alternative example of an AI system developed by Indigenous experts that operates according to different, nonextractive ideologies. I close by calling for TPC to take up the lens of extraction more often and, using that lens, imagine more just ways forward for our field. There are, I argue, ways to design AI systems according to nonextractive ideologies, and TPC practitioners have both the skills and a disciplinary responsibility to make substantial contributions to these efforts.

Defining Extractive AI

Extraction and technologies have been intertwined for millennia. Mining has long been a prominent form of resource extraction. And in recent decades, data mining has become a widely used metaphor for extracting valuable information in digital technologies. But for today's technologies, the logics of extraction—and their pernicious consequences—reach well beyond the extraction of resources and data. Communication scholar Cram (2022) has described this more expansive “extractivism” as the “cultural and ideological rationale that either motivates extraction or is the consequence of it.” Understanding extraction through the broader lens of culture and ideology reveals the wider range of ways that these extractive logics pervade western technologies and exploit vulnerable communities, leaving us with what Cram called a “violent inheritance” (p. 4). Extraction, in other words, is not just pulling resources out of the earth; it is also a set of norms and values that shape ways of thinking and acting. For the cultural critics Mezzadra and Neilson (2019), conceiving of extractivism as more than resource extraction is essential:

Lifting the concept of extraction away from its sectoral or literal association with mining and other forms of resource extraction allows an analysis that is attentive not only to Indigenous and antiracist struggles but also to more general predations of capital. (p. 138)

In a world dominated by the market economy, capital is everywhere. So too is the opportunity for extraction, exploitation, and resistance. Understanding extraction in this broader sense—as the siphoning of value, not just the dispossession of resources—draws attention to the menacing ways that extractivism animates many of today's most popular technologies, including AI.

Crawford (2021) exposed the extractive logics that undergird mainstream AI systems. She described how these AI systems pillage the earth in the search for precious metals; drain local ecosystems of water and natural resources; exploit underpaid laborers who work in poor conditions in the Global South; steal intellectual property from authors, artists, and creators without their consent; classify data in ways that reinforce existing inequities; and grease the wheels of state-sanctioned violence. She concluded that “AI systems are built to see and intervene in the world in ways that primarily benefit the states, institutions, and corporations that they serve” (p. 211). These systems define extractive AI: They rely on exploitation both to remove resources from ecosystems and to leach value from laborers and creatives all with the aim of redirecting profits from local communities to a distant elite few. As I will show next, understanding mainstream AI as extractive AI is essential to fully grasp the consequences of AI for TPC.

Contrasting Examples of AI

Of the many examples of extractive AI, Otter.ai underscores the challenge that these technologies present to TPC. Founded several years before the release of ChatGPT, Otter.ai, a subscription service that provides real-time transcription, promotes itself as an AI meeting note taker. Otter.ai can integrate easily with Zoom, Microsoft Teams, and Google Meet. “Never take meeting notes again,” says the home page (Otter.ai, n.d.). Sounds good, right? Technical communicators often have meetings that require note taking. Having an AI note taker would improve efficiency, and efficiency has long been an important—and fraught—concept in TPC (Clark, 2023). Plus, TPC scholars have advocated for years for audio and visual content to have transcripts in order to improve accessibility (Zdenek, 2014). But when seen through a social justice lens, such transcription tools tend to lack linguistic diversity. Transcription tools are often better at English than less widely spoken languages, including Indigenous languages. Driven by a market economy, transcription tools like Otter.ai tend to privilege languages spoken by more people because more speakers equal more profits. On the back end, AI transcription tools also rely on widely spoken languages because, for accuracy, these AI tools must be trained on massive amounts of text and recorded speech. Typically, these data are scraped from the internet without consent from, or compensation given to, the speakers and writers who provide their voices and words to the AI transcription tool. For sound studies scholars Sterne and Sawhney (2022), this kind of AI is the manifestation of a “techno-colonial will to datafy” (p. 290) that catalogs our online activities, transforming them into valuable data and selling the data back to consumers in order to bolster corporate profits. This extractive AI poses a problem for social justice efforts in TPC.

A traditional social justice response to this kind of inequity and lack of diversity would call for inclusion and accessibility: include more languages in the AI transcription tool, add more voices to the training data, expand the variety of English it supports, make AI tools accessible to speakers and readers of all languages, and so on. But there is a catch. First, market-driven companies are unlikely to pay for inclusive efforts that only fractionally grow their customer base, especially when those companies pilfered most, if not all, of their English-language training data from the internet. Sure, it would be good to make the internet a more inclusive space with a wider range of voices and languages taking center stage. But to encourage more marginalized voices to write and speak online would risk exposing those voices to being used without consent by extractive AI like Otter.ai. Second, Otter.ai uses the recordings and transcripts that it produces for users to further train its AI (Privacy & Security, n.d.). Users of this transcription service, then, are themselves mined for value by Otter.ai, which extracts users’ words and voices without compensation other than access to a tool that may not fully serve their needs. The logic of inclusivity would urge more marginalized speakers to use Otter.ai in order to help improve its transcription of marginalized languages, but doing so only opens those speakers to more exploitation. In this situation, Sterne and Sawhney (2022) cautioned that the “principles of inclusivity and access that critical scholars have long propounded … can also contribute to the very problems they claim to overcome” (p. 293). And these problems are not just limited to machine listening technologies like Otter.ai. Many of the mainstream AI systems—chatbots, text generators, image generators, multimodal AI, and so on—operate by the same extractive logics. By exposing the knowledges, languages, and voices of vulnerable groups around the world to be mined for corporate gain, extractive AI problematizes TPC's efforts toward inclusivity and accessibility.

To be clear, by interrogating how the logics of accessibility and inclusivity serve extractive AI, I am not attempting to undermine the validity of accessibility and inclusivity as essential values in TPC. Rather, I am suggesting that TPC scholars and practitioners must be wary of how their commitment to accessibility and inclusivity might be co-opted by extractive logics that profit from adding languages, dialects, and diverse representations to their data sets. As Sterne and Sawhney (2022) explained, “the idea here is not to deny the need for accessible speech technologies … rather it is to ask which technologies might best provide access for people who need it” (p. 298). They suggested, for example, that instead of employing unjust AI technologies to do voice-to-text transcription, employers could appropriately compensate human transcribers to do the same work (p. 303). Yet such solutions that could better serve both users and laborers often struggle to gain traction because they are not as lucrative as extractive AI.

In finding an alternative to extractive AI, then, the TPC field might look at an AI tool developed by Te Hiku Media that functions according to altogether different logics. Te Hiku Media is a Māori-owned Indigenous media organization in Aotearoa (New Zealand). As part of the long-running movement to revitalize te reo Māori (the Māori language), Te Hiku Media saw the need for a speech recognition tool for te reo Māori, but it knew the importance of data sovereignty. For Indigenous communities, data sovereignty (Taiuru, 2023a) and network sovereignty (Duarte, 2017) are essential, especially in the context of AI (Taiuru, 2023b). In Aotearoa, Māori data sovereignty means that “Māori data should be subject to Māori governance” (Te Mana Raraunga, n.d.) and not the governance of others, especially multinational for-profit corporations. Te Hiku Media did not want a non-Indigenous company from Silicon Valley to create and own a Māori-language speech recognition tool that extracts value from Māori-language speakers and sells it back to Māori as an AI transcription tool. So Te Hiku Media decided to build its own. To do so, it relied on decades of relationships built through Māori community radio and activism. Through these networks, Te Hiku Media crowdsourced a large and diverse data set of Māori voices by asking speakers from different iwi (tribes) to voluntarily record themselves speaking te reo Māori (Mahelona, 2020).

Crucially, Te Hiku Media does not consider itself as the owner of these voices and their words. Rather, it views itself as a guardian and steward of the Māori language and the voices that participated in this project (Mahelona, 2020). As such, Te Hiku Media told the contributors that their voices would not be datafied for profit by overseas technology companies but that the Indigenous communities who gave their voices, time, and knowledge to develop this tool would be the beneficiaries. By prioritizing Indigenous data sovereignty and crowdsourcing knowledge from local communities—choosing stewardship over ownership, community over profit, collaboration over extraction—Te Hiku Media models an alternative way of building generative language technologies without the damaging logics and practices of extraction. Rather than advocating for international companies to make their tools more inclusive and accessible to Māori speakers, which would have exposed Māori communities to exploitation and extraction, Te Hiku Media found a better way to achieve its goals on its own terms, according to its own values. So when faced with the threat of extractive AI, a discipline such as TPC that has a deep commitment to social justice has much to learn from Indigenous-led initiatives like this one.

For TPC and its neighboring fields, these two examples underline the importance of using an extractive lens when studying AI systems. The MLA-CCCC Joint Task Force on Writing and AI (2023, 2024) has released working papers that summarize the benefits and risks that AI systems pose to students, teachers, and our professions. These documents mark the first of many steps in our collective efforts to grapple with the implications of AI for our disciplines. Yet among the many AI-related dangers outlined in these documents, extraction goes unmentioned. The authors gesture toward the logics of extraction when they describe the “biased outputs, privacy and copyright violations, and environmental costs” of these AI systems (MLA-CCCC, 2023, p. 6). But I worry that readers will not fully comprehend the pervasive reach of extraction without an explicit discussion of the extractive ideologies that buttress mainstream generative AI systems. Neglecting the logics of extraction in discussions of AI systems risks treating only the symptoms of oppression and not the disease itself. Without using an extractive lens, technical communicators gain only a limited picture of AI systems, and such a narrow perspective threatens to expose vulnerable communities to harm. Studies of AI systems in TPC and its neighboring fields must interrogate not only the biases, inequalities, and injustices of these systems but also the extractive ideologies that bolster these systems. For the field of TPC to thrive, the ideologies of extraction—in AI systems or elsewhere in technical communication—must be laid bare.

Elevating the Use of an Extractive Lens in TPC

As an approach for viewing technical communication, the use of an extractive lens deserves more prominence in TPC scholarship, teaching, and practice. For researchers, using an extractive lens can highlight forms of exploitation in technical communication that might otherwise go unnoticed. Through this lens, researchers can also evaluate whether or not advocating for inclusivity and accessibility are appropriate responses to unjust communication technologies. Researchers with expertise in environmental rhetoric have already led the charge in examining some of the ways that technical communication can be extractive. They have studied, for instance, the relationship between extraction and unequal risk communication (Pflugfelder et al., 2023), the master narratives that obscure petrochemical harm in Puerto Rico (De Onís, 2021), and the “digital damage” caused by big data infrastructures (Edwards, 2020). These projects go hand in hand with “apparent feminist” (Clark, 2023, pp. 35–67) and “apparent decolonial feminist” (Haas & Frost, 2017, p. 170) approaches to risk communication, which make visible the need for feminist and decolonial interventions in environmental and other “slow crises” (Clark, 2023, p. 74). Building on this body of work, future projects can extend studies of extraction beyond the literal dispossession of resources to also consider the reach of extractive ideologies in less obvious aspects of technical communication. Such an enlarged scope would bolster the discipline's commitment to social justice by illustrating how extraction “alters and intensifies the social dimensions of exploitation” (Mezzadra & Neilson, 2019, p. 167).

Exposing the logics of extraction also involves using the lens of extraction to view the discipline of TPC itself. We do not yet know the extent to which extractive ideologies have penetrated the field, so researchers need to revisit the histories of technical communication in order to interrogate how TPC has been complicit with extractive industries by helping to further the exploitation of both human labor and nature. Pflugfelder et al. (2023) have offered one such historical investigation in their analysis of water access in the Colorado River Basin. And Pflugfelder et al. (2025) are currently working on a book that further examines the discipline's relationship to extractive industries by analyzing the sixteenth-century Latin mining textbook De Re Metallica and the extractivism implied in cases studies of petrochemical companies that TPC has uncritically presented. These are important investigations for the field, and TPC needs more of them. However unsettling it may be, now is the time to re-examine the histories, theories, methods, practices, and pedagogies of TPC to uncover their relationship with extraction.

At the same time that we reflect on our past and present, we must also imagine what kind of future we want for TPC in light of extractive AI. The field must invent and enact ways to “move against and beyond extractive world making” (Cram, 2022, p. 24). Hart-Davidson (2018) argued that we should teach rhetoric to robots (p. 253). As part of that effort, we must also work toward building robots and AI systems that adhere to ideologies that are different from the other parasitic EdTech tools on the market. Eli Review is a model of a collaborative, scholar-led technological platform that operates according to such a nonparasitic ideology (McLeod et al., 2013). What else might we be able to build when we think collectively, learn from Indigenous communities, and resist the damaging logics of extraction? “How might we,” then, as Edwards (2021) asked, “imagine and build an infrastructure otherwise?” (p. 10). Author and activist Klein (2014) has suggested countering the tentacles of extractivism by adopting a “worldview based on regeneration and renewal rather than domination and depletion” (p. 424). And critical internet researcher Noble (2021) has asked, “What would it look like to create resources and runaway investments in communities that have been tremendously harmed by a variety of different kinds of [extractive] projects?” Both Noble and Klein posited alternative futures that could be fruitful for TPC.

Technical communicators must discuss these questions more widely and urgently en route to articulating a collective vision for the field moving forward. A declaration of our disciplinary commitments relative to the logics of extraction—akin to the seminal “antenarrative of technical communication” (Jones et al., 2016)—could orient the field's response to the threat of extractivism. But we must not rely on any single document to resist extractivism. Rather, we must, as a field, keep the lens of extraction near at hand and use it to critique the insidious technologies that foster and augment injustices in our world. And like our colleagues at Te Hiku Media, we must know when to set aside the extractive lens all together and build new technologies according to more just values.