This study examines how the language of posts on Amazon’s Mechanical Turk (MTurk), an online global crowdsource marketplace, rhetorically depicts human agency and identity. Launched in 2005, MTurk allows individuals and organizations to solicit a distributed pool of workers principally for menial jobs that are impractical for computers. Examples of such "Human Intelligence Tasks" or "HITs" include identifying objects in images, correcting transcription errors, and verifying product descriptions. The reward for each HIT is often only a few cents, which has led to controversy about compensation on the platform (Hara et al., 2018; Semuels, 2018; Bergvall-Kåreborn & Howcroft, 2014; Cushing, 2012); nevertheless, it is quite robust, with thousands of jobs and workers (or "turkers") active at any given time (Difallah et al., 2018). Although there are now several similar crowdsourcing services, many suiting niche markets, MTurk remains one of the most significant, partially because of Amazon's salience.
MTurk's curious moniker derives from The Turk, a chess-playing automaton that was publicly exhibited to great interest throughout Europe and the Americas during the 18th and 19th centuries. Although the device, which won several matches against accomplished chess players and luminaries including Napoleon and Benjamin Franklin, was billed as being completely mechanical, in reality it incorporated human confederates within and without, and functioned as a complex performance of people and objects.1 In contemporary usage, the phrase mechanical turk has pejorative connotations, signifying something that purports to be driven by sophisticated computer technology but covertly relies on a concealed human labor pool—a digital sleight-of-hand by barkers hoping that the glamour will gain them Series A funding. Amazon has explicitly embraced these implications, with chief executive Jeff Bezos describing MTurk as "artificial artificial intelligence" (as quoted in Pontin, 2007, para. 3).
MTurk is an active research topic for disciplines within the social sciences (Hauser & Schwarz, 2015; Casler et al., 2013; Paolacci et al., 2010), particularly economics (Hara et al., 2018; Ipeirotis, 2010a; Ipeirotis, 2010b; Ross et al., 2010), as well as psychology (Cheung et al., 2017; Paolacci & Chandler, 2014), sociology (Shank, 2016), and education (Follmer et al., 2017). However, studies from rhetoric and adjacent fields such as computational linguistics are limited, and most of them focus on using MTurk to produce research rather than address it as a subject of study (Fort et al, 2011; Gibson et al., 2011; Sprouse, 2011; Schnoebelen & Kuperman, 2010). This is unfortunate, because MTurk comprises a large, dynamic body of written texts manifesting contemporary rhetoric concerns including persuasion and ethos. Most applicable for our purposes, the platform functionally distinguishes humans from non-humans through written language acts driven by pragmatic commercial needs. This is a valuable counterpoint for conversations about such distinctions that commonly hinge on speculation about humanity’s elevated intellectual or ethical traits. In rhetoric studies, the opposition between exalted essential human qualities and impoverished mimetic object traits is evident as far back as Socrates’s denigration of writing as a prosthesis in Plato’s Phaedrus, as scholars including Walter Ong (1986), Gregory Ulmer (1985), and Jacques Derrida (1983) identify. Although the influence of new materialism has resisted this framing by positioning digital technology as humanity’s beneficial complement (Brooke, 2009; Clark 2004; Hayles, 1999) or productive partner (Kelly 2011), it nevertheless often operates at the level of abstract faculties. It would seem equally productive to shift our attention to the other end of the spectrum, as it were, and examine concrete, mundane tasks—operations that are trivial for humans but difficult for machines. MTurk provides an exemplary site for such study, as the ongoing progression of HITs establishes a fine-grained catalog of characteristic human actions in practice. MTurk provides a means to build a bottom-up depiction of the human rather than a top-down one that inherits pre-existing frameworks that reify venerated human capacities. The language of HITs provides concrete, material distinctions between humans and machines not based on high-flung intellectual or moral capacities but quotidian behaviors vetted by market forces.
A realignment of labor along an axis spanning humans and digital machines would produce a corresponding revised human ethos—a group identity based on what we do and thus are. Such a model would be akin to others that make similar reclassifications. David Autor’s (2014) “Polanyi's Paradox and the Shape of Employment Growth” draws from Michael Polanyi’s philosophical assertion that humans’ embodied functional knowledge exceeds their conscious explicit understanding to explain contemporary labor market polarization: the simultaneous growth during the period of digital automation of high-wage jobs that require creative problem solving and low-wage manual labor that demands situational adaptation to sensory data. Autor contends that these poles are united because they are difficult to articulate fully, which makes them resistant to automation. This establishes a new binary labor classification in a digital era that obviates conventional distinctions predicated on education or skill level. Something similar has occurred during the labor reclassification of essential and inessential work in response to the COVID-19 pandemic. The essential category includes occupations such as physicians, machinists, delivery people, politicians, cashiers, and pharmacists; it too muddles established separations between high- and low-skill and compensation disparities. Our goal is to determine the ways MTurk makes a commensurate change in the labor calculus and thereby revises “human” and “digital machine” as coherent classifications.
This study thus engaged the research question of how language in MTurk HITs rhetorically constructs distinctions between humans and digital machines based on labor and agency, and what discernable characterizations of the human and non-human emerge from these rearticulations. Because MTurk produces a large, continually changing corpus of brief written documents, we addressed these issues through a method based in quantitative text mining. This allowed us primarily to work inductively, drawing conclusions from emergent patterns in the language, rather than attempting to suit data into a pre-existing essentialist framework. In attempting to avoid abstract presumptions about humans’ intellectual or ethical capacities, an aggregate reading of HITs’ concrete particulars is advantageous. We acknowledge that any truly neutral positioning is likely impossible, and the data never speak for themselves; nevertheless, we made intentional method choices to align with how MTurk itself functions and serve as a needed mundane counterpoint in disciplinary conversations about humans’ distinctions from digital machines.
This study's findings suggest that HITs depict humans as sophisticated receptors for sensory data and emotional resonance with keen judgment about social constructions—yet hard distinctions between humans and non-humans are somewhat elusive. The following sections provide details about our observations, and it is crucial to acknowledge that we do not position this study as an attempt to discover pre-existing objective characterizations; MTturk is not a window that provides insight into true essences, nor is it a subsidiary symptom of a broader political, social, or economic prime mover. This study’s goal is to trace, in a circumscribed way, how MTurk continually remakes distinctions between humans and non-humans. In this we have followed Marshall McLuhan’s (1967) guidance in Hot & Cool: “I don’t explain — I explore” (p. xiii). This study, as part of the larger Following Mechanical Turks research project, here seeks to map some of MTurk’s convolutions.
The degree to which audiences were "fooled" by The Turk is debatable. Some contemporary reports demonstrate credulity, but many others, including Edgar Allan Poe's (1836) well-known essay "Maelzel's Chess Player," focus more on investigating how the effect was created. As with any good magic trick, it would seem that many if not most observers understood that they were encountering an illusion, but that in no way undercut their fascination with the performance, and it should not devalue the complex labor undertaken by multiple humans and non-humans working in concert. ↩