The Living Machine: A Computational Approach to the Nineteenth-Century Language of Technology

Abstract

This article examines a long-standing question in the history of technology concerning the trope of the living machine. The authors do this by using a cutting-edge computational method, which they apply to large collections of digitized texts. In particular, they demonstrate the affordances of a neural language model for historical research. In a deliberate maneuver, the authors use a type of model, often portrayed as sentient today, to detect figures of speech in nineteenth-century texts that portrayed machines as self-acting, automatic, or alive. Their masked language model detects unusual or surprising turns of phrase, which could not be discovered using simple keyword search. The authors collect and close read such sentences to explore how figurative language produced a context that conceived humans and machines as interchangeable in complicated ways. They conclude that, used judiciously, language models have the potential to open up new avenues of historical research.

Introduction

Machines have always occupied an ambiguous position in the human imagination: their motion and action tempting us to figure them as alive, even though we know they are not. These small acts of anthropomorphism—common to the way humans relate to the planets or the weather—are heightened thanks to the apparent similarities between humans and machines.² These parallels have evoked much cultural and political discussion, with machines imagined variously as uncanny, sinister, or out of control; as trans-formative, superhuman, or the archetypal nonhuman.³ By drawing on these similarities and differences, writers have reaped a rich harvest of descriptive figures, symbols, and metaphors.⁴ This language of machines always exists in a concrete historical reality into which new machines emerge, further shaping—but also shaped by—the stock of phrases in circulation. Historians of technology have long explored this language. However, the uniquely complex semantic field in which humans and machines have been juxtaposed, as well as the ubiquity of the phenomenon, create difficulties in gathering and interpreting the evidence. This article explores this complex and ubiquitous exchange of meaning in a systematic way, using a method not previously attempted. We use a language model to "read" large, digitized collections of texts, thus employing a living machine of our own, in an intriguing twenty-first-century recapitulation of this same historical trope. An advanced computational method helps find figures of speech not discoverable by simple keyword search, which we then close read and contextualize, to suggest new avenues for research in the history of technology.

In the nineteenth century, machines could be seen to move by themselves, but they could not think and so occupied a position distinct from humans and animals—in some sense "animate," even agential.⁵ But aside from infamous hoaxes such as Maelzel's Mechanical Turk (a chess-playing machine that concealed its human operator), the properties of learning and thinking were understood to mark the limits of machine capability. The mathematician Alan Turing would complicate this boundary in his classic 1950 essay, "Computing Machinery and Intelligence," by positing a machine that could pass as human; an increasingly fuzzy boundary as machine learning algorithms and so-called "artificial intelligence" become commonplace.⁶

Our article explores nineteenth-century English-language texts using a probabilistic representation of language known as a neural language model.⁷ Such models have themselves been widely figured as alive in the contemporary imagination because of how they simulate human language; an impression increased in the case of the model we use thanks to its playful name, BERT, derived from its learning architecture (Bidirectional Encoder Representations from Transformers).⁸ The discussion of language models has centered on what they can do: notably, produce text automatically, mimicking human communication; this was the focus of intense debate when OpenAI's GPT3 ChatBot was launched in late 2022. Thanks to skillfully orchestrated public relations, this was the first language model to receive widespread attention, raising ethical, epistemological, and ontological questions, and fueling debate about the nature of creativity itself.⁹ To date, science and technology studies (STS) has witnessed little theoretical discussion of language models and the nature of the knowledge they produce.¹⁰ Such epistemological inquiry lies beyond the scope of this article; however, our demonstration of the affordances of language models for humanistic inquiry highlights the potential for such new forms of research and raises questions for historians and theorists to consider.¹¹

Humans and machines have been interrelated discursively for millennia, inviting reflection on their substitutability, above all in relation to work.¹² This thematic concern gave machine language a degree of consistency through the early modern period.¹³ However, the rapid industrialization experienced in nineteenth-century Britain accompanied a language of machinery that reflected a new mass phenomenon. Instead of the rarefied Swiss mansions of eighteenth-century philosophes, this was the age of the railway journey and the factory floor. In such a context, we ask: How did Britain's vibrant print culture report on the massification of machinery? The motif of the living machine has long been noted in literary and philosophical texts but rarely explored in more commonplace genres, which nonetheless discussed at length the coming of rail, the impact of new inventions, and safety in the workplace. When newspapers reported an 1854 debate on chimney sweeps, the discussion focused concretely on the boundary between human and machine work: "there were no chimneys in existence" understood one parliamentarian, which "might not be cleaned by machine."¹⁴ Sedimented in this simple sentence is a hidden seam of historical meaning because, to a language model trained on nineteenth-century English, it is very surprising that it ends with "machine." This article applies probabilistic insights to everyday texts from the high-water mark of industrialization, to capture such changing cultural and political expectations.¹⁵

We explore the linguistic contexts of the word "machine" to make better sense of how machines more generally were being conceptualized.¹⁶ The task is therefore akin to conceptual history; however, our approach makes use of a wider sample of language, much closer to its everyday form than the elite or literary texts on which historians of technology have previously tended to rely.¹⁷ Our aim is to complement the work of scholars of mechanical discourse in Victorian Britain by approaching this cultural history if not from below, then from the perspective of big data.¹⁸ A keyword absent from our experiments is "technology," a twentieth-century term not yet operational during our period, but one whose shadow looms over it. The unavailability of "technology" to contemporaries was astutely described by the historian Leo Marx as a "semantic void," marking the lack of a concept to describe the emerging and omnipresent system of machines in general.¹⁹ This article does not add to the growing genealogy of technology but ventures deliberately into the murky semantic space Marx mapped out.²⁰

We combine a computational method for finding instances of discourse about machines with close reading and interpretation. Our approach identifies recurrent patterns embedded in language by going beyond simple keyword search. Figurative language in relation to machinery reveals deep-seated attitudes, including cliché and emergent turns of phrase beyond the literary canon, and so we use a large sample of English-language sources (ca. 25 million pages, described below), encompassing a broad range of contemporary voices. Our understanding of industrialization is guided less by the influential Culture and Society tradition associated with Raymond Williams and more by Humphrey Jennings's wild and open-ended work of montage, Pandæmonium, which sought to capture "the coming of the machine as seen by contemporary observers."²¹

Our findings reveal the complexity of machine language, not least because of the myriad substitutions and transfers of meaning from machines to humans or other living entities and back. "A horse and an ox are machines," wrote the anonymous author of a popular working-class history published in the Northern Daily Times in 1859, "substituting manual labour as much as a steam engine, with this inconvenience: that they are living machines, requiring to be fed with or from what would feed men."²² These figures tapped into widely observed continuities between horse, ox, steam engine, and manual laborer—implying their historical interchangeability and noting the irony that any device that saves labor must nonetheless be fed. A parallel dilemma applies to language models that must also be fed, not only with a stock of words but also electricity, to animate the machine learning on which our own form of work relies.²³ In particular, the task of identifying semantically similar words speaks directly to this imagined equivalence between humans and machines that saw Victorian authors substitute them in all sorts of ways.

Machine Models

Figurative language involving machines provides rich cultural evidence of their conceptualization, especially in relation to humans. Finding such phrases in order to analyze and close read them is not a straightforward task, despite the availability of large, digitized text collections. Traditional keyword searches are unlikely to capture the variations that occur in language. N-gram searches centering on common phrases, or short strings of words, are a crude tool and miss many instances of the phenomenon we are seeking.²⁴ Our proposal is to use a semantic property of this language that underlies the discursive interchangeability of humans and machines, involving a sort of game in which we challenge a language model to guess which word it would expect to fill a gap in a sentence based on the context. What the language model "expects" is the result of complex calculations, offering a novel way to describe a body of texts quantitatively. The interplay between the model's expectation and the human reader's interpretation allows historical inferences to be drawn about the cultural context of those texts. The specific methodological challenge, therefore, is to find instances of figurative language—motifs that are hard to specify using fixed n-grams or a particular syntax. This builds on our previous method for detecting instances of machines described as if they were alive, which identified the linguistic phenomenon of "animacy" in "atypical" entities; that is, ones not in fact alive.²⁵

The choice of texts for computational linguistic research is a critical determinant of the analysis and so requires explanation. Our criteria in creating this corpus were representativeness, diversity, and availability. Representativeness is an increasingly thorny topic in digital scholarship, due to concerns that despite their vast size, certain datasets do not adequately represent the phenomenon in question: size alone cannot solve the problem of unrepresentative data, nor absolve the digital historian of the need to conduct source criticism at scale.²⁶ Diversity reflects our desire to go beyond a single genre such as fiction (the focus of several nineteenth-century datasets) and include a breadth of language closer to that used in the period. The need for diversity goes beyond ensuring a sample (e.g., of novels) adequately represents the wider population (of novels), but that it represents the heterogeneity of social reality. Availability is the nontrivial issue of which datasets can be used freely and openly by colleagues without paywalls or special access arrangements to replicate or explore our results. Our analysis is restricted to open data and eschews the privatized text corpora offered by academic publishers or genealogy corporations.²⁷

Our corpus of texts consists of three parts:

1. Ca. 48,000 books made freely available by the British Library (hereafter British Library Books or BLB), mainly nineteenth century and containing ca. 5.1 billion words. This collection is the result of an abortive digitization initiative in 2011 by Microsoft, who gave up on the project at an early stage and made the underlying text and images available to the library free of

Figs 1, 2 & 3.

Machine Sentences Over Time. The top line shows the number of sentences containing the word "machine(s)" (y-axis) by year (x-axis). The bottom line shows the annual sentence count after filtering. The upward trends reflect the increase in data over time. Filtering tends to remove more sentences when the underlying text is noisier. This explains the wider gap between the two lines for newspapers, which suffer worse OCR quality, as compared to the other two datasets.

copyright (unlike subsequent projects like Google Books, which remain unavailable en masse to researchers). Microsoft did not select volumes to scan systematically, resulting in a collection whose contents if not random are at least arbitrary, and which includes novels but also geography, history, and other nonfictional genres.

2. Fourteen national or regional nineteenth-century British newspapers (hereafter Heritage Made Digital or HMD14), containing ca. 2.4 billion words, also freely available from the British Library thanks to a publicly funded digitization initiative. These titles are not a comprehensive selection but result from curators' attempts to fill gaps in existing digital collections; although idiosyncratic, these titles offer breadth as well as coverage from ca. 1800 to 1871.

3. A much smaller set of texts expected to contain significant discussions of machinery, comprising 5,169 articles from a long-running technical and scientific journal published by the Society of Arts (hereafter JSA), spanning roughly 1783 to 1908, containing ca. 60 million words, and freely available from JSTOR Data for Research.²⁸

The three collections constitute a vast amount of data. We filtered the sentences to those containing "machine" or "machines" in noun form, published between 1783 and 1908; material of low quality was then removed—digitized historical collections often include illegible text resulting from poor Optical Character Recognition (OCR) and incomplete sentences. Postprocessing filters helped catch and remove as many of these as possible.²⁹ The end result was a set of 185,100 high-quality sentences for our investigation.³⁰ As this remains an impossibly large amount of data to read, we identified figurative instances of living machines using a method based on the idea of distributional semantics.

The distributional hypothesis of language—often illustrated with the line "you shall know a word by the company it keeps"—postulates that words occurring in similar contexts have similar meanings.³¹ This intuition implies that the meaning of a word consists in the contexts where it appears. By quantifying these contexts, language models produce a probabilistic representation of linguistic usage. Techniques to create such models have evolved from simple counts of words and co-occurrences to use neural networks that "learn" the contextual meaning of words from large text corpora and produce dense vector representations of each word, approximating to its meanings.³²

The language of machinery, with its multifarious metaphors, is a good example of heterodox or atypical linguistic usage not previously amenable to computational methods. However, the contextualized word embeddings used by BERT models hold out the prospect of analyzing even such difficult turns of phrase systematically. Linguists have explored how "being alive" has transposed figuratively from core instances (e.g., humans) to ostensibly inanimate entities through phrases that confer properties like motion or intelligence, such as in this sentence from BLB: "the most ingenious machinery, working with the precision of an intelligent being, is that of the paper box machines."³³ In this very common formulation, the machine acquires the animate properties of ingeniousness and intelligence typically reserved for human agents. Among inanimate entities, machines are a challenging border case, largely because of their semantic proximity and (often uncanny) similarity to humans. The phenomenon linguists describe as "atypical animacy" is complex to detect and may be time specific—it might become typical to describe certain inanimate entities as alive over time, a historical change to be expected in the long nineteenth century, but not yet investigated.

Our approach to detecting atypical animacy involved training contextualized embeddings with a BERT masked-language model.³⁴ Because language is anchored in its geographical, historical, and ideological context, such models are considered faithful representations of the corpora on which they are trained. To this end, we fine-tuned a base model on nineteenth-century works from the BLB corpus published ca. 1760–1850, containing ca. 1.3 billion tokens, which we call "BLERT.³⁵ Given a sentence containing a hidden or "masked" term, BLERT was used to predict the most likely word under the mask. "Machine" was masked in this sentence: "A [MASK] pushes along a strip of cardboard beneath a punch, or rather group of punches, fixed on one block or keyboard."³⁶ Using contextual information learnt by the model, BLERT produces a ranked list of predictions, shown here in decreasing order of probability: "man," "woman," "boy," "female," "person," "hand," "girl," "weaver," "gentleman," "child," "lady," "tailor," "blacksmith," "carpenter," "master," "negro," "worker," "servant," "smith," "machine." From this list of candidate words, we calculate the extent to which one would expect the entity behind the mask to be alive by producing a weighted animacy score.³⁷ If the model predicts typically animate entities in the subject position, this reveals a discrepancy because we know the subject of the sentence, by contrast, to be a machine. Given the totality of the language in the corpus, for a machine to be in a position where one would expect a living being is the mark of a heterodox semantic usage and a representational mismatch in the language of the period.

We extend this method to provide a coarser-grained, more robust analysis better suited to historical inference and generalization; as well as obtaining a fine-grained animacy score for a given expression, we are also able to quantify the semantic similarity between BLERT's predictions and a higher-level concept or category. An example (this time from HMD14) is a sentence, again with the word "machines" masked: "The Duke of ARGYLL said he understood there were no chimneys in existence which, by opening traps, might not be cleaned by [MASK]."³⁸ Our BLERT model predicted "boys" (0.91) as by far the most likely word, followed by "men" (0.03), "children" (0.01), and "them" (0.01). Interestingly (with due respect to the Duke of Argyll), "machines" does not appear in the top twenty most likely candidates. These probability scores reveal how anomalous or surprising it is, from the model's perspective, for "machine" to complete the sentence, given the characteristics of the nineteenth-century corpus on which it was trained.

This extends our previous method, which provided a single animacy score per masked expression, by using word embeddings that situate BLERT's predictions in relation to a broader set of concepts. These allow us to see beyond the top prediction and get a more robust sense of what the model takes the sentence to be about. This is done by calculating the semantic proximity between the probabilities of a given list of predictions and each one of a defined set of concepts, chosen for their historical interest, for example: "child," "workforce," "energy," or "slave."³⁹ For a historian exploring changing nineteenth-century suppositions about the substitutions of labor—registered in the semantics of work—these categories offer a more useful typology for entities expected to be the subject of the sentence than a single search term. This technique is thus a powerful diagnostic tool for identifying the leading edge of social and economic change, captured within the language in our corpus. Filtering by category makes it possible to identify the ways that surprising, novel, or unusual linguistic formulations involving the token "machine" vary by topics like "energy," "transport," or "females." For the masked expression above, the compatibility scores with the concepts "child" and "workforce" are 0.947 and 0.966, respectively (1 is the highest possible score), while "energy" and "transport" score significantly lower. Semantic analysis pitched at the conceptual level helps find relevant examples more readily and focus on the most surprising or unexpected results.

Machine Figures

Our approach was to identify the many scattered examples of figurative language in relation to machines and group these "machine figures" for analysis as a body of sentences. These sentences subvert typical semantic practices, sometimes subtly, in relation to the animacy or agency of their subject. This commonly occurs through anthropomorphism, or the personification of a machine described as if it were alive, by using metaphor or verbs, adjectives, or adverbs typically reserved for a living entity. For instance: "The machine is fed by a compassionate girl with papers, which are eagerly swallowed, folded, gummed, and dried by this voracious mechanism, and evolved in a perfectly finished packet form in an indescribably short space of time."⁴⁰ In this characteristic example—an 1893 account of the Wills cigarette factory in Bristol—we hear of machines being "fed," swallowing "eagerly," and being "voracious." The descriptions pertain to the animal act of eating with appetite; they do not exist in the body of the machines but in the mind of the author and consequently in those of his readers.

Such metaphors are so normalized that we barely notice them; however, it is these everyday linguistic formulations that effectively transfer animacy and agency from humans to machines. The political stakes are significant. Figuring machines as active agents helps create a framework for understanding the relationship between society and technology where human agency, action, and intention are elided and rendered invisible, removing responsibility for decision-making from the people in charge, in a form of technological determinism: "The machine did it!" The specter of self-acting or autonomous machines, charted by the political scientist Langdon Winner forty years ago, has returned in recent times.⁴¹ The specific nineteenth-century instances of living machines we identify are like those "hackneyed vignettes of technologically activated social change" familiar to us today. They are, nonetheless, the historical building blocks for the more generalized twentieth- and twenty-first-century concept of technology in general, which Leo Marx also described as "hazardous," because of how the term takes on "a thing-like autonomy and a seemingly magical power of historical agency."⁴² Marx and others have traced the general development of "technology" as part of a critique of determinism (and its hazards). We offer examples to help substantiate such critique with a broader evidence base.

Figurative language by its conspicuous nature reveals assumptions about agency and animacy that might otherwise remain unnoticed. However, it remains difficult to detect the phenomenon of metaphor computationally.⁴³ To make systematic sense of such language requires combining the power of text mining for discovery, with close reading and interpretation for the task of analysis. Figurative language in relation to machines includes metaphors (both deliberate and accidental) but also clichés and even slips of the tongue.⁴⁴ These varying linguistic forms have received much literary and hermeneutical attention, from psychoanalytic and Marxist critics to the very different approach of pragmatics—the study of how people achieve communicative goals.⁴⁵ A comprehensive account of linguistic practice in relation to machines is not possible here, but borrowing insights from a combination of these approaches helps develop a form of search well suited to our target phenomenon. At the very least we must acknowledge that figurative language has always helped constitute scientific and technological discourse, if only in its conspicuous disavowal by those such as Thomas Sprat, who in 1667 attempted to codify a pure form of language that rejected "swellings of style" and insisted instead on "a close, naked, natural" communication unlike what "Wits, or Scholars" preferred.⁴⁶ Myriad histories of science and technology written since Sprat's injunction have by contrast highlighted the centrality of metaphor in framing not only the popular but also the expert discourses of Darwinian biology, genetics, and economics, among many others.⁴⁷

Before characterizing the many instances of such language, you first have to find them. Our aim was to devise a method for detecting generic rather than specific figures of speech, which deliberately or unwittingly bestow animacy on a machine. A procedure described as "broad winnowing" and "harvesting" of data (detailed below) involved filtering many sentences before a manual selection of salient examples for close reading.⁴⁸ An illuminating example of a different, but widely recognizable, form of anthropomorphism appeared in a book describing southern Africa's gold fields; ostensibly from a geographical work, the passage is one of many that merits detailed excavation. Having brought a delegation from Matabeleland in British colonial South Africa to visit London, the narrator recounts how "what most astonished them was the telephone. They were placed a mile apart, and talked together. Afterwards they declared that they could imagine such a machine might talk English, but how it could be taught to speak the Kafir language they could not understand.⁴⁹ The experience of wonder and astonishment at this encounter with new machinery was a common trope in late Victorian culture.⁵⁰ The shock of the new was discursively heightened by recounting these "primitive" reactions on witnessing Western technologies for the first time.⁵¹ The mocking humor trades on the assumption of racial or civilizational difference, that no (white) modern person would make such a blunder as to assume that a telephone really talks. Yet the wondrous nature of this new machinery is precisely the register in which it was also described in Britain; the supposedly nonsensical idea of the machine being alive, already sedimented into widespread linguistic usage.

Whether or not writers intended it, our language model consistently captured the way that, at some level, people widely posited machines as alive. In addition, it is no coincidence that themes of race and empire recur in technological discourse, which often centered on the exploitation of natural or human resources. It is ironic that buried within the naive visitor's question—how a machine could possibly learn an African language—lies a present critique of language models: by being "stochastic parrots," they have been "taught" to speak English much more than other languages. The deleterious effects of this problem are now better understood and clearly acknowledged as a residual feature of the tools used here.⁵²

Our method for detecting atypical animacy revealed examples of "classic animacy" in relation to machines, where they are personified or subject to anthropomorphism: they act, move, or work by themselves. Future work could classify the varieties of this classic animacy, either in relation to specific machines or in terms of the animate entities selected for comparison. However, a key feature of our approach is that although it is inevitably shaped by our preconceptions or hypotheses, it remains open to unexpected or serendipitous findings analogous to those of historians working in a traditional archive. An unexpected outcome was that our model also detected many more examples in which the comparative figure flowed in the opposite direction—with a being described as if they were a machine, in what we might call dehumanization if applied to a human subject, or abstractly, "technomorphism." From the perspective of the model, a comparison in either direction—human to machine or vice versa—involved a similar semantic slippage from the expected grammar.

Such examples abound in our corpus, often in surprising contexts, sometimes deliberately to dehumanize the subject but not always. Thus, we see the value of such large collections in allowing the discovery of relatively rare phenomena. The notion of surprise indicates material worthy of closer inspection in the technical sense that it appears as surprising to the model, rather than prima facie to a historian. In a sentence from BLB, Lord Byron expresses wonder at the creativity of the Irish orator John Philpott Curran, claiming "there never was anything like it that ever I saw or heard of. He was a machine of imagination, as some one said that Piron was an epigrammatic machine."⁵³ The use of "machine" as a superlative is surprising, not only because Byron was well known for opposing mechanization but also because the property being underscored is imagination, archetypically cast as its opposite.⁵⁴ More common examples involve authors figuring an organization, institution, or something abstract as mechanical, with deliberately pejorative intent. In his History of England in the Eighteenth Century, William Lecky quotes Edmund Burke's opinion of the early 1700s anti-Catholic penal code as "a machine of wise and elaborate contrivance, and as well fitted for the oppression, impoverishment, and degradation of a people, and the debasement in them of human nature itself, as ever proceeded from the perverted ingenuity of man."⁵⁵ The metaphor of an unflinching or rigid entity as a machine became commonplace but is developed powerfully in Burke's connection back to the deliberate and appalling "ingenuity" necessarily involved in its human construction.

Even this cursory close reading gives a sense of the complex semantic field in which the machine figure is situated. Our initial exploration points to the inherent bidirectionality of the descriptions as a key feature of this evolving language, highlighted by the way our model did not distinguish between sentences that conferred a mechanical attribute to a nonmachine entity and vice versa. This serendipitous feature of the model had an unexpected yet highly relevant outcome. Going further, a subclass of cases involves both comparisons at once, such as the rhetorical question in this sentence: "Are we merely machines, living automatons, aggregations of atoms?"⁵⁶ The author invokes two chimerical figures, which demarcate the essence of the human by displaying its opposite—the mechanized human and the living automaton.

This comparative and contrastive language has been used in ways that merit more attention than is possible here; but suffice it to say that in a significant number of machine sentences, the meaning flows in a "technomorphic" direction.⁵⁷ A final example illustrates the most common trope, explored in the following section. An author, describing the degeneration of the French Army ahead of its trouncing in the Franco-Prussian war of 1870, describes how "the system of drill was so devised as to give no play to the reasoning power of the officer. He was a machine and nothing more."⁵⁸ Instead of its positive attributes as tireless, steadfast, and reliable, the machine is deployed as a figure to depict a dehumanized or sub–human being, lacking an essential part of its nature: a simulacrum, no more than a machine, "a mere machine."

Machine Compounds

The figurative search outlined above winnows down sentences that share a thematic and linguistic pattern in order to identify relevant material for close reading and interpretation. Across the books, newspapers, and journals in our corpus, the trope "mere machine" recurs with striking frequency. Authors use this complex, technomorphic figure with pejorative intent and often in a negative formulation, such as in "The laborer should not be a mere machine."⁵⁹ This sentence from the Journal of the Society of Arts appears in a report about a new training institute for artisans. It indicates in what type of discussion this figure features, as well as its deceptively complex syntax of double negation. John Ruskin uses the same figure to discuss the effects of architectural style: "whether the workman shall be a living, progressive, and happy human being, or whether he shall be a mere machine, with its valves smoothed by heart's blood instead of oil—the most pitiable form of slave."⁶⁰ This arresting passage invokes a stark opposition between a full vision of the human and a mechanical travesty, fleshed out further by comparison to a figure considered still less human: the slave. The rhetorical entanglement between slave and machine runs deep in nineteenth-century language for many reasons, not least the force with which one could illustrate the dehumanized nature of the other. In Ruskin's time, with the abolition of slavery, the figure took on the power to shock, pointing to the anachronism that the promised liberation of workers by machines had left them in a pitiable state of metaphorical slavery.

A more genteel version of the same figure appears several times in the writings of Charles Dickens, who seems to have found it both topical and congenial to his satirical style, trading on the tragicomic image of a mechanical subhuman.⁶¹ In Sketches by Boz, we hear that "Mr. Samuel Briggs was a mere machine, a sort of self-acting legal walking-stick."⁶² The figure in this set piece provides an absurd personification of unflinching bureaucratic or commercial rigidity and is reprised in A Tale of Two Cities when bank manager Jarvis Lorry explains his remoteness from his clients: "I have passed from one to another, in the course of my business life, just as I pass from one of our customers to another in the course of my business day; in short, I have no feelings; I am a mere machine."⁶³ The emotional dynamic of the "mere machine" is not always negative. More complex constructions involve nested meanings, such as this reflection: "We were told that universal benevolence was what first cemented society; we were taught to consider all the wants of mankind as our own; to regard the human face divine with affection and esteem: he wound us up to be mere machines of pity."⁶⁴ Here, the machine figure plays a double role: creating distance from an inherently human act, feeling pity, but imagined being performed mechanically and thus as egregiously inauthentic. Taken from the Works of Oliver Goldsmith, this instance illustrates a notable philosophical and theological context for machine language. These examples occur earlier, even before our period entirely.⁶⁵ We conjecture that the locus of machine language shifted from eighteenth-century discussions of theology and metaphysics—with the importance of mechanism for debates around dualism—to everyday and industrial contexts of labor by the late nineteenth century.

To quantify the incidence of this unusual turn of phrase, in which the adjective "mere" modifies the noun "machine," we used a sequential tagger to annotate the role played by each word in a sentence. The tagger identifies dependencies between words, allowing us to see all the different adjectives

Fig 4.

A Mere Coincidence? This table shows the frequency of words appearing in a sentence alongside "machine" and acting as modifiers. For the two major datasets in our corpus (in the left-hand columns) we see that "mere" is among the most frequent modifiers and so "mere machine" constitutes a very significant figure of speech.

used with the noun "machine." Listing these adjectives by frequency allows us to explore their prevalence (see the table in figure 4).

In the BLB and HMD14 corpora, the adjective "mere" occurs 1,065 and 319 times, respectively. This makes it the fifth most frequent modifier of "machine," behind "other" (which might be considered a stop word); "threshing," revealing the persisting importance of agriculture as a site of mechanization; "infernal," referring to new forms of weapon; and "new," marking the perennial association of novelty with machines. In the JSA corpus, "mere" does not feature in the top fifty adjectives, reflecting a different genre, perhaps less concerned with the dehumanizing effects of technology. Therefore, in the two large subcorpora containing most of our texts—which come closest to everyday language—"mere machine" is a very significant figure of speech.⁶⁶

If richer and more detailed metadata are created, it will become possible to explore large text corpora in a more granular fashion. For example, if properties such as date and genre can be discerned reliably, long-standing issues in the literature about the changing cultural valence of machines and mechanization will become amenable to text mining.⁶⁷ The dismissive discourse of "greasy mechanics," allegedly promulgated by those whom C. P. Snow described as "cultural Luddites" in the nineteenth-century canon, could be weighed against the more everyday discourse of machinery in the broader literary and print culture of the period.⁶⁸

Machine Affinities

This final section outlines an approach to seeking affinities between our model predictions and concepts of historiographical interest. Observing how our model systematically predicts masked words incorrectly enables us to explore certain characteristics of the corpus. Consequently, it is interesting to classify what most confounds the model, because it provides a marker of surprisingness. To recap: based on its training data, if the model expects the masked term to be "boy," "girl," or "slave" with a high degree of confidence, when the term is in fact "machine," this counts as a surprising sentence, which warrants close reading. As proximity within semantic space is a marker of affinity, we collated all the tokens predicted in each masked machine sentence and computed an affinity score that gauges its content relating to each of thirteen concepts of our choosing.

In relation to the concept of child labor, for example, this method identifies those sentences for which the model predictions are located in the same semantic neighborhood. Examining these sentences together reveals a rich array of material, such as in the following example from the BLB corpus, which has conceptual affinities with both "child" and "workforce," with scores between 0.8 and 0.9. This sentence, in which the model incorrectly predicts the token "children" (instead of "machine") with high confidence (0.75), is one of the most surprising in the child category. It comes from a firsthand account of the celebrated New Lanark—a utopian industrial colony on the River Clyde in Scotland—that nonetheless continued to employ "barefooted" and "ragged" child labor. "The work of these children, as of all the others employed at the Cotton Mills, of all ages, and sexes, is comparatively easy; they have only to set the [MASK] to work, watch them, and feed them with a constant succession of materials."⁶⁹ This sentence is an instructive example of how the language model appears to work. Notwithstanding the prevalence of child labor, in predicting "children" behind the mask, it more readily expects children to be in the company of other boys and girls than machines. The model prediction is perhaps strengthened by the way the author also anthropomorphizes the machines, figuring them as working and eating beings. Similarly, we can explore the concept of female labor by using the same method to identify relevant sentences and exploring in depth the reasons for the model predictions. The following sentence comes from a wildly popular and whimsical Victorian novel: "Four sewing [MASK] stood near the wall where grated windows admitted sunshine, and their hymn to Labour was the only sound that broke the brooding silence."⁷⁰ The most confident prediction for the masked word was "girls" (0.67), followed by "women" (0.19), then a long tail including "children," "men," "boys," "ladies," "hands," "maids," "mothers," and "sisters," but surprisingly no "machines." Like the previous example, we note an element of atypical animacy at play in these mistaken predictions. The sewing machines are figured as alive by virtue of their "hymn to Labour," which the model may pick up on, and that the gerund "sewing" appears to be collocated overwhelmingly with girls and women.

In addition to women and children, we used word embeddings to create eleven more concepts of interest, including work, transport, the body, and animals, reflecting our thematic interests and offering the prospect of coarse-grained search. Other concepts could be used for particular research interests by selecting a basket of tokens related to a particular occupation or technology. Our approach is well suited to detecting instances where one term can substitute for another over time, as when a machine replaces a human. This follows from the nature of the masked language model, which explores possible substitutions between words and hence the extent of their synonymy within the culture. As shown above, this works particularly well in cases where laboring bodies have been subject to a discourse that commensurates them with machines.

Our method is well suited to exploring forced labor and the disturbing ways in which the discourse of slavery interoperated with that of machines. The notion of a machine being comparable or interchangeable with a slave is found in Aristotle and continued in the nineteenth century in economistic terms, which strike us as offensive. The sentences in our slave category include this example from an 1825 parliamentary debate on slavery, where the model predicts "slave" with 0.78 confidence: "It might be sufficient to reply to the latter argument, that the loss in the wear and tear of a slave is taken into the account at the time of purchase, and diminishes the purchase money, like the chance of death, or the wear and tear of a [MASK]."⁷¹ The prediction is less surprising given the repetition of "wear and tear," perhaps reproduced in the model, notwithstanding the jarring equivalence being made. A similar line of reasoning, based on an economistic argument for abolition, features in the following example from an 1862 newspaper report of a parliamentary debate on agriculture:

Our steam plough engineers are men, however, of world wide views. They aim not only to double … the produce of our own cold clays, but to supersede the hoe of the slave, and to drive slavery out of the tropics. Mr. FOWLER mentions South Australia, Egypt, Cuba, South America, and Porto Rico, as places whither he has sent his [MASK] to triumph. … So surely as they succeed there … the "great question of negro slavery," he says, "must feel the effect of steam cultivation" … and "so surely as the sun would rise to-morrow, so surely must slavery succumb before the more mighty but silent change."⁷²

The model predicts "slave" but without great certainty (0.16), ahead of "people," "sons," "brethren," and "colonists," but not "machines." This paean to the progressive power of mechanization to sweep away slavery is compelling; but as the speaker was John Fowler, the inventor and patent holder of the steam plough, his claim could also be read as the familiar labor-saving rhetoric of the utopian or profiteering inventor seeking new export markets. The complexity of such texts in the popular press merits further study. Future work could explore how British factory discourse mobilized notions of wage slavery during the American Civil War, and so explore the political complexities of northern England's cotton famine.⁷³

A final example illustrates the legacy of the slave metaphor in seemingly unrelated contexts. An obscure volume of late Victorian essays contains the claim that "All wealth may be due to labour, but it is due to the trained skill that directs the labourer what he is to do. Until the labourer thinks, he is but a [MASK] in the hands of the capitalist."⁷⁴ Though the word "slave" does not appear in the context of the sentence, it is the second-ranked prediction (behind "tool" and ahead of "child" and "puppet"). That the model finds semantic traces of slavery demonstrates the powerful rhetorical triangulation of the master-slave-machine triad, describing the balance of power under Victorian capitalism. This language is rooted in a host of overlapping and complementary forms of historical exploitation, which produced comparative machine figures for making sense of modern industry and its novel forms of waged and coerced labor. A further legacy of historical slavery can be found in computer science and engineering discourse, where the master-slave-machine metaphor lives on despite longstanding efforts to remove it from the language of these disciplines.⁷⁵

This article has shown some of the ways that neural language models can be used as a targeted form of search for figures of speech not discoverable any other way. The results of this procedure in conjunction with close reading and contextualization have the power to illuminate well-known themes in the history of technology. Our approach allows us to integrate vast amounts of new source material of the kind not traditionally used in cultural or conceptual history. This exploration shows the potential to generate much future work in relation to the language of machines during industrialization. Besides the path taken here, other possible lines of inquiry suggest themselves, particularly around the language of work and the ways it has been figured discursively over time.⁷⁶ The question of who was expected to do the work was a politically fraught issue in our historical texts and remains so today. That a technique using "artificial intelligence"—a technology cast as a living machine in the popular imagination—is well suited to exploring this theme does not diminish the inherent irony this involves.

Conclusion

An affordance of using language models to make sense of a large text corpus includes their potential to capture change over time. The examples here involved models trained on texts from the early Victorian period; however, the date range for training can be restricted to specific periods, creating models corresponding to particular time slices. Taking the sentence by John Ruskin mentioned above, but this time comparing the predictions by two different models, gives a sense of what is feasible (see figure 5): "whether the workman shall be a living, progressive, and happy human being, or whether he shall be a mere [MASK]." A model trained on texts from 1760–1850, which we might call "Revolutionary BERT," made the predictions in the left-hand column, whereas a model trained on texts from 1890–1900, which we might call "Fin-de-Siècle BERT," made those on the right.

While several tokens overlap, the earlier model imagines Ruskin to be talking about slaves, whereas the later model correctly predicts "machine," with "slave" in nineteenth position. There is a lot to unpack in this example, but it is striking that such a degree of change is detectable even in a narrow time window. An 1869 article in The Sun newspaper contains another illustration: "He lost his control over the [MASK] and it ran with great force against a low stone wall, against which Mr. Edwardes was pitched head foremost."⁷⁷ Whereas 1760–1850 BERT predicts "horse," 1890–1900 BERT predicts "bicycle." The story's headline, "A velocipede accident, which it is feared will be fatal—and if so it will be the first of its kind," marks it as

Fig 5.

Exploring Change over Time. By deliberately restricting the date range of the training material, we can create models which specifically represent a given period. This table shows the top five predictions made by two different models for the masked word in the sentence from John Ruskin, above. The left-hand column shows results for a model trained only on texts published in the period 1760–1850; the right-hand column shows those from a model restricted to 1890–1900.

historic, displaying a certain symmetry, its publication coinciding with the arrival of this revolutionary new machine, halfway between the time periods of the two models.

None of the techniques used here are free from critiques around bias and representativeness, or the hermeneutic opacity underlying word embeddings.⁷⁸ However, if used carefully and reflectively, in transparent and reproducible ways, these tools can help create better forms of search and analysis for the large and increasingly ubiquitous digitized text collections that will be used whether we like it or not. Historians of technology are especially well placed to engage critically with these tools—and the political economy underpinning them—thanks to our understanding of processes such as social shaping. Today's breathless language of "A.I." has uncanny resonances for scholars of nineteenth-century technology. Looking across the centuries from his vantage point in 1879, Samuel Butler's vision of present-centered enthusiasm has a familiar resonance if we substitute "language model" or "A.I" for Butler's "steam engine" in this sentence: "The simplest steam engine now in use in England is probably a marvel of ingenuity as compared with the highest development which appeared possible [to seventeenth-century engineers], while our newest and most highly complicated engines would seem to them more like living beings than machines."⁷⁹