Read an Excerpt from “Divination Engines: Natural Language Processing, Artificial Intelligence, and the Making of Algorithmic Culture” by Xiaochang Li

Since the advent of computers, society has fantasized about conversing with machines. In her eye-opening book Divination Engines: Natural Language Processing, Artificial Intelligence, and the Making of Algorithmic Culture, technology expert Xiaochang Li shows readers how that dream both fueled the demand for data and set the stage for today’s generative AI. With original research and clear explanations, Li elucidates the origins of what’s known as natural language processing (NLP) and the heated twentieth-century debates between computer scientists, linguists, and communication engineers that shaped today’s technology. Starting with early devices that recorded, analyzed, and attempted to interpret human speech, she demonstrates how computer speech recognition, particularly efforts led by Bell Labs and IBM, advanced technology by deemphasizing linguistic meaning in favor of statistical prediction. In other words, researchers gradually abandoned systems that sought to understand human language, opting instead for workarounds that simply predicted patterns in speech and text data. That solution became incredibly and surprisingly adaptable. As Li reveals, transforming linguistic questions into engineering ones ushered in the routine operation of search engines, spam filters, and the varied content sorting and recommendation mechanisms that regulate the access, circulation, and legitimacy of information across every platform. But this has all come at the cost of forever requiring copious and ever-growing amounts of new data.
 
At its core, Divination Engines illuminates how the artifacts of human communication—speech, text, and images—have become both the fodder for and products of computers. This connection between communication and computation, Li shows, has given rise to data-driven analytics, machine learning, and today’s algorithmic culture.

Read on for an excerpt from the Introduction of Divination Engines.  

---

In November 2016, London Mayor Sadiq Khan kicked off a press event featuring Google’s chief executive, Sundar Pichai, with a small jest. Later reported in The New York Times, the joke went something like this: “A friend, he began, had recently told him he reminded him of Google. ‘Why, because I know all the answers?’ the mayor asked. ‘No,’ the friend replied, ‘because you’re always trying to finish my sentences.’” The punchline rests on a misidentification that appears at once ridiculous yet all too plausible. Anticipating a compliment, Mayor Khan blithely conflated the core functionality of Google’s search engine— which sifts through hundreds of billions of indexed web documents in response to user queries to select and rank the most “relevant” results—with its far less sophisticated “auto­complete” feature, a superficial interface element that attempts to predict the completed search query as it’s being typed. The mayor believed himself to be proffering knowledge; his friend knew that he was just guessing at words.

Considered in hindsight a decade on, the confusion between providing answers and predicting text no longer seems like much of a joke after all. Text completion was introduced as a default setting across Google’s search platform in 2008 under the name Google “Suggest” and was initially promoted as a dynamic reference index, which conveniently offered spelling corrections and recommended keywords to search queries in a drop­down list. But in 2010, Suggest was rebranded “Autocomplete,” and the language of reference guides and spelling assistance was dropped from official product descriptions to reflect a new emphasis on “predictions, not suggestions.” This rebranding was part of Autocomplete’s integration with “Google Instant,” a new incremental search feature that continuously updated search results as the user typed, which replaced the drop­down list of suggested keywords. Autocomplete and Instant, while technically separate functions, were merged under what Google portrayed as its newly predictive search experience.

This seemingly minor interface change was a decisive move to reconceptualize the user experience from one of suggestion to one of prediction. In previous iterations of Google Suggest, not only were possible text completions offered as a drop­down list that the user could select from or ignore, but each suggested query was also displayed alongside the number of search results it was expected to retrieve. Doing so presented the user with a glimpse, however superficial, into the internal logic of the search index, which had positioned Suggest as a guide for using the search function to its full effect. By eliding the offer of suggestions in favor of directly returning search results as the user typed, based on the anticipated text completion, the integration of Autocomplete and Google Instant worked to obscure the process of information retrieval, giving only the appearance of answers in place of options. From the user’s perspective, query suggestions and search results were consolidated into a single operation. At the Google Instant launch event on September 8, 2010, Marissa Mayer, then Google’s VP of search products and user experience, presented Google Instant as “a fundamental shift in search,” clarifying that Instant search wasn’t “search­as­you­type, but search­before­you­type.” Google Instant, in other words, provided results not to partially typed queries but to their predicted text completions—no longer an index, but an oracle.

The deliberate interdependence cultivated between information retrieval and text prediction has only intensified with the recent onslaught of generative AI, as a growing number of users turn to Chat­GPT and its many Large Language Model (LLM) brethren in search of answers. These systems, which generalize linguistic patterns from ever­greater quantities of data, are capable of producing strings of answer­like text that appear nearly indistinguishable from the real thing. These tools have ushered in a new epistemological uncanny, generating sophisticated text predictions that readily assume the familiar posture of knowledge to an extent both remarkable and unsettling.

In just a few short years, generative AI chatbots have begun to visibly encroach on the search engine, prompting sensationalized hand­wringing over if and when they may take over the role of search platforms as the presumptive method of gathering information. Whether chatbots realistically posed an imminent threat to Google Search in all its monopolistic splendor, the company opted to preempt the issue by simply making the search engine behave more like a chatbot. In 2023, Google introduced a new “Search Generative Experience” that integrated a custom version of its own ChatGPT competitor, Gemini (then Bard), “to reimagine and expand what Google Search can do.” This was most visible in the new AI­generated summary that was prominently displayed above actual research results, all but completely overshadowing them. Despite claiming to have “meticulously honed our core information quality systems . . . so you can get information you can trust in the blink of an eye,” Google’s new “AI Overviews” feature was as prone to regular “hallucinations”—fabricated or inaccurate content—as its chatbot counterparts. Following on the work of Google Instant before it, AI Overviews collapsed the distinction between finding answers and finishing sentences more brazenly than ever.

Over a century earlier, in 1906, Russian mathematician Andrei Andreyevich Markov added a new branch to the theory of probability that expanded its authority to the prediction of successive, dependent events. Markov’s concept of chain dependence calculated the probability of dependent events that were strung together like links in a chain, crucially introducing temporality to the prediction of random events. In other words, Markov processes estimated the probability of any result in a stochastic series with respect to the one that immediately preceded it, at each point calculating the chances for what would follow next. Chain dependence thus rendered probability explicitly predictive, offering statistics a new sense of the future.

Over the ensuing decades, Markov processes were taken up by mathematicians, physicists, economists, and engineers, opening a vast new array of real-world phenomena to probabilistic speculation. The concept also spawned a growing assortment of algorithmic cognates, such as Markov chain Monte Carlo, hidden Markov models, and Markov random fields, that appear throughout computational analysis and simulation tasks, particularly in artificial intelligence and machine learning. Yet, when Markov himself conducted statistical trials to verify the theory of chain dependence in 1913, he chose a rather curious source of empirical data. He demonstrated his eponymous chains using a copy of Alexander Pushkin’s Eugene Onegin, applying them to the prediction of successive letters in the text.

Although text prediction served simply to illustrate a general mathematical principle, the suggestion that language could be subject to probabilistic interpretation had wide­ranging implications. Whatever motivated Markov’s initial application, the use of chain dependence to analyze language gained fresh urgency through the wartime confluence of cryptanalysis and early electronic computing. Of particular consequence was Markov’s influence on information theory through the contributions of Claude Shannon, who applied Markov chains to formulate a mathematical theory of communication for the encoding and transmission of speech and text. Shannon’s work established a means to express information in quantitative terms, measurable in discrete “bits” that could be enumerated, forming the bedrock of signal processing, telecommunications networks, and digital computing. In recent decades, Markov processes have grown ubiquitous, used for computational modeling across diverse forms of knowledge production, generating everything from protein sequences to financial instruments to, rather fittingly, web search rankings. A new province of computational knowledge was laid open, in other words, through the very task of guessing at words.

This book tells a story about what happens between these two moments: a story about measuring speech and predicting text, communication and computing, and the quest to bring language under the authority of algorithmic processing. Broadly defined as an applied branch of artificial intelligence, natural language processing (NLP) describes a subdomain of computational research dedicated to the analysis, interpretation, and generation of speech and text. It is distinct from word processing programs, which aid in the digital production of text documents by allowing typewritten input to be formatted, manipulated, and stored as numerically encoded character strings. Language processing tasks, in contrast, seek to make text input legible to computer processing in its capacity as language. Today, NLP comprises a diverse assembly of applications of varied purpose and sophistication, including voice assistants and chatbots, machine translation and transcription services, and text and sentiment analytics tools that inform predictions about everything from purchasing habits to public health. Aside from its more overt applications, NLP also serves as an essential facet within the operational infrastructure of the digital knowledge economy, enabling the unassuming and omnipresent churn of search engines, spam filters, and content sorting and recommendation mechanisms that regulate the accessibility, circulation, and legitimacy of information across every platform.

In recent years, NLP has surged into the public spotlight with the release of OpenAI’s GPT­3 language model in 2020, which signaled a massive leap in the scale of deep learning and prompted a sudden deluge of much­-publicized “generative” AI tools that seemingly transformed the digital information landscape overnight. Despite what the breathless headlines and staggering financial investments might suggest, however, NLP’s position as machine learning’s defining task domain is not a new development. Artificial intelligence has been infatuated with the possibility of natural language processing long before the formal advent of computational linguistics, and its continuous pursuit has shaped how the parameters of human and computational reason were imagined and implemented. Its recent rise in public standing aside, NLP has in fact played a constitutive role in the emergence of machine learning, predictive analytics, and related forms of “data­driven” algorithmic decision­making as a privileged and pervasive form of knowledge work. Taken together, the mayor of London’s prophetic joke about failed conjecture and the Russian mathematician’s unlikely source of statistical data bring into focus a powerful and strangely persistent reciprocity between algorithmic knowledge production and the deceptively superficial work of predicting text.

---

Xiaochang Li is assistant professor in the Department of Communication at Stanford University and affiliate faculty in the Program in Science, Technology, and Society and the Program in Modern Thought and Literature.

Divination Engines is available now! Save 30% when you use the code UCPNEW at checkout on our site.