Abstract
This article showcases a speculative methodology for recreating interactions between a human and Google Search’s Auto-Predict interface as conversations, to explore how AI-based systems are both persuasive and deeply personal. Using ethnomethodology tools and a symbolic interactionist lens, the paper presents three versions of a single Google search, each variation building a slightly different angle on the plausible utterances and interpersonal dynamics of the human and nonhuman partners. This thought experiment emerges from a decade of classroom-based digital literacy exercises with young adults, training them to analyze their lived experiences with digital media, algorithms, and devices. Transforming information exchanges into personal conversations provides a creative method for analyzing how relations are co-constructed in the granular processes of interaction, through which mutual intelligibility is built, meaning about the world is made, and identities are formed. This critical analysis extends methods for human–machine communication studies and elaborates notions of algorithmic identity.
Keywords
Introduction
On the surface, micro-level AI entities like Google’s auto-predict or Netflix’s recommendation systems are helping humans make decisions. Likewise, through their human responses, people are helping AI agents get better at their job. Under the surface, these entities are increasingly enmeshed in the enactment of everyday life, operating less like software systems and more like life partners. Conceptualized as intimate interpersonal partners, how do symbolic meanings emerge in these interactions, and with what influence on a human’s decision making and self-identity formation? How do these micro-entities make their human counterparts feel about themselves?
I’ve been asking young adults in classrooms and workshops these questions for years, as part of a now long-standing critical digital literacy project where participants conduct what I’ve called “guided autoethnography” (Markham, 2018, 2022) of their own digital lived experiences. If the auto-predict functionality of Google Search was analyzed as a person with whom you are in an interpersonal relationship, what sort of person would it be? The aim of this article is to showcase one technique emerging from this project, speculative recreation of a single Google Search as a plausible yet imagined conversation between the human and the nonhuman agents in the interaction, focusing on interpersonal aspects of the relationship. Through presentation and discussion of this thought experiment, I extend and specify recent discussions about the interaction patterns and persuasive aspects of human–machine communication (e.g., Coleman, 2021; Dow Schüll, 2018; Guzman, 2018; Guzman and Lewis, 2020) by focusing on the granular level of interactions and the symbolic interaction processes therein.
The analysis demonstrates the heuristic value of creatively reverse engineering swift and seamless human–machine interactions, particularly as questions of relationality, influence, and manipulation gain more relevancy in the era of conversational types of generative AI. That is, if features, defaults, or affordances of interfaces invite users to respond in particular ways, we can ask: how are these invitations phrased, especially when they are not phrased directly? If self-tracking or quantification devices are “rhetorically energetic” as a “means, a medium, and a cipher for human experience, self-understanding, and sense of agency” (Dow Schüll, 2018), how does this rhetorical energy manifest symbolically in the human/tech encounter? If ubiquitous technologies like surveillance cameras can function as teachers (c.f., Duffy and Chan, 2019), what is happening in the micro-moments of the lessons? From a symbolic interactionist perspective, we would say that at some point, there is an encounter and interaction, and here, one can explicate how the features and patterns of these moments are functioning symbolically, affectively, and persuasively.
The article proceeds by first providing background of this speculative methodology as it has been developed through a long-standing digital literacy project, as I work with young adults in the university classroom. The method is specifically situated within symbolic interactionism with an inspirational nod to the granular focus and techniques of conversational analysis. The overall argument of this piece, including the justification and analytical lens, is positioned within two domains: first, relational theories of the self, emphasizing both classic symbolic interactionist concepts and more contemporary discussions of algorithmic identity, and second; critical data studies that focus on how infrastructural elements of AI-driven platforms like Google Search have agency and influence. I then present the speculative methodology through three different variations of “conversation” between a Searcher and Auto-Predict, to illustrate some of the ways the technique can work as a tool for exploring how machinic agency is enacted interactionally, specifically as a form of symbolic interaction.
Background of the overall project and methodology
This analysis emerges from a long-term critical data and algorithm literacy project with young adults in higher education classrooms and standalone workshops. Through what I’ve termed “guided autoethnography,” young adults (mostly students) have engaged in a variety of activities to track and reflect on their lived experience of and with digital media and devices. The overall project has involved many types of exercises and experiments since 2012, conducted under the general umbrella approach of participatory action research. To date, more than 1600 students have participated, from bachelor, master, and PhD level courses primarily in the United States and Denmark. As a brief note on how ethics are built into the design of this project: Students were aware at the outset of the course that these exercises were part of a long-standing project. They were assured that none of their materials would be archived or used unless they elected to become “study participants” after the course ended. They were required to conduct the guided autoethnography but given many alternative ways to perform, record, and submit their analyses. None of the materials they produced and submitted during the course were automatically archived for further analysis. Instead, after courses ended and grades were finalized, some students voluntarily became participants by donating their data to the project and went through informed consent processes. Elaboration on the larger setup of this project, including exhaustive details of the ethical considerations, can be found elsewhere (Markham, 2022; Markham and Pronzato, 2023).
The content of this article is inspired by a small part of this larger project, when participants have been invited to recreate a moment of an interaction with Google Search as if it were a dialogue, focused on the personality of the “Auto-Predict” function as a nonhuman participant in the conversation. Below, I present a thought experiment in three variations, whereby a single search is unpacked and explored as an interpersonal interaction between the human Searcher and the nonhuman micro-entity, Google Auto-Predict.
This creative-analytical technique is inspired by Lucy Suchman’s (1987) extraordinary use of ethnomethodology and Conversation Analysis (CA) to unpack the human–xerox machine interaction. Here, I use the most general of CA techniques rather than more specific conventions of notation to slow down an interaction to the level where it can be examined as a series of two parties taking turns. “Turn-taking” is the term given to what sociologists and linguists recognize as one of the basic protocols that organize interactions (Sacks et al., 1974). In reviewing this and other interaction-organizing aspects of robot-human interactions, Skantze (2021) returns to the basics to remind readers that “[s]ince it is difficult to speak and listen at the same time, speakers in dialogue have to somehow coordinate who is currently speaking and who is listening” (p. 1). As a basic foundation, I have found turn-taking to be a useful starting point in an ethnomethodological project 1 of reverse engineering the human–machine interaction to reconstruct a speculative conversation. When added to other forms of analysis, this focal point can tease out patterns or power dynamics. In this project, when one of the partners is a nonhuman whose voice is imagined, the technique is intended to be more generative and exploratory than accurate. Participants in this project have noted that this lens helps them clarify what happens in the micro-moment of the interaction and more importantly, also elicits new feelings about the relationship that were previously not recognized.
We know there are deeply meaningful interactions occurring between humans and machines that function symbolically on people’s sense of self as well as their understanding of information ecologies gives sign evidence of this. Of course, since these influences are situated within ongoing and often tacit interactions between human and machinic actions, as well as within larger dynamic sociotechnical ecologies, these need to be surfaced and translated into forms whereby the kinetic or rhetorical energies can be analyzed. Any time a person is part of an informational exchange system, we can analyze various parts of this interactionally, using utterances as the focal point to explore what is being said and with what possible impact on the person, especially when this is not an actual utterance. To address the question of whether it is a viable move to translate machinic actions to utterances, we can say that the complications of this translation work are similar to those associated with asking a person to translate what they think, see, or feel into words. In the effort to extract meaning-making from ongoing processes, there is a lot of slipperiness, arbitrariness, and fuzziness. This is not a problem, since the effort is not comprehension but consideration, to make sense of what is plausibly happening and to follow the ethnographic impulse to explore how it happens when it happens.
This lens is usefully applied to responses of platforms to our taps, swipes, or clicks, the infrastructural properties that manage or modulate our messaging. The results of, or responses to, our actions are communicative, uttering messages at the direct level and indirect, or meta, level, which means they become meaningful “participants in a continual symbolic interaction process whereby our understandings of self, other, and our social worlds are co-constituted” (Markham, 2015: 5).
Returning to the overall project within which this experiment is situated: The exercise is a prompt to facilitate greater critical data literacy by generating reflexive analysis of the Self, speculative analysis of the technology, and critical analysis of the relation. First, assigning interpersonal characteristics to nonhuman or machinic interlocuters raises many critical engagement points for participants. Second, the method requires the participant-as-analyst to slow down fleeting interactions that have significant impacts on what might be later described as a relationship. The technique affords scrutiny on how this relation does not simply exist but is an ongoing accomplishment, by both machine and human, at the granular level of the communication process. An interpersonal communication lens, developed within the field of communication studies, “is uniquely positioned to provide an understanding of the relational aspects of people’s interactions with machines that function as subjects designed to form relationships with people” (Guzman, 2018: 13).
The case of Google Search’s Auto-Predict function is used here because it operates visibly as a series of utterances, useful for detailing some of the communicative processes at the level of the conversation. Although many other aspects of the situation can be considered “communicative,” like browser and device, visuals and design, language settings, previous use patterns, personal history, and so forth, I bracket these in order to focus attention on the value of exploring sociotechnical relationality as emerging through something we can operationalize as exchanges of utterances between humans and AI systems. This example of Google Search’s Auto-Predict, a by-now typical and embedded style of AI, is ever more relevant for studying how sensemaking happens through overtly conversational AI forms, like ChatGPT.
Before proceeding to the examples of reconstructing Google Search as conversations, I delve into some basic symbolic interactionist premises for understanding the human–machine relation as this connects to identity formation of individuals. I then situate the argument more centrally in recent critical data studies discussions of agency and infrastructures, focusing on the well-known problems of platforms curating and filtering knowledge, specifically Google.
The human–machine relation
There is by now a long-standing body of literature that algorithmic systems are not only persuasive, luring, training, or entrapping users in subtle ways, as Seaver (2019) adroitly encapsulates in his work, but also transformative; that is, they are changing the nature of being. This is a point that posthumanist scholar Katherine Hayles has documented evocatively for many decades. If, as Hayles wrote back in 2005, the human is a “hybrid creature that enfolds within itself the rationality of the conscious mind and the coding operations of the machine” (p. 192), a still open question remains: How does this enfolding happen when it happens? Hayles (2017) makes an implicit call for exploring these enfolding processes at granular levels to locate the relevant inflection points (p. 204). I suggest that the concepts of symbolic interactionism are well suited to identifying these inflection points as critical junctures in identity formation processes. The exercise of building relational frameworks emphasizes the collaborative, co-constructed, and symbiotic aspects of human–machine configurations.
Relationality is by now a well understood concept for describing identity formation as well as the social order. Particularly in fields influenced by the symbolic interactionist tradition, the Self is interactional and relational. What might be considered a relatively stable Self-Identity, often just labeled “self” or “identity,” is a continuously repeated performance that is both relational and social, involving responding to the continuous response of Others.
Focusing on the Self, used synonymously here with “self-identity,” from a relational lens brings attention to how one’s sense of identity is never static or complete, but rather a continuous negotiation over time, in response and relation to Others, or other elements of the general environment or specific situation. Identity formation has been understood to be dramaturgical, not least because culturally-situated norms set the stage for particular performances or “interaction orders” (Goffman, 1959, 1983), and provide the “frames of reference” within which we find ourselves acting and reacting (Garfinkel, 1967). Whatever we understand to be our identity is a momentary sense that must be continually repeated to be maintained (Butler, 1990), as it is relational, an ever-shifting assemblage emerging in continuous (inter)actions, in situations. This constant dynamic of identity-in-the-making mostly goes unnoticed unless some anomaly disrupts the “natural” flow, or relative balance in one’s immediate environment. This brief definition of self-identity draws on a number of classic symbolic interactionist theorists such as Cooley (1902), Mead (1934), Blumer (1969), Goffman (1959), and Laing (1969), as well as more recent scholars such as Judith Butler (1990)—whose work on gender identity deftly advances how core characteristics of self-identity are deeply social and relational, and Kenneth Gergen (1991), whose ideas of multiple, fragmented, or “multiphrenic” identities in media saturated contexts explicate the temporary and fragmented characteristics of the relational self.
Within studies of digitally-mediated identity, core features of social media use have highlighted for researchers how fragments of information cohere temporarily, configuring what one might experience as one’s own identity or understand as another’s. This occurs in more temporal and momentary assemblages. Theresa Senft (2012, 2015) discusses this as “the Grab,” a concept that is intended to suggest both the literal act of grabbing ahold of something, and a counterpoint to “the gaze.” Through this, we can understand identity as identity-for-a-moment, filtered through algorithmic systems to yield “profiles,” repeatedly but in variation of particularities. Various system processes and components are functioning with agential force in this process, with varying degrees of visibility. The more infrastructural elements can seem like banal system level responses that convey information that a message was delivered or received, an image was uploaded, a post was viewed and liked, favorited, or forwarded.
Despite their surface utility or behind-the-scenes functionality, these micro-entities operate actively as interactants with consequences on identity formation, as I have elaborated elsewhere (Markham, 2013, 2021b). Every interaction presents a series of options for defining and redefining the Self. In other words, what may seem like only a user’s reaction to the system or the system’s response to the user is an enactment of a particular reality, whereby the interactional partners are selecting what seems to work, discarding what does not work, and over time, retaining certain meanings, or ways of doing or being, which become everyday ways of being. 2 This process of sensemaking is entirely relational, but not exclusively between humans, as has been pointed out in various ways over the decades by scholars exploring the human-technology relation (c.f., Guzman, 2018; Haraway, 1991; Hayles, 1999, 2005, 2017; Miller, 1978; Suchman, 1987).
Focusing on the sequences and cycles of interaction—that is, repeated cycles of actions and utterances, responses, and responses to responses—is useful for pulling apart different aspects of meaning-making in the micro-processes of communication between entities. It can highlight patterns and tendencies of automated actions as conversational elements that are functioning with what Coleman (2021: 14) has called “rhetorical energy.” It can also help unpack how machinic entities operate as interlocuters, whether or not they are generating utterances.
The figure of the “Other” or the interlocutor of the interaction is worth explicating briefly here, as it has often been taken to refer to another person, but among symbolic interactionists, the Other has long been a generous figure. There are inevitably multiple types of Others in the moment-to-moment sensemaking that contribute to the ongoing defining of Self. These can be actual or perceived. After all, much of the communicative process occurring in situations of meaning-making is intra-personal, as one carries on sensemaking through self-talk, self-reflection, or other modes that involve comparing, considering options, and making sense. For example, this intra-communicative process—of positioning in relation as a form of ontological security or finding a cohesive narrative about the Self—undoubtedly adds historical Others who contribute prior knowledge about roles or norms; perceived Others, often as audiences or imagined interlocutors; and a whole host of “roles,” which in part derive from the Self’s perceptions about their own and others’ desired or actual personal or social status and responsibilities.
Extending the notion of algorithmic identity, coined by Cheney-Lippold (2011, elaborated in 2017), we can explore how the combination of datafication, plus large-scale data aggregation, plus algorithmic processing, generates materiality that is fed back to individuals to influence sensemaking about the Self. I’ve argued elsewhere that whether the relevant material is an individual unit of information or a temporary assemblage of meaning, this material outcome of an algorithmic process becomes a dynamic Other in the symbolic interplay of everyday interactions through which people make sense of themselves, as well as other aspects of the lifeworld (Markham, 2021b).
The mediation or modulation of interaction, as well as the outcomes of interactions themselves, generates the need for more complex frameworks for how this relational Other might be characterized. Hayle’s (2017) work is again instructional in building the idea that computational media are not just functioning with what we might label agential force, but with “nonconscious cognition,” whereby the machinic operates from different origins than conscious cognitive entities (humans) but with similar functions. This perspective can open trajectories for exploring how machines become intimate partners and collaborators, which helps specify how agency develops or plays out in situ.
Taking the relationality of machinic agency in a slightly different direction, Neff and Nagy (2018) develop the useful concept of “symbiotic agency” to “capture the dynamic nature of human and technological agency in human–machine communication, in which users simultaneously influence and are influenced by technological artifacts” (p. 103). Thus, anything we consider to be “self-identity” will be an outcome of a deeply intertwined symbiosis of both human and machine. Algorithmic, combined with datafication processes, add more complexity to what counts as interaction in this process, since there are many elements that will exert powerful influence in the process of sensemaking about one’s identity, in interaction with everyday multi-entity agents.
The development of the field of human–machine communication takes up a similar trajectory, as Guzman (2018) notes, addressing core questions such as, “whether and to what extent machines are able to be communicative agents in their own right” (Gunkel, 2018: 223), or “who or what communicates, who or what can be responsible for generating original content, and who or what occupies the position of ‘Other’ in social interactions and relationships” (Gunkel, 2018: 223). Generative AI heightens the importance of analyzing the specificity through which the human–machine interaction becomes a relational partner, impacting how humans think of themselves in sometimes obvious ways, but more importantly, in rhetorically powerful forms of “call and response” that people likely never notice.
Naming Google auto-predict as a conversational partner
Google Search’s automated suggestions, as augmented by the auto-predict system, provide a good example of a call-and-response interaction that functions as conversation between a human and a nonhuman micro agent. Google’s auto-predict highlights important core features of what are already becoming commonplace interactions with generative AI, where the back and forth of turn-taking occurs similarly to human–human conversations. This call and response can be both powerful and unnoticed in the microscopic accomplishments of Search: Each letter we type into a search box sends a ping that results in an ever-changing list, echoes in the form of suggested search terms and phrases, and whether or not we pay close attention, each carries informational value about the potential object of our search, or the relevant directions we might go in our search. Often, the appropriate phrase is suggested long before we have finished typing in the search box. (Markham, 2021b: 7)
Google Search more generally is understood as a powerful moderator of experience. Early work by Introna and Nissenbaum (2000) focused on how search engines like Google would privilege certain content through indexing and ranking practices. Later, researchers such as Vaidhyanathan (2011) Gillespie (2014), and Noble (2018a) developed keen analyses demonstrating that the algorithmic processes were anything but neutral. While the system is centrally based on “relevance” and the outcomes are promoted by Google as simply reflecting trending patterns, both the inputs and outcomes of the algorithmic system have been found to be not only capricious and manipulable (Gillespie, 2014) but also biased in many ways (cf., Noble, 2018b). This matters for many reasons, not least because search and discovery is central to contemporary society (Halavais, 2018: 227) and Google is arguably the world’s most powerful curator of information.
Auto-Predict is a now taken-for-granted part of the infrastructure of information retrieval and selection. The flood of suggestions for how one should complete their search phrase delimits and shapes not just the information, but also the user, in multiple ways beyond the immediate moment of the search itself. It invites users to act in particular ways and trains users to think in particular ways about search terms and searching as a process. Over time, as Vaidhyanathan (2011) argued more than a decade ago, people come to have a close and trusting relationship with Google that borders on blind faith, a reliance on Google Search to hold, locate, filter, and curate information, without much questioning at all. Among other ways of looking at this situation, Auto-Predict can be seen as a powerful Other in the interaction, functioning with a type of persuasion that builds from relational power. 3
This matters. As Miller and Record (2017) have argued, autocomplete in search engines “inevitably and irreparably induces changes in users’ epistemic actions, particularly their inquiry and belief formation” (n.p.). This influence moves from the individual to the cultural level as norms, trust relations, and dependencies emerge from the continuous repetition of this type of interaction between humans and Google Search in larger ecologies of everyday media use. These algorithmic processes function dialogically and, more specifically even, as dialogue, between human and nonhuman. Rhetorically powerful sequences of utterance occur that, if attributed to a person, would be considered proactive, instrumental, or manipulative. Karapapa and Borghi (2015), in discussing this quality from the perspective of liability and accountability, detail how the predictive qualities of search engines create a snowball effect. Auto-predict initially orients users in ways that direct them to certain informational outcomes, and when users click on these outcomes, this in turn influences the algorithm to produce more of the same (Karapapa and Borghi, 2015: n.p.). Turmerello (2012) calls this a “vicious cycle” of algorithmic influence, whereby, for example, “if Google autocomplete suggests a certain search term . . . people are more likely to search it; the more people search a certain term, the more likely Google autocomplete is to suggest it” (n.p.). This is not reflective but directive.
These critical analyses of the agential force of Auto-Predict in Google Search are usefully connected to the discussion of algorithmic identity above, whereby algorithmic processes of prediction function at a level of interaction that can be understood as discursive, persuasive, and personal. Cheney-Lippold’s (2011) early arguments about algorithmic identity focused on how individuals are identified through datafication and algorithmic processes, which would influence how they conceptualize their own status as object, product, or subject. A slightly shifted angle on this process emphasizes how, in these algorithmically-saturated situations, one also self-identifies through the process of the interactions themselves (Markham, 2015).
The following thought experiment operationalizes this idea to explore what happens when the human/nonhuman interaction is transformed deliberately into conversational form. It focuses less on what is said and more how it is said, to reveal plausible personalities or affective components of the rhetoric valences. The examples are followed by further analysis and discussion.
Recreating auto-predict as a conversation: three variations and analysis
A common, even banal, sort of Google search is presented below in three variations, each progressively more relationally explicit and involved. The search phrase is “masking in around elderly people” (the strikethrough text here is deliberate, since the edit from “in” to “around” comprised the actual search phrase). The point of Variation 1 is to demonstrate what can be revealed when one simply slows down the actual interaction to open up what is happening in each turn of the interaction. Variations 2 and 3 build on this same interaction by adding actual side comments and articulated thought processes of the human (the article author) and some plausible affective elements, personality traits, and commentary of the nonhuman agent (Google Auto Predict). Each variation is explained in turn.
Variation 1
Variation 1 highlights the algorithmically-driven machine responses to the human typing on a keyboard. To note, this is not intended as a precise analysis using the methodologies and conventions of Conversational Analysis. The excerpt simply delineates the designed interface interaction. The actual interaction, which took less than 2 seconds, yielded many more interruptions, but this is indicative.
To guide the reading: quotation marks are added to highlight these as utterances in a conversation. Interruptions are highlighted by a long em dash [–]. The Searcher’s inner-voice thoughts or actual vocal reflections to themselves are included, but deliberately not distinguished in font type here.
“M –”
“Moving to Florida”
“Moisturizing at night”
“Michelle Yeoh”
“a –”
“Maxing out your credit card”
“March Madness”
“Maslow’s Hierarchy of Needs”
“Macy’s”
“Masters of the Universe Action Figures”
“s –”
“Massage near me”
“Massage for sports injuries”
“Mashed Potatoes”
“Massive Attack
“king –”
“Masking tape
“Masking theory”
“Masking autism”
“Masking liquid”
“Masking”
[space]
“in –”
“Masking in Switzerland”
“Masking in airports”
[backspacing]
“around –”
“Masking around newborns”
“Masking around curves”
“Masking around the world”
“Masking around meaning”
“Masking around head and ears”
“Masking around electrical outlets”
“elder –”
“Masking around elders”
“Masking around eldercare centers”
“Masking around elderly people”
“Masking around elderly populations”
This interaction is not unusual at all; nothing stands out as unexpected, glitchy, or inappropriate about the autocomplete suggestions. I mention this because when Google’s auto-predict interface was still novel, or at least novel in how it glitched results, glitches in predictive suggestions were noticeable, not least because they were strange, sometimes humorous, often offensive, noticeably racist, and otherwise non-neutral. Pressured to fix the glitches and guided by the clarity and tenacity of scholars like Safiya Noble (2018a, 2018b), who spent years repeatedly pointing to overt and covert racism in Google search, Google’s Auto-Predict is now designed to be much more circumspect. For example, searching for “Republicans are . . .” or “Why do white people . . .” will, most of the time, return zero suggestions. Of course, the effort to minimize inappropriate search results is not the same as being neutral, not least because the function of Auto-Predict is to curate, as discussed by authors mentioned previously in this article, which steers users to think and respond in particular ways. New problems continue to emerge as scholars continue to explore Auto-Predict functions in use cases (cf., Chonka et al., 2023).
If I were to ascribe a personality trait to Google Auto Predict at this stage of the speculative exercise, I would say that it is trying to be helpful, even a sort of kindness, by giving shortcuts for the user to skip the typing and select one of the pre-chosen options. On the other hand, as Auto-Predict’s many suggestions are transformed into a conversation in this way, the helpfulness is accompanied by, or perhaps even accomplished by, a startling amount of interruption.
Variation 2
This variation pushes further into the question of personality to ask: if Google Auto Predict was a person, what sort of person would it be? The same search is repeated in Variation 2 below to dive deeper into this question. To explain how this speculative recreation was generated, I return briefly to the methodology underpinning the larger guided autoethnography project. Users first make screen recordings of themselves conducting a simple Google Search. They then view this screen recording and make a “think aloud” audio recording describing what is happening and how they feel. After some time has passed (sometimes hours, sometimes days, depending on the setup of the specific intervention), they produce a second audio voiceover, reflecting on what they think of the personalities of both the human entity (themselves) and the nonhuman entity (Google Auto-Predict), if both were imagined as “people.” After some further reflection on how they feel about the relationship between themselves and Google Auto-Predict, participants recreate the interaction as a written dialogue, transcribing Google Auto-Predict’s part of the interaction and adding their own inner or actual voice as dialogue, based on their recordings and reflections. They may also add more of the plausible/speculative “voice” of Google Auto-Predict, separating this from the actual statements made by Google Auto-Predict. For this article, I composed all three variations of the speculative dialogues, drawing on many years of reading student productions of written dialogues and my own experiences conducting the same exercise with them. The notation style herein is peculiar to me, in a style I find useful when I do this exercise myself.
To guide the reading: quotation marks are added to the letters that are typed into the search box. Interruptions are highlighted by a long em dash [–]. Three ellipses [. . .] indicates that the searcher feels there is a notable silence (absence of response). The Searcher’s inner-voice thoughts or actual vocal reflections to themselves are included (not distinguished by font style or type).
Ok, let’s Google something. My browser is open to the Google search screen.
Oh, I just noticed I call it “googling” instead of “searching.” That’s weird when I stop and think about it. I never think about it. I just call it that, I guess, because everybody else does.
Why are you called Google, anyway?
. . .
Never mind. That’s beside the point.
The other day I wanted to look up masking and wearing masks around elderly people and also the efficacy of various kinds of masks.
Ok Google, let’s talk about masking
. . .
So let me start typing it out for you:
“m –”
–How about Moving to Florida, Moisturizing at night, or Michelle Yeoh?
“a –”
–How about Maxing out your credit card? March Madness? or Maslow’s Hierarchy of Needs?
Or are you wanting to shop for something at Macy’s?
How about Masters of the Universe Action Figures?
“s,” and by the way, why are you mentioning Maslow? Am I such an academic geek that you would just recommend this to me?
–Massage near me, massage for sports injuries, Mashed Potatos, Massive Attack
“king –”
–Masking tape, masking theory, masking autism, masking liquid, masking.
Wow, thanks Google. Weird mix. I don’t generally stop to look at all the results like this and they’re totally weird.
. . .
“in –”
–Masking in Switzerland? Masking in airports?
Seriously? Switzerland? I would expect you to offer an option I’ve seen before. Why Switzerland?
What am I really looking for?
[backspace].
Sorry, I meant “around”
–Masking around newborns, masking around curves, masking around the world, masking around meaning, masking around head and ears, masking around electrical outlets?
Ok, I’m really glad you are not a person because in less than five seconds, you are driving me nuts!
“elder –”
Masking around elderly person
Yes! Thank you!
This variation foregrounds the liveness of the dialogue and gives voice to the participants in an interaction. At least at the surface level of the Searcher’s side of the interaction, more of the user’s constant intra-personal, or self-to-self communication is revealed, which also articulates non-uttered side comments to or about an interlocuter. The commentary expressed in Variation 2 illustrates a core theory of symbolic interaction: that communication always functions at two levels; the content and relational level. That is, in any interaction, there will be direct communication and indirect meta-levels of communication, where statements “speak to” ideas about the relationship itself (Watzlawick et al., 1967). Here, we begin to learn some things about how the interlocuters feel about the Other as well as this relationship overall, especially from the Searcher’s perspective.
As for the utterances of Auto-Predict, this is kept at a fairly mundane level. While punctuation is added along with a bit of enthusiasm, nothing more is deliberately added to augment the personality of the nonhuman. Even so, by simply transforming actions of a user and the responses of Google Auto-Predict into vocalizations, their personalities and their potential relationship start to be revealed.
In this example, Google Auto-Predict interrupts with continual suggestions for where the user should turn their attention, covering a vast range of possible informational directions, most of which are considered by the Searcher to be completely off-base. The Searcher expresses feeling confused, distracted, and annoyed by these interruptions. Google Auto-Predict seems relentless and equally enthusiastic about all possible directions. Google Auto-Predict seems to enjoy the guessing game. It excels at making instant and swift decisions. It also just blurts these out, whereas the Searcher seems to be plodding along, typing one letter at a time.
This is useful as an analytical exercise, to experimentally or iteratively surface certain features or patterns of communication exhibited by a nonconscious cognitive entity. It is also useful as a reflexive exercise to consider one’s own tendencies, reactions, and patterns, especially over time and development of habitual relations with a nonhuman entity. The results of such analyses are interesting to compare, since the experiences and depictions have interesting variance, but this is not included here as it is not the central point of this analysis. Still, it is worth mentioning that by far the most common description given to Auto-Predict by my participants between 2012 and 2020 is “helpful.” When this is attached to an anthropomorphized identity, common phrases would be helper, guide, investigator, explorer. Sometimes this helpful entity is described by participants as awe-inspiring, even godlike for its predictive abilities (“fountain of knowledge,” ‘godlike knowledge generator “the Information God”). At other times, Google’s Auto-Predict is qualified as “helpful, but. . .,” such as a “know it all,” (“knows everything,” “bossy,” “parental,” “always has an answer, even if it’s not the one I want”).
Variation 3
Variation 3 extends the interaction in an even more conversational style, which builds the personalities of the participating interlocuters more fully, and in this case, also highlights the interruptive characteristics of the Auto-Predict/human interaction. This variation also adds plausible meta level commentaries from both the interlocuters, which acknowledges that both the human and the nonhuman in this conversation might be annoyed. To clarify, quotation marks and italics have been added to separate what is analytically more related to direct/content level communication (quotation marks) from indirect meta-communication about the relationship (italics): 4
To guide the reading: quotation marks are added back into this variation to indicate imagined as well as actual statements by both the human and nonhuman. Italics are added to distinguish inner voice and thoughts of the human as well as imagined inner voice and thoughts of the nonhuman. As with Variations 1 and 2, interruptions are highlighted by a long em dash [–]. Three ellipses [. . .] indicates that the searcher feels there is a notable silence (absence of response). Material in ALL CAPS indicates an imagined raise in volume of actual, inner, or imagined voice.
“Why are you called Google, anyway?”
. . .
Fine, don’t give me an easy answer. Make me search for something. If you were Alexa, I could have gotten an answer already.
. . .
“Never mind. That’s beside the point.”
“The other day I wanted to look up masking and wearing masks around elderly people and also the efficacy of various kinds of masks.”
“Ok Google, let’s talk about masking.”
“Ok! I’m a blank slate, just waiting for you to ask me a question! I’m so excited my cursor is blinking!”
“I’ll start typing it out for you.”
“m –”
“What are you asking for? Huh? Huh? Are you trying to say ‘Moving’?”
“Let me guess! Are you trying to say ‘Moisturizing?’ At night, maybe???”
“Are you trying to say ‘Michelle Yeoh’???”
“Are you –”
“—a –”
“– Oh! How about Maxing out your credit card? Is that it?”
“Huh? Huh?”
“Let me guess again! March Madness?”
“Maslow’s Hierarchy of Needs? Or are you wanting to shop for something at Macy’s? How about Masters of the Universe Action Figures?”
“Did I get it?” Did I? Why is this human so unhelpful? “What are you looking for?”
Jeez, what’s with all the hasty interpretations? Just let me finish! And by the way, “Why are you mentioning Maslow?” Am I such an academic geek that you would just recommend this to me out of the blue?
. . .
Oh, so now you’re quiet. No answer for that?
“s –”
“Massage near me! Massage for sports injuries!”
“I still think it might be Maslow’s Hierarchy of Need!”
“I know! Mashed Potatoes!!”
“No, wait, maybe it’s Massive Attack! Yes!” That would be good to know about
If you would just get on with it, and type a little faster, I would be able to guess more accurately!
OMG. Really? Could you be any more RANDOM? “Mashed Potatoes?” Really???
Ok, now just hang on. Stop going so fast! Let me type four more letters. Don’t interrupt!!!
. . .
“king –”
Ok, finally! “‘Masking’—That’s easy. You finally made a word, so I can narrow down the possibilities! I’m pretty sure you want something like masking tape, masking theory, masking autism, masking liquid, masking.”
Wow, thanks Google. Whatever. Weird mix.
. . .
“in –”
“Masking in Switzerland! No? How about Italy? Japan? USA?”
“Oh, wait, I got it! Masking in airports!” Yeah, airports!
No.
Sigh . . . I think I am asking this the wrong way. I guess I need to be more precise.
What am I really looking for??
[backspace, backspace].
How about I try “around” instead of “in?”
“around –”
“Great! I can make that work! Masking around newborns?”
“Or . . .” let me process that “. . . Maybe you’re asking about photoshop masking? Like ‘masking around curves!’”
No? Yes? “Maybe you’re asking about masking around the world, masking around meaning, masking around head and ears, masking around electrical outlets?”
WHAT IS IT YOU’RE WANTING TO KNOW ABOUT???
Ok, I’m really glad you are not a person because in less than five seconds, you are driving me nuts! SHUT. UP!!!
“elder –”
“ELDERLY PERSON! MASKING AROUND ELDERLY PERSONS!!”
“Yes! Thank you!”
Seriously. You are so annoying.
You are so annoying! . . .. and boring! You’re not anywhere near enthusiastic enough at the whole world of possibilities I could have showed you!
. . .
“Um . . . by the way, here are some other search results I’m showing you now. It’s about 44,400,400 results (0.47 seconds)
. . .
This variation further develops the personalities of each entity and brings forward certain relational patterns. Readers will notice that the Searcher’s attitudes and perceptions continue to build and complexify from Variation 2, but the more remarkable transformation is in the interaction style of the nonconscious cognition entity. When put into a conversational form with quotation marks, Google Auto-Predict takes on more humanesque character and becomes even more obviously interruptive. Adding punctuation allows the Searcher to add a range of emotional qualities to the entity. This particular depiction of Google Auto-Predict uses a lot of question marks and exclamation points, which ramps up the affective intensity of the interaction. Auto-Predict is portrayed as excited most of the time, and as much as the entity is offering suggestions, it is also questioning whether this or that is the “right” or “best” suggestion. Google Auto-Predict appears easily excitable, bossy, and yet also needy. It is enthusiastic but also impatient. On the flip side, when we examine Variation 3 from the imagined point of view of Auto-Predict, it becomes rather obvious that the nonconscious cognition entity is incredibly, even endlessly patient with the human’s plodding progress toward figuring out what they want to search for, not to mention the excruciatingly slow typing, overly linear thinking, and limited variety of input.
Taken together, Variations 1, 2, and 3 are striking in showcasing some of the ways turn-taking occurs during a Google Search, and the disconnects or missed connections between the parties in the interactions. These features can help users identify what sort of relationship this is, within what seems on the surface to be an instrumental relationship, yet is also a deeply personal partnership. On what grounds, or logics, are the two partners interpreting the relationship? And for the human, how does this impact their sense of self? Take the illustrated extent to which an algorithmic system like Google Auto-Predict appears to respond to only those parts of the interaction that are technically informational. When a human types a letter or backspaces, this contains information Auto-Predict can use to respond. It then changes its response, offers new ideas for the user. Auto-Predict ignores pauses. However, behind the scenes, or as part of the larger ecosystem of algorithmic processing, Auto-Predict is responding to its coded injunctions to reference this search against other searches, by this user and others. Its response is based on other assumptions it is making about the user and the situation based on background information available about this user, including at least location, and likely many other markers, available through its own and third-party data sources, used to produce personalized content and targeted advertising. While the user might realize on an abstract level that this is occurring in the moment of search, this is most likely overridden by the priority to get the information one needs, as quickly and effectively as possible.
The responses are therefore not equal from both interlocuters, since the machinic is based on both an extremely literal frame of reference based on the searcher’s technical accomplishment of an action it can recognize and also a black boxed set of frames of reference that build a predictive logic about who this user is and what it wants. And there are countless factors influencing the Searcher’s actions, not discussed here. This is to say, a range of logics will inform the expectations and actions of each entity in this interaction.
While identifying the personality and conversational style of Google Auto-Predict is a productive part of this analysis and has proven quite satisfying to student participants in the larger project, the elements of these interactions function in ways that are not easily broken into component parts. Not discussed in this particular article, an important next step in the analytic process is to zoom in on specific trios of utterances between conversational partners, to identify and then analyze how cycles of sensemaking are not dyadic but triadic, involving an action, a response, and a response to the response. This is meant to essentially translate what Miller (2007: 147) calls the “kinetic energy of the rhetorical situation” into ever-moving cycles of conversational interactions so their symbolic properties and functions can be analyzed by participants. This is a valuable analytical move for human participants who want to make sense of how they are unconsciously interpreting the actions of AI as symbolic exchanges in their everyday lived experiences. This is challenging work, since whatever might be considered “influence” is not easily broken down into language or even interaction, but emerges “more broadly in the energies that are ‘at work’ in the interaction between entities” (Coleman, 2021: 14).
Conclusion
Given how much we interact with micro-entities like Auto Predict throughout our work, leisure, and home lives, it is vital to explore these everyday interactions from the perspective of symbolic interactionism and interpersonal communication, which opens the opportunity to explore what counts as utterance and how relations develop in conversational interactions between humans and their devices, software, or platforms.
Through micro-level cycles of human–machine communication, entities enact and gain agential force. Identities emerge, whether these are self-identities or identities-for-others. In the instrumental action of typing a search into a box on a screen, it may be difficult to see something that seems like a personality, much less a significant life partner. Generative AI is helping to change this picture somewhat, as it is easier to anthropomorphize. An interpersonal assessment of more mundane applications like Auto-Predict brings needed attention to how a machinic entity can develop as a communicator over time and interaction with individuals or multiple other entities. Even if an individual dismisses the auto-predict function in the moment of a Google Search, it communicates, through direct utterances, not just in the form of suggestions but in presenting ideas that become frames of references about the nature of the relationship. And people respond in kind. Over time, the nonconscious cognitive entity develops a “voice,” in that it is uniquely “heard” by its partner human user. They develop an interpersonal relationship. The micro entity need not be alive in the traditional sense to have liveness, or be acting with independent will to have agency. Agency is the outcome of a continuous interactive process, as well as an assignment of the attribute of intention, will, or control. The specific conceptualization will therefore shift back and forth in different types of interactions, or over time. As Carolyn Miller (2007: 137) writes, “Agency is not located with individuals, or agents, but is the property of the rhetorical event.” By becoming a relational partner and intimate Other, the machine in the human–machine communication interaction participates actively in a user’s development as a person.
Beyond Google, this is how many automated micro-helpers work. It is at the basis of generative AI. It is the very core of what makes personalization not merely effective but hegemonic, to the extent that its power is hidden beneath the surface of utility and function, taken for granted, and neutralized over time. The power is in the very invisibility of the process, which builds systems that are taken for granted as just the way things work, a process of “discursive closure” whereby one might be annoyed at the outcomes, or even the relationship, but alternatives are not really open for discussion, and radically different imaginaries are not considered possible (Markham, 2021a).
The heuristic presented here is just one way of using an ethnomethodological lens to explore the deeply symbolic interactions occurring between humans and machines. This lens can be developed significantly. Understanding how these interactions impact people’s sense of self as well as their understanding of information ecologies could be further strengthened by elaborating on other aspects of these interactions as speech acts or through the more elaborate tools of Conversational Analysis, or by exploring perceptions and actions within these interpersonal relationships over time, to examine how relationships develop, stabilize, or decline over time.
Through this methodology of speculative reimagining mundane human–machine interactions as dialogues, we can broaden our understanding of how relationality happens and can creatively explore the granular ways these automated features of digital technologies function as interactional partners that, over time, influence the baseline normative frameworks of reference for how we think. Reverse engineering contributes vitally to shaping working relationships with technologies that can serve mutually beneficial purposes.
Footnotes
Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.
