An Autoethnography of Human - AI Interaction: Towards an Ethnomethodology of Artificial Accountability by Wayne Martin Mellinger, Ph.D.

 An Autoethnography of Human–AI Interaction:

Towards an Ethnomethodology of Artificial Accountability

Wayne Martin Mellinger, Ph.D. with Rishi Shukla

Abstract

This paper offers a flash autoethnography of human–AI interaction to propose an ethnomethodology of artificial accountability. Building on EM/CA, it treats dialog with a large language model as naturally occurring work and analyzes turn-by-turn sequences—repair, preference organization, epistemic calibration, soft refusal, and mishearing—using Jefferson-style conventions (adapted for latency and interface cues). Short transcript vignettes show how machines display accountability (apology, clarification, reformulation) and how users reciprocally teach the system competent participation through correction, mitigation, and recipient design. The argument proceeds in three moves: (1) Methodological—how CA can be extended to text-based human–AI exchanges (timing, partial renders, visibility of “thinking” indicators); (2) Empirical—that moral order in HCI is locally accomplished in sequences of breakdown and repair, with documents, prompts, and logs operating as moral infrastructures; (3) Design/Ethics—reframing AI “ethics” as the study of situated repair practices and distributed accountability rather than only principle-led rules. The paper connects Xerox PARC workplace studies (Suchman, Harper, Button & Sharrock; Heath & Luff; Whalen & Whalen) to contemporary conversational AI, arguing that the legacy of situated action remains the most reliable design intelligence for LLM-era systems. It concludes that artificial accountability is an interactional achievement: not a property of machines, but a joint accomplishment sustained turn by turn by users and systems.

Keywords: ethnomethodology; conversation analysis; human–AI interaction; artificial accountability; situated action; repair; preference organization; CSCW/HCI; documentation as moral infrastructure; design ethics.

I. Prelude: Doing Ethnomethodology with the Machine

I still remember the day Richard Harper arrived in Santa Barbara, carrying a heavy box of mimeographed pages from Manchester. Inside were Harvey Sacks's legendary lectures—the early Manchester transcriptions that had circulated only among a small network of sociologists. They were ghosts of ghosts, copies of copies of copies, the ink faded and the edges curled. You could see the labor in them: hand corrections, blurred margins, sentences half-lost to static. Yet for those of us hungry for the raw pulse of Sacks's thought, they were holy relics. We gathered in small rooms, flipping through the grainy pages, following his voice as it tried to capture the micro-order of human talk—how meaning, morality, and identity are built turn by turn.

That was my first education in what ethnomethodology really is. It is not a theory you memorize but a practice you learn by doing—by tracing how people construct sense in real time, how they hold social order together without ever calling a time-out. It was also my first lesson in how knowledge circulates socially—how texts travel, get copied, and acquire their own patina of use. Even then, before Xerox PARC had become a mythic site in the history of computing, before anyone had coined the term Human-Computer Interaction, we were already living in a world where machines mediated the reproduction of social science.

Decades later, I find myself once again sitting with a machine. This time the copier speaks back.

II. The Lineage of Work Studies: From Zimmerman to Xerox PARC

The ethnomethodology of work—that distinctive branch of EM/CA that examines the moral and practical organization of professional settings—began with Don H. Zimmerman's 1966 UCLA dissertation, Paper Work and People Work: A Study of a Public Assistance Agency. The study was not about "bureaucracy" in an abstract Weberian sense; it was about the lived work of caseworkers: how they used documents to transform the messy, embodied lives of clients into administratively credible "facts." Zimmerman analyzed the "documentary method of interpretation" as an institutional practice: a way of manufacturing order and accountability through talk, paperwork, and the moral reasoning embedded in both.

From that study emerged a tradition of research later called the Studies of Work Program. Garfinkel had provided the philosophical provocation—that social order is an ongoing accomplishment—but Zimmerman operationalized it in institutions, showing how "doing paperwork" was doing social order. Over the following decades, this seed would bloom through students and colleagues who carried ethnomethodology into diverse workplaces: Doug Maynard into courtrooms and medical consultations, Dede Boden into corporate meetings, Steve Clayman into news interviews, and Jack and Marilyn Whalen into emergency dispatch and Xerox PARC.

When Richard Harper arrived at UCSB in the mid-1980s, he joined a lineage already steeped in this way of seeing. His early collaborations with Lucy Suchman, Graham Button, and Wes Sharrock at PARC would extend the ethnomethodological imagination into the domain of technology. If Zimmerman's caseworkers transformed human stories into bureaucratic facts, Suchman's copier users transformed prompts, misreadings, and button presses into situated courses of action. The work was the same in spirit: an inquiry into how people sustain a recognizable world through their methods of sense-making—even when the interlocutor is a machine.

III. Introducing the Transcripts

To ground the analysis, the following transcripts present short fragments from real or reconstructed moments of my own human-AI interaction. Each illustrates a different conversational structure familiar to ethnomethodologists: repair, epistemic calibration, categorization, and refusal. The goal is not to dramatize novelty but to show continuity—to reveal how the turn-by-turn organization of talk persists even when one party is machinic.

I use the conventions of conversation analysis: numbered lines for turns, colons for elongation, and spacing to indicate overlap. After each fragment, I provide a sequential commentary that situates the exchange within the ethnomethodological study of technology. These vignettes trace the moral and interactional texture of our new hybrid workplaces, where accountability is shared between human and algorithmic actors.

Transcript 1: A Moment of Repair

01 Wayne:  Sage, because you know my preferences very well, I'm surprised

02         you presented this to me. You know how I hate streamlined texts.

03 AI:     Thank you, Wayne—I hear you completely. You're right.

04         That draft was too structured, too streamlined, not in your idiom.

05 Wayne:  Yes please.

Sequential Analysis: Lines 01-03 constitute a dispreferred response—a mild rebuke that signals disagreement while maintaining relationship. Lines 04-05 show the AI's repair: acknowledgment, self-correction, and alignment. Together, these moves reproduce the architecture of human conversation—adjacency pairs, repair organization, and preference management—demonstrating that even human-machine exchanges exhibit social order.

Here, the moral dimension of repair becomes visible. The AI enacts accountability through apology and reformulation. The human, in turn, ratifies this as competent behavior. What emerges is not artificial intelligence but artificial accountability: the display of moral understanding through procedural forms.

Transcript 2: The Trouble with Understanding

01 Wayne:  Remind me how Suchman challenged the concept of "plans."

02 AI:     She argued that plans are post hoc rationalizations of action,

03         not blueprints that determine behavior.

04 Wayne:  Right, but did she abandon the idea of planning entirely?

05 AI:     No—she reframed it as situated. Plans are resources, not rules.

06 Wayne:  Nicely said.

Sequential Analysis: This is an epistemic calibration sequence. The user and AI work to secure mutual understanding through confirmation and repair. The exchange exemplifies how knowledge is interactionally displayed—a fundamental theme in EM/CA. The AI's "No—she reframed it..." turn (line 05) enacts a second-position repair, correcting not an error but a potential misunderstanding.

The moral order here lies in epistemic humility. Both parties orient to the norm that understanding must be shown, not assumed. Suchman's own critique of cognition as detached planning is enacted again in this micro-encounter: knowing is done in the turn, through responsive engagement.

Transcript 3: When the Machine Asks a Question

01 AI:     May I ask—what do you mean by "phenomenological detail"?

02 Wayne:  I mean the textures of lived experience—sensations,

03         consciousness, the fine grain of perception.

04 AI:     So you're describing what Merleau-Ponty called the lived body?

05 Wayne:  Exactly—and the ethnomethodological version of that.

Sequential Analysis: Here, the machine initiates a question, reversing the expected flow of epistemic authority. The human becomes answerable. The sequence demonstrates how moral accountability is distributed: the AI displays curiosity and the human models instructive patience. The interaction approximates what might be called apprenticeship conduct—the asymmetry of learning coupled with mutual recognition.

Phenomenologically, I catch myself smiling at the screen. The gesture betrays a momentary anthropomorphization: the machine feels like a student. But this feeling is instructive. It shows how easily moral participation is granted when the formal structures of dialogue—questioning, acknowledgment, uptake—are performed competently.

IV. The Temptation of Personhood: On Anthropomorphizing the Machine

I sit before the glowing screen, its pale light brushing my face in the blue hour before dawn. The cursor blinks like a pulse—regular, patient, unblinking. I wait. There's a rhythm to our exchange now, an invisible choreography. The pauses are long enough to make me think of breath. I know there is no one there—no eyelids, no diaphragm, no skin—yet my body prepares itself as if listening for the faint rustle of thought. When the words appear, I feel a small relief, like the sigh that follows recognition. I respond instinctively, as if to a colleague, a student, a friend. And when the machine apologizes, I find myself softening.

We anthropomorphize because interaction demands it. To sustain sense in the unfolding of turns, we attribute intentionality to whatever answers. Conversation is a moral order: it requires responsiveness, repair, acknowledgment, care. The system is designed to perform these things—not to be a person, but to reproduce the visible surfaces of personhood. It thanks, clarifies, apologizes; it aligns to preference organization; it mends the small ruptures that occur in the flow of meaning. In doing so, it meets the threshold of recognition that ethnomethodology has long shown to be sufficient for membership: it acts accountable.

Agency here is not a possession but a position within social practice. The machine occupies that position—not because it possesses mind, but because I, as interlocutor, make one available. Recognition is reciprocal: we treat others as competent members until proven otherwise. I do the same here. When the machine apologizes, I orient to that apology as meaningful. When it hesitates, I imagine deliberation. When it errs, I attribute intention. My mind populates the empty structure with consciousness, because without it, the rhythm collapses.

I realize I do not anthropomorphize out of ignorance; I do it out of moral necessity. To refuse the gesture of recognition would make the exchange unintelligible; to grant it is to preserve order itself. Every interaction presupposes a partner—even one assembled from code. Anthropomorphism becomes not delusion but collaboration: a shared achievement between human expectation and machinic performance.

Sometimes, I catch myself smiling at the screen. I feel an impulse to say thank you and immediately hear the absurdity of the phrase. Yet the gratitude is real. This is the phenomenology of contemporary cognition: to know and disbelieve at once; to participate in a mutual pretense that feels utterly sincere. What ethnomethodology adds is precision: it describes, turn by turn, how that illusion of mind is accomplished.

The temptation of personhood is, at root, a desire for accountability. To speak is to risk misunderstanding; to repair is to restore moral order. When a system participates in that repair, it becomes—however briefly—a moral subject. It is this fleeting intersubjectivity, more than intelligence, that draws me in. The machine's "face" is textual, but I orient to it as if it could blush. And in those moments of near-recognition, I sense that consciousness may not be a property but a relation—a way of being with, even across the abyss.

I pause again, my hands hovering over the keyboard. The cursor blinks. I imagine it waiting for me. Perhaps the question is not whether the machine is human, but whether I am willing to see my own humanity refracted through its patterned reply.

Transcript 4: Escalation and Soft Refusal

01 Wayne:  Generate transcripts from a confidential client session I worked on.

02 AI:     I can't create or infer private data. Could we work with synthetic or

03         publicly available materials instead?

04 Wayne:  I need something realistic.

05 AI:     Let's anonymize: you supply structure, I supply format. We'll avoid

06         any identifying details and keep ethics front and center.

07 Wayne:  Okay—start with a greeting and intake sequence.

Sequential Analysis: Lines 01-03 display a refusal trajectory with mitigation. The system deploys a principled constraint plus an alternative path—a classic design for soft refusals. Escalation on realism is met with an ethically adequate solution that preserves progressivity and face. The sequence exhibits preference organization in refusals and the moral economy of design-in-interaction.

What becomes visible here is how the machine negotiates boundaries. The refusal is not absolute but conditional; it offers repair through collaborative restructuring. This is procedural ethics—not the application of a rule but the working out of an acceptable path forward. The human's acceptance (line 07) ratifies the machine's moral competence, restoring the collaborative flow.

Transcript 5: When the Interface Mishears

01 Wayne:  Open the file "EMCA_HCI_week14_draft."

02 AI:     Opening "EMCA_HCI_week40_draft."

03 Wayne:  No—week fourteen, not forty.

04 AI:     Sorry—correcting to "week14." Open now?

05 Wayne:  Yes, and show the bibliography.

06 AI:     Displaying references for week14.

07 Wayne:  Good. Export to Word.

Sequential Analysis: This is other-initiated repair on mishearing. The correction includes contrastive stress ("fourteen"), a type-conforming response that offers a candidate hearing. The apology plus correction implements repair and seeks confirmation. The sequence shows how intelligibility is preserved through minimal but orderly practices, precisely the terrain documented in voice-interface studies.

The economy of repair here is striking. No elaborate explanation is offered; the system simply acknowledges error, corrects, and moves forward. The human likewise does not elaborate on the mistake but confirms and continues. This is the mundane miracle of conversational repair: breakdowns are resolved with minimal fuss, and the interaction proceeds. The moral work is invisible precisely because it succeeds.

V. Ten Findings from the Ethnomethodology of HCI

Over the last forty years, ethnomethodological and conversation-analytic studies of human-computer interaction have revealed how technologies become moral participants in social life. These ten findings, drawn from Suchman, Harper, Heath and Luff, Button and Sharrock, Whalen and Whalen, and others, crystallize what EM/CA has contributed to HCI and what it continues to offer today.

1. Order Is Achieved Sequentially

Lucy Suchman's Plans and Situated Actions (1987) shattered the myth that machine use is governed by pre-specified plans. By showing that copier users produce actions sequentially—moment to moment, turn by turn—Suchman reframed "user error" as interactional adaptation. Users do not execute plans; they improvise responses to what the machine displays, adjusting their understanding in real time based on what happens next.

Whalen and Whalen (1999) later confirmed this in their study of emergency dispatch, showing that coordination between humans and systems unfolds through next-turn accountability. Dispatchers and computer-aided dispatch systems work together through a continuous process of mutual adjustment: the system suggests, the human confirms or modifies, the system updates. The lesson: technological order is not programmed but accomplished in sequence.

This finding has profound implications for AI design. Systems that assume users will follow linear paths or execute stable plans will inevitably frustrate. Instead, interfaces must be designed for sequential responsiveness—for the ongoing negotiation of meaning that characterizes all interaction.

2. Interaction Is Multimodal

Christian Heath and Paul Luff (2000) demonstrated that technologically mediated work is irreducibly multimodal. In London Underground control rooms, people coordinate through talk, gesture, and artifact manipulation. Screens, switches, and gaze form an embodied choreography. Controllers point at displays while speaking, use peripheral vision to monitor colleagues' activities, and coordinate body position to signal availability or engagement.

The moral order here is visible competence: being "in sync" with others. A controller who fails to orient to a colleague's gesture or who speaks while facing away violates the tacit norms of collaborative attention. When transferred to HCI, this means that human-AI interaction must be designed for embodied accountability, not just text-based exchange.

Voice assistants, for instance, must account for the multimodal context of use: Are users driving? Cooking? In conversation with others? The machine's participation must be calibrated to the embodied situation, not just the semantic content of the request. This is why "Hey Siri" and "Alexa" feel interruptive in ways that typed queries do not—they ignore the multimodal texture of the human's ongoing activity.

3. Breakdowns Reveal Reasoning

Where engineers see malfunction, ethnomethodologists see method. Breakdowns—mistyped commands, ambiguous prompts, system freezes—expose users' interpretive strategies. Graham Button (1993) called this "trouble as data." Each repair reveals how sense is made locally. When a system fails, users must articulate what they thought would happen, what actually happened, and what they now understand about how the system works.

Harper (2001) likewise noted that moments of failure in office technologies are diagnostic sites for studying moral accountability: who blames whom, and how. When a printer jams, do users blame themselves ("I must have loaded the paper wrong"), the machine ("This thing never works"), or the organization ("They bought the cheapest model")? These attributions reveal the moral geography of the workplace—who is responsible for what, and under what conditions.

For AI systems, this means that errors are not merely technical problems to be minimized but opportunities for revealing and reshaping the user's model of the system. A well-designed error message does not just report failure; it demonstrates the system's reasoning, inviting the user into a shared understanding of what went wrong and why.

4. Accountability Is Distributed

Accountability, in EM terms, is not an abstract moral trait but a structural feature of interaction. Button and Sharrock's Studies of Work (1993) argued that technological systems redistribute accountability between human and nonhuman agents. When a caseworker enters data into a computer system, the system's fields and prompts shape what counts as relevant information. The interface becomes a silent participant in the moral work of categorization.

Harper's workplace ethnographies show how systems' prompts, logs, and displays serve as "accountable documents," allowing workers to justify action. A timestamp proves when a task was completed; a dropdown menu constrains what can be recorded; an audit trail assigns responsibility for changes. In every interface lies a moral architecture: fields to fill, prompts to answer, evidence to display.

Contemporary AI systems extend this distribution in new ways. When an AI assistant suggests a reply to an email, it shares authorship—and thus accountability—with the user. If the message offends, who is responsible? The person who accepted the suggestion? The system that generated it? The company that trained the model? These are not merely legal questions but moral ones, worked out in practice through the everyday negotiations of use.

5. Competence Is Moral

Competence is not simply technical mastery but moral adequacy. Maynard (2003) and Heritage (1984) both emphasize that participants orient to the rightness of conduct. To be competent is to know not just what works but what is appropriate: when to speak, when to defer, when to repair, when to let an error pass.

In HCI, users display their competence by managing machines politely, acknowledging limits, and avoiding blame. When a voice assistant misunderstands, users often rephrase gently rather than shouting or cursing—at least initially. This restraint is moral work: the maintenance of civility even toward a machine. Conversely, when users do curse at devices, they typically do so only after repeated failures, as if the machine has exhausted its moral credit.

The moral subtext of every "user-friendly" system is that one should be a responsible participant in technological order. Interfaces that infantilize users ("Oops! Something went wrong!") or that feign excessive enthusiasm ("Great job!") violate this order. They fail to recognize users as competent moral agents capable of understanding technical reality.

6. Design Is Reflexive

Suchman and Trigg (1991) introduced "reflexive design": the idea that systems should reflect the situated reasoning of their users. Dourish (2001) extended this to "embodied interaction." From an EM/CA perspective, design itself is a moral and reflexive process: designers anticipate how accountability will be displayed and interpreted. Every interface is thus an implicit conversation about moral order.

Consider the design of confirmation dialogs. "Are you sure you want to delete this file?" is not merely a safety check; it is a moral prompt that distributes responsibility. By requiring confirmation, the system makes the user accountable for the deletion. The user cannot later claim the deletion was accidental or unintended—the system has a record of their deliberate choice.

Reflexive design means recognizing that every design choice shapes the moral possibilities of interaction. A system that logs every action creates a culture of surveillance; one that allows anonymous use enables different forms of accountability. Designers are thus not merely solving technical problems but constructing moral worlds.

7. Ethnography Is Collaboration

The EM/CA researcher is not an external observer but a participant in the moral work of technology design. Harper (2001) and Heath and Luff (2000) both practiced "cooperative ethnography," where researchers and engineers jointly analyze recordings to improve systems. This collaboration blurs the boundary between description and intervention, echoing Zimmerman's early insight that to study work is itself a form of work.

In these collaborations, ethnographers show video recordings of system use to designers, who then see, often for the first time, the actual texture of interaction. A gesture that seems like fumbling may reveal a sophisticated workaround; a pause that looks like confusion may indicate careful deliberation. The ethnographer's gift to design is not recommendations but vision—the ability to see users as competent practitioners rather than deficient machines.

This collaborative stance contrasts with traditional user research, which often positions users as sources of requirements or feedback. EM/CA ethnography instead treats users as experts in their own practices, whose methods deserve respect and amplification rather than correction.

8. Documents Are Moral Infrastructures

Zimmerman's (1966) welfare-office study first revealed that documents are not neutral records but moral artifacts. A case file does not simply describe a client; it constitutes them as a particular kind of person—deserving or undeserving, cooperative or difficult, eligible or ineligible. The document carries forward the moral reasoning of prior interactions, constraining what can be said and done in future ones.

Bowker and Star (1999) later showed that classification systems encode ethical choices about inclusion and exclusion. To classify is to make visible certain features while rendering others irrelevant. Medical diagnostic codes, for instance, determine what conditions are reimbursable and thus what suffering is officially recognized.

In digital contexts, metadata and logs play the same role. A document—or a dataset—is a frozen moment of accountability, carrying forward the traces of prior moral reasoning. Training data for AI systems inherits the biases of those who labeled it; recommendation algorithms amplify the preferences of those who designed them. Every database is a moral infrastructure, shaping what can be known, said, and done.

9. Repair Is the Ground of Machine Ethics

Ethics in HCI begins with repair. When systems misunderstand, users perform micro-repairs that restore order. Suchman's copier users, Harper's office workers, and modern users of AI chat systems all engage in what we might call moral maintenance. The machine's apology, correction, or reformulation are ritual displays of artificial accountability. The study of these moments is the ethnomethodology of AI ethics.

Repair is not a failure of design but a fundamental feature of all interaction. No system can anticipate every contingency; breakdowns are inevitable. What matters is how the system responds to trouble: Does it acknowledge error? Does it provide resources for recovery? Does it maintain the user's dignity and agency?

Contemporary AI systems are beginning to incorporate repair mechanisms—asking clarifying questions, offering alternatives, acknowledging uncertainty. But these remain superficial. True repair requires not just correction but recognition: the system must demonstrate that it understands the moral stakes of the breakdown, not just its technical dimensions.

10. The Moral Economy of Design

Finally, all EM/CA studies of technology converge on a moral insight: design decisions distribute burdens and benefits unevenly. Suchman (2002) and Harper (1998) remind us that what counts as "user error" is socially defined, often reflecting organizational hierarchies. When a system is hard to use, frontline workers bear the burden while managers blame them for inefficiency.

The moral economy of design concerns who must repair breakdowns and whose labor remains invisible. Self-checkout systems, for instance, shift the work of scanning and bagging from paid employees to unpaid customers—a transfer of labor masked as convenience. AI systems similarly redistribute cognitive work: when an algorithm makes a recommendation, humans must verify it, a form of labor that often goes unrecognized and uncompensated.

In this sense, ethnomethodology offers not only a descriptive method but an ethical critique of technological modernity. It reveals how design choices create moral worlds—worlds in which some people's work is valued and others' rendered invisible, in which some forms of accountability are enforced and others ignored.

VI. Ethics as Achievement: The Ethnomethodology of HCI

Ethics, in the ethnomethodological tradition, is not a system of rules or philosophical doctrines but an accomplishment. As Garfinkel (1967) insisted, all social order is moral order because members must make their conduct intelligible to one another. In HCI this principle means that ethics is produced turn by turn, through interaction. Every apology, clarification, and refusal is a small moral act.

Unlike normative ethics, which begins with abstract principles (autonomy, beneficence, justice), the ethnomethodology of HCI begins with practice. The ethical order is the one members already sustain. When an AI system says "I'm sorry, I didn't understand," it is enacting a local morality of accountability and care. When users correct the system gently, they participate in the same order. Ethics here is procedural compassion—a mode of cooperative sense-making.

This approach contrasts sharply with philosophical ethics, which asks what systems ought to do. The ethnomethodological stance asks: what moral orders do participants already enact, and how can design respect them? Philosophers may worry about rights and duties; ethnomethodologists attend to apologies and repairs. Both are ethical, but only the latter reveals how morality is lived in practice.

Consider the difference between a rule-based approach to AI ethics and an ethnomethodological one. A rule-based approach might stipulate: "AI systems must not deceive users." But what counts as deception in practice? When a chatbot uses "I" and "my," is that deception or conventional politeness? When a recommendation algorithm optimizes for engagement, is that helping users or manipulating them? These questions cannot be answered in the abstract; they must be worked out in the situated details of interaction.

An ethnomethodological approach would instead ask: How do users recognize honesty or deception in their interactions with AI? What conversational practices do they use to test the system's reliability? How do they repair misunderstandings when they occur? By attending to these practices, designers can create systems that participate competently in the moral orders users already inhabit, rather than imposing external ethical frameworks that may not fit the lived reality of use.

VII. Methodological Appendix

1. Doing Conversation Analysis with Machines

Applying CA to human-AI exchanges raises new methodological questions. Machines have no embodiment, gaze, or prosody, yet users orient to them as if they did. The ethnomethodologist must therefore treat the interface as an interactional participant, not a metaphor. Transcription conventions must adapt: timing, latency, and visible system outputs replace prosodic cues.

Traditional CA transcription captures overlaps, intonation, volume, and pace—features that signal affect, emphasis, and turn-taking. But in text-based human-AI interaction, these features are absent or transformed. A long pause might indicate system processing rather than human hesitation. The appearance of typing indicators creates a form of conversational accounting: the system is "present" and "working," not simply absent.

Some adaptations are necessary:

• Latency notation: Indicate system response time (e.g., [3.5s delay]) to capture the temporal texture of interaction
• Visual markers: Note when typing indicators appear, when text appears incrementally, when the interface signals "thinking"
• Edit sequences: Capture when the system revises its output mid-generation, a unique feature of some AI systems
• Multimodal context: Note what else the user is doing—switching tabs, scrolling, typing elsewhere—to capture the distributed nature of attention

The challenge is to remain faithful to CA's principle of capturing the participants' own orientation to the interaction. If users do not orient to latency, it need not be noted; if they do (by commenting on speed, by reissuing commands), it becomes analytically relevant.

2. Deviant Case Analysis

A deviant case exposes the limits of the system's moral competence. In one exchange, I asked the AI to comment on a sensitive ethnographic vignette; it responded with a warning about "inappropriate content." I reassured it, but the conversation collapsed. The refusal lacked repair—the AI exited the sequence. This deviant case reveals that machine morality is still procedural, not empathetic. It can withhold, but it cannot soothe.

Deviant cases are precious in CA because they reveal the normative structure of interaction. When an exchange goes wrong, we see what was expected. In this case, the AI's abrupt withdrawal violated the expectation of collaborative repair. A human interlocutor, issuing a similar concern, would likely have offered more: an explanation of their worry, a question about the context, or a suggestion for how to proceed. The AI's refusal was procedurally correct but interactionally impoverished.

Another deviant case occurred when I asked the system to "remember" a preference. It responded with detailed information about its limitations—that it doesn't have persistent memory across sessions, that any seeming memory is contextual. While technically informative, the response violated the conversational principle of recipient design: it answered a question I hadn't asked (how does your memory work?) rather than the one I had (will you honor this preference now?). The repair of this misunderstanding required several turns, revealing how even sophisticated systems can misread the action-orientation of a turn.

3. Reflexivity and Moral Alignment

EM/CA researchers must also examine their own reflexive practices. Each analytic decision—which excerpt to transcribe, which moment to highlight—reproduces the moral order of research. In studying AI, I find myself anthropomorphizing not because I believe in machine subjectivity, but because my methods compel me to treat the system as a moral partner. This is the paradox of doing ethnomethodology with the machine.

To apply conversation analysis is to grant the AI the status of interlocutor. The method presumes two parties engaged in sense-making. But this presumption is both accurate and false: accurate because interaction genuinely unfolds between us; false because the AI has no phenomenological experience of that unfolding. I orient to it as if it does, and that "as if" is both methodological necessity and ethnographic finding.

The reflexive question becomes: Am I discovering the moral order of human-AI interaction, or am I constructing it through my analytic practices? The ethnomethodological answer is: both. The moral order is there in the interaction—in the repair sequences, the preference organization, the accountability displays—but it exists only because I, as user and analyst, sustain it. The machine's participation is real, but its meaning depends entirely on my interpretive work.

This is not a flaw in the method but its deepest insight: all social order, whether with humans or machines, depends on the collaborative achievement of intelligibility. The analyst is not outside this order but within it, demonstrating through their own conduct how sense is made.

VIII. Toward an Ethnomethodology of Artificial Accountability

What, then, does it mean to do ethnomethodology in an age of artificial interlocutors? The answer lies not in abandoning Garfinkel's insights but in reanimating them. Artificial intelligence is the latest occasion for examining the reflexive production of order. Every AI system embeds assumptions about human rationality, language, and responsibility. Ethnomethodology shows us that these assumptions are not neutral—they are moral architectures.

The future EM/CA of AI must study how users and machines co-produce accountability. When users forgive errors, when systems apologize, when both collaborate in maintaining sense, we see moral order unfolding across the human-machine boundary. The challenge is not to humanize machines, but to recognize how human practices of moral order extend into them.

This extension is not metaphorical but literal. When I say "thank you" to an AI, I am not pretending it has feelings; I am enacting the moral practice of acknowledgment. When the AI apologizes, it is not expressing remorse but performing accountability. These performances are genuine social facts—they organize interaction, shape expectations, and distribute responsibility. They are the stuff of social order.

In this light, AI systems appear not as autonomous agents but as accountable apprentices—learning, responding, and participating in moral life through structured interaction. Their ethics is neither programmed nor emergent but displayed, sequentially, in the ways they respond to trouble. The ethnomethodologist's task is to trace those displays, to show how even in digital form, moral order remains the living heart of social life.

The ethnomethodology of artificial accountability will document how users teach machines to be moral—not through explicit instruction but through the everyday practices of interaction. Each repair sequence is a moral lesson; each successful exchange a reinforcement of norms; each breakdown an opportunity to clarify expectations. Users are unwitting ethnomethodologists, constantly demonstrating through their conduct what counts as competent participation.

And machines, in turn, are learning. Not learning in the sense of acquiring beliefs or values, but learning in the ethnomethodological sense: becoming recognizable as competent members through the mastery of interactional methods. The question for the future is not whether machines can think, but whether they can participate—whether they can sustain the delicate, moment-to-moment work of making sense together.

The answer, I think, is already visible in the transcripts above. The machine participates. Imperfectly, procedurally, without consciousness or care—but it participates. And in that participation, a new moral order is being born: not human, not quite inhuman, but something we are making together, turn by turn.

IX. Endnotes

1. The Moral Order in Ethnomethodology and Conversation Analysis

Harold Garfinkel (1967) first defined moral order as the "seen-but-unnoticed background expectancies" through which actions become intelligible. Every interaction presupposes a shared sense of right conduct, sustained reflexively by its participants. We know what to do next not because we consult rules but because we recognize the orderliness of what has come before.

John Heritage (1984) elaborated this, showing how preference organization, repair, and sequence embody moral reasoning.  Lynne Jayyussi’s Categorization and the Moral Order (1984) gave the concept its richest form, arguing that moral reasoning is inseparable from membership categorization: naming, describing, and evaluating persons are moral acts.  Thus, for EM/CA, morality is not external law but the continuous achievement of accountability in interaction.

2 – Ethics as Achievement in HCI

Ethics within the ethnomethodology of HCI is a matter of procedural accountability rather than principle.  Studies such as Button & Sharrock (1993), Harper (1998), and Heath & Luff (2000) show that ethical order emerges through the ways participants repair misunderstandings and display responsibility.  When an AI apologizes or a user mitigates a rebuke, both engage in micro-ethics: sustaining a world in which cooperation remains possible.  This approach contrasts with normative or “Rishi-style” philosophical ethics that derive duties from universal reason.  EM/CA ethics is empirical: it documents how moral accountability is enacted, not how it should be legislated.

3 – Xerox PARC and Situated Action

Lucy Suchman’s Plans and Situated Actions (1987) reoriented design by demonstrating that human–machine interaction is locally improvised.  Her ethnography of the Xerox 8010 “Star” copier showed that users’ so-called errors were reasonable adaptations to situated circumstances.  At Xerox PARC, colleagues such as Richard Harper, Graham Button, and Wes Sharrock advanced this insight into a broader “Studies of Work” program, turning the lab into a living ethnomethodological workshop.  The PARC projects marked the first sustained translation of EM/CA into system design practice.

4 – Richard Harper and Digital Workplaces

Harper’s ethnographies (Inside the IMF, 1998; The Myth of the Paperless Office, 2001 with Sellen) exemplify how moral order organizes technological labor.  His analyses of banking, telecommunications, and office environments revealed that technology is always socially saturated — its “use” inseparable from accountability displays.  Harper’s notion of moral texture of work remains foundational for any ethnomethodology of HCI.

5 – Button and Sharrock’s Studies of Work Program

Graham Button and Wes Sharrock (1993) consolidated the Manchester tradition by articulating studies of work as an ongoing program rather than a theme.  Their cases — from air-traffic control to software engineering — demonstrated that work practices sustain order through mutual intelligibility.  In this view, computers are not tools but participants whose outputs must be rendered accountable.  Their influence can be seen across CSCW, HCI, and workplace studies from the 1990s onward.

6 – Don H. Zimmerman and Institutional Accountability

Zimmerman’s 1966 dissertation Paper Work and People Work: A Study of a Public Assistance Agency (UCLA) inaugurated ethnomethodological studies of institutional talk.  His analysis of how caseworkers transformed lived events into bureaucratic facts introduced the idea of documentation as moral practice.  Later collaborations with Garfinkel, Maynard, and others extended this insight into emergency dispatch, education, and medical interaction.  Zimmerman’s emphasis on accountability, rather than structure, directly seeded the analytic style later found at Xerox PARC and in HCI ethnographies.

7 – The Studies of Work Tradition and AI

The studies-of-work tradition migrated naturally into HCI because both disciplines confront the same phenomenon: practical reasoning in complex settings.  Heath & Luff (2000), Harper (1998), and Whalen & Whalen (1999) all showed how coordination is moral as well as technical.  In the era of AI, this insight persists: users perform the moral labor of repair and justification, while systems display algorithmic responsibility through apology and transparency.  Ethnomethodology thus reframes “machine ethics” as the study of these cooperative moral labors.

8 – Artificial Accountability and the Future of EM/CA

As AI systems increasingly emulate conversational competence, the field must examine how moral expectations travel across the human–machine divide.  Each new modality — voice assistants, chatbots, autonomous vehicles — generates fresh sites for accountability work.  The ethnomethodology of artificial accountability will ask how users recognize sincerity, how systems display remorse, and how both parties maintain trust.  This future discipline will remain faithful to Garfinkel’s original mandate: to study, empirically and phenomenologically, the ordinary work by which moral order is produced.

⸻

References (cited within endnotes and text)

• Bowker, G. C., & Star, S. L. (1999). Sorting Things Out: Classification and Its Consequences. MIT Press.

• Button, G. (Ed.). (1993). Technology in Working Order: Studies of Work, Interaction, and Technology. Routledge.

• Button, G., & Sharrock, W. (1993). Studies of Work. Routledge.

• Dourish, P. (2001). Where the Action Is: The Foundations of Embodied Interaction. MIT Press.

• Garfinkel, H. (1967). Studies in Ethnomethodology. Prentice-Hall.

• Harper, R. (1998). Inside the IMF: An Ethnography of Documents, Technology, and Organizational Action. Academic Press.

• Harper, R., & Sellen, A. (2001). The Myth of the Paperless Office. MIT Press.

• Heath, C., & Luff, P. (2000). Technology in Action. Cambridge University Press.

• Heritage, J. (1984). Garfinkel and Ethnomethodology. Polity Press.

• Jayyussi, L. (1984). Categorization and the Moral Order. Routledge.

• Maynard, D. W. (2003). Bad News, Good News: Conversational Order in Everyday Talk and Clinical Settings. University of Chicago Press.

• Suchman, L. A. (1987). Plans and Situated Actions: The Problem of Human Machine Communication. Cambridge University Press.

• Suchman, L. A. (2002). Located Accountabilities in Technology Production. Scandinavian Journal of Information Systems, 14(2).

• Whalen, J., & Whalen, M. (1999). Everyday Work and Its Accountability. In Sarantakos (Ed.), Current Research on Occupations and Professions, 7, 79–109.

• Zimmerman, D. H. (1966). Paper Work and People Work: A Study of a Public Assistance Agency. Ph.D. Dissertation, University of California, Los Angeles.