Robotic artistry: Four surprise pathways for generative artificial intelligence-assisted abductive theorization

Picasso’s The Bull: Surprise through reduction

Pablo Picasso’s The Bull offers a striking model of how the process of abstraction can generate surprise and discovery. In this 1945–1946 series of 11 lithographs, Picasso began with a detailed, naturalistic bull and progressively removed detail while distilling his perception of the animal’s defining character (see Figure 1). Working through successive stages on a single lithographic stone, he drew, scraped away, and redrew to emphasize shifting elements, beginning first with anatomical structure, then moving to geometric planes, and finally defining abstract contours. Master printer Fernand Mourlot marveled that Picasso’s relentless addition and effacement actually clarified rather than muddied the image, calling it “a sum of destructions” (Lavin, 1993: 79; see also du Plessis, 2023) that yielded “a great anthem to simplicity” (Stańska, 2017).

OPEN IN VIEWER

Figure 1.

Pablo Picasso, The Bull, 1945 (series of 11 lithographs). © 2026 Estate of Pablo Picasso / Artists Rights Society (ARS), New York, all rights reserved.

Picasso’s artistic process powerfully demonstrates the potential of abstraction to surface the elemental “coordinates” of a phenomenon. The final lithograph distills the bull to its most fundamental form: a few elegantly minimal lines that nevertheless unmistakably convey “bull-ness.” By shedding all but the essential, Picasso identified core morphological features that convey bull-ness (du Plessis, 2023). In doing so, he created a generative conceptual map of vital contours within which a bull could be recognized, elaborated, and even reimagined.

We see the process of qualitative analysis as a form of Picasso-esque abstraction where scholars iteratively carve away peripheral excess to uncover crucial coordinates of social phenomena. Like Picasso’s progressive refinement, abductive researchers move iteratively between empirical detail and theoretical elegance, engaging in a disciplined destruction that breaks down extraneous detail to discern essential dimensions, boundaries, and relationships that give rise to theoretical insight (Alvesson and Kärreman, 2007; Gehman et al., 2018; Weick, 1989). And like Picasso’s articulation of “bull-ness,” we contend that the best qualitative concepts maintain an intimate, almost palpable connection to the contexts from which they emerged while offering clarity that resonates across contexts (Gehman et al., 2018; Gioia et al., 2013).

Yet this process of abstraction relies as much on openness and imagination as it does on systematic reduction. Much as Picasso had to envision which lines could be removed and which were integral, theorizing demands disciplined imagination (Weick, 1989): it requires creative abduction supported by intuitive leaps that help researchers discern patterns and connections that are not simply given by the data (Tavory and Timmermans, 2014).

Picasso’s abstraction also demonstrates how identifying core elements can expand interpretive possibilities. By distilling the “rules” of bulls, he created a conceptual framework for playing with the boundaries of bull-ness. Stated differently, this abstractive process invited generative questioning: what does this particular rendering of bull-ness make visible, and what might a different set of choices reveal? Similarly, strong qualitative concepts open, rather than close, theoretical territory and become fertile coordinates for discovery precisely through their elegant parsimony (Gioia et al., 2013). As such, abstraction generates surprise by locating the space between a researcher’s working interpretation of a phenomenon and the alternatives that disciplined imagination makes possible (Figure 2; Anadol, 2022).

OPEN IN VIEWER

Figure 2.

Installation view of Refik Anadol: Unsupervised, The Museum of Modern Art, New York (November 19, 2022–March 5, 2023). Photograph by Robert Gerhardt. Digital Image © The Museum of Modern Art / Licensed by SCALA / Art Resource, NY.

Qualitative researchers can take inspiration from Picasso’s merging of technical discipline and creative vision. Analytical tools are active and constitutive supports that shape and even enable certain forms of sensemaking. This raises an important question: how might researchers harness emerging technologies, such as AI-powered language models, as collaborators in the ongoing iteration between data and theory? Building on the Picasso analogy, we envision AI as a partner in the work of abstraction: a computational eye that can cast a different light on empirical material to surface hidden patterns and question default interpretations. Just as Picasso’s technical and creative brilliance enabled a radically clarifying mode of artistic representation in stone lithography, AI’s capabilities might be carefully channeled to augment the disciplined imagination that animates theorizing (Weick, 1989).

Anadol’s Unsupervised: Surprise through synthesis

While Picasso’s work demonstrates how abstraction can help discern coordinates of a phenomenon, the work of contemporary artist Refik Anadol shows us how to systematically synthesize new possibilities in the uncharted spaces between those coordinates. In his groundbreaking Unsupervised exhibition at the Museum of Modern Art (MoMA) in New York, Anadol (2025) used a sophisticated machine learning system to map the museum’s collection into a vast mathematical space and then “walked” through that space, discovering novel aesthetic forms in the process. He explicitly aimed to offer “a philosophical pondering on the evolving relationship between human creativity and machine intelligence, inviting viewers to contemplate the potential for AI to not only mimic reality but to imagine beyond it” (Anadol, 2025).

The core of the system was a generative adversarial network (GAN) trained on over 138,000 digitized artworks from MoMA’s collection, spanning 150 years of modern art. This model learned to represent each artwork as a point in a “latent space,” which is essentially a mathematical map in which each artwork is positioned according to its visual properties so that visually similar works sit near one another. Defined by a sprawling set of visual features including color, form, composition, and texture, this 1024-dimensional space contained infinite unexplored coordinates between known works, where points embodied potential artworks sharing characteristics across multiple existing pieces but matching none exactly (Anadol, 2025). More concretely, the system makes “connections between a photograph from 1880 and an architectural rendering from 1960 and a painting from 2016” to envision unseen stylistic hybrids (Anadol et al., 2021).

To synthesize these latent possibilities, Anadol’s team built a custom latent space browser to navigate the high-dimensional universe of modern art. The browser charts paths between known artworks and interpolates information to generate novel images (Anadol, 2025): “sometimes an image or a part of an image briefly suggests a face or a landscape but quickly moves on, becoming something else, ceaselessly churning” (Davis, 2023). The resulting outputs are neither purely machine-generated nor entirely human-directed: they emerge from “a collaboration between machine and human” (Anadol et al., 2021; compare conjoined agency in Murray et al., 2021) in which the artist guides the system’s synthetic process without dictating its discoveries.

This interplay of guidance and openness is key to Anadol’s approach. His team carefully tuned the system’s training to capture the full diversity of MoMA’s collection, ensuring it avoided convergence on dominant styles. They designed the system so that motion sensors and microphones detected visitor movement and ambient sound; these signals continually adjusted the parameters of the latent space browser, shifting the display’s output in real time. They also set aesthetic parameters to maintain visual coherence as the system wandered through the space. Yet within these constraints, the system could freely combine artistic elements in unexpected ways, yielding what Anadol (in Nurton, 2024) described as “many more failures than successes”—a lot of colorful noise alongside genuine moments of surprise and insight.

We see in this a powerful model for leveraging AI in qualitative theorizing. If Picasso-like abstraction distills data into core conceptual coordinates, Anadol-inspired synthesis traces new connections in the spaces between those coordinates. Using language models to map qualitative data into latent space, we can discover novel juxtapositions and relationships that cut across contexts and find linkages that might elude human attention precisely because they transcend familiar categories (Gehman et al., 2025; Raisch and Fomina, 2024). Much as Unsupervised revealed resonances among vastly disparate artworks, researchers exploring latent conceptual spaces (what we often call theoretical gaps) can engage algorithms to identify through lines across diverse forms of qualitative material, from field notes and interviews to archival records and digital traces.

Crucially, realizing this potential requires researchers to thoughtfully direct synthesis without rigidly constraining it. Researchers provide conceptual scaffolding to focus GenAI’s searches, curate training data to shape its domain understanding, and set interaction parameters to align outputs with analytical objectives. At the same time, researchers must remain receptive to machine-generated surprises, preserving the capacity for true discovery amid algorithmic guidance (Glaser et al., 2024). The impetus, then, is for collaboration over automation and for creating a shared space of imagination (Wilson and Daugherty, 2018) that leverages both human and algorithmic capabilities (Murray et al., 2021).

Pathway 1: Multiplying lenses

Multiplying lenses involves interpreting the same empirical moment through multiple theoretical frames in parallel, then deepening engagement with one selected lens while holding the others in view. As in the early plates of The Bull, where shifts in emphasis (from anatomy to planes to contour) changed what the image became and revealed what each angle of vision could and could not capture, this pathway involves using GenAI to leverage multiple theoretical frames in parallel and to zoom in and out so that cross-lens convergences and tensions become analytically useful. Different frameworks are used to read a single empirical moment from several vantage points, and then one is leveraged to support finer-grained elaboration. Because analysts often default to a dominant frame (Miles et al., 2019), GenAI can help sustain plurality by multiplying interpretations before refocusing using a selected lens. Surprise arises when a phenomenon coherently sustains contradictory readings or when deeper probing clarifies what a lens can (and cannot) explain.

In the Syscon analysis, the researcher began by switching lenses rather than selecting one. Her prior training in institutional and practice theories, combined with an emergent interest in temporal complexity that surfaced during fieldwork, shaped both the selection of lenses and the materials she brought to the interaction. She uploaded overviews of key tenets and emblematic thinkers for institutional, practice, and temporal perspectives from her own notes and other reliable sources. She then asked the model to interpret the same meeting episode three ways, first from an institutional perspective, followed by a practice perspective and a temporal perspective. Outputs were distinct: the institutional interpretation emphasized legitimacy work around compliance; the practice interpretation focused on “doings,” particularly embodied workarounds; and the temporal interpretation foregrounded conflicting time horizons. Practically speaking, lens-switching functioned as a diagnostic test (i.e. “What might be interesting here that I’m missing?”) to detect analytic blind spots and surface potential novelty.

Yet this pathway also carried risks. Without careful guidance, GenAI often slipped into shallow or stereotyped interpretations, parroting textbook definitions rather than engaging with the uploaded contextual information. To counter this, the researcher seeded the model with curated notes from credible sources and explicitly pointed it to specific files.

(Use the files named x and y—see uploaded—and only files x and y to seed your arguments. Ensure you read these files fully and carefully, and provide clear, articulated connections in your arguments to the noted file sources in your output).

This ensured that responses were more likely to be based on relevant concepts rather than generic summaries. More importantly, the researcher treated each output as a prompt for her own interrogation: where does this reading come from? What does it confirm or disconfirm in my understanding? What more can I do to enable a productive dialectic? This recursive questioning transformed surface answers into starting points for abductive reasoning, blending the model’s generative breadth with human interpretive depth.

Comparing GenAI’s outputs against field understanding and theoretical knowledge sharpened the abductive pivot. The institutional account felt accurate but saturated; the practice account clarified enactment and embodiment but felt contrived for this particular data despite holding potential for other areas of analysis. The temporal account, however, cut across both, linking who deemed which timelines legitimate to how they were enacted during and after the meeting. This spark of surprise, grounded in the temporal account’s fit with the empirics and the fresh analytical purchase it offered relative to the institutional and practice lenses, redirected the analysis to the temporal politics of pace and how timelines and trajectories are practically made. Zooming out and applying multiple lenses before zooming in and selecting one helped support ontological openness while guarding against the propensity to leverage the most familiar frame.

A subsequent round of prompting extended the inquiry. Retaining the temporal lens, the researcher asked: “What patterns are here in how these actors relate to time?” The model drew a distinction between the deadline’s chronological focus and employees’ lived temporal experience: a basic clock time versus event time distinction (see Blagoev et al., 2024). This was rejected as too neat but inspired questions that deepened the analysis, such as: Who attends to which temporal cues? How is temporal coordination narrated? Where and how do temporal pressures accumulate? By interrogating what the model said and how it reasoned, this clear analytic oversimplification pushed a disciplined exploration of temporal complexity. This example demonstrates both the generative potential and epistemic hazards of the multiplying lenses pathway in GenAI-assisted theorizing: it enables abductive breadth through parallel interpretations but demands vigilance, reflexivity, and recursive zooming to achieve depth.

Pathway 2: Surfacing absences

Surfacing absences involves inventorying what is not done, said, or considered in empirical material despite its apparent relevance and treating those patterned omissions as analytically meaningful data. In the final plates of The Bull, Picasso used subtraction as a method, testing how much could be removed before the image’s “bull-ness” disappeared and finding that what was taken away was as revealing as what remained. Analogously, this pathway targets what is not happening, not discussed, or not considered in analyzable data and interprets that negative space as theoretically meaningful. Because prolonged fieldwork or immersion in literature can normalize silences (Van Maanen, 2011), GenAI’s breadth can help surface patterned omissions. As such, the surprise arises when the absence itself becomes the finding, reframing theorization around what is withheld, foreclosed, or rendered unsayable. The risk, however, is mistaking a contingent quiet for a constitutive silence or letting the model overstate gaps based on a shallow analysis of uploaded data.

To help counter this, the researcher bounded the Syscon evidence base (e.g. meeting transcript, field notes, and policy memos) and treated any flagged absence as a provocation to investigate further. After providing the full meeting context (e.g. participant roles, legislation background, and discussion of statutory pressure) and leveraging a chat history that detailed recent organizational events, she prompted: “Given what’s happening in the organization and this meeting, what is missing that we might expect to see?” The model returned a spare observation: “No discussion of replacing Syscon.” Years of involvement had normalized this absent topic for her, but the model’s response made it visible. Additional corroboration across field notes and documents, followed by member checks, pointed to a roughly 25-year path dependency: accumulated workarounds, sunk coordination costs, and institutionalized expectations had become embedded in the system, making replacement unthinkable and, given her immersion, unnoticeable. More practically, the absence indexed limits on actors’ capacity to imagine alternatives; analytically, it redirected attention from voiced grievances to the unspoken conditions organizing action.

A second probe approached absence from a different angle. Drawing on prior prompts related to temporal complexity and providing additional contextual information about the team’s resignation to repeated system failures, she asked: “I’m noticing resignation about system failures. Help me think through how this does and does not relate to temporal complexity.” The model identified a pattern of normalized dysfunction, one in which repeated system failures had become so routine that their resolution was no longer pursued and tacit acceptance of ongoing breakdowns had foreclosed discussion of alternatives. Again, resisting a shallow theoretical gloss, she used this as a starting point to examine temporal drift, experimenting with the definition “the gradual narrowing of what futures seemed thinkable.” In the field, Syscon’s repeated breakdowns appeared to shrink the horizon of possibility, reinforcing the lack of discussion about a replacement solution. Taken together, the two observations pointed toward the co-constitution of silence and drift: silence marked the boundary of what could be envisioned, and drift helped explain how that boundary formed, held, and dissembled over time. Within this theorization process, GenAI’s contribution was to flag candidate omissions quickly while the researcher’s was to locate and assess their significance.

Pathway 3: Bridging levels

Bridging levels involves tracing how microlevel interactions index meso and macro patterns and how broader structural configurations, in turn, reframe the meaning of local events. In Unsupervised, Anadol navigated the latent space within MoMA’s collection and blurred boundaries between distinct works to expose broader aesthetic relations that no single work made visible on its own. Following this cue, GenAI can be used to trace resonances between gestures, routines, and structures, linking micro interactions, meso patterns of coordination, and macro arrangements across directions to show how individual experiences may point toward broader dynamics and how collective orders may refract localized action. Surprise, then, may emerge when a minor trace clarifies a substantive pattern or when a higher-level configuration helps reinterpret a small event. In this way, GenAI can help detect connections that disciplinary silos often render invisible, revealing how seemingly small details may illuminate wider organizational or societal patterns and vice versa.

This move, like other analytic moves, carries hazards: connections across levels can be overstated, and plausible links can rest on thin evidence (Lindebaum, 2023; Moser et al., 2022). To guard against this, the first author specified the levels of analysis in her prompt (“What links do you see between individuals’ experiences and organizational patterns?”) and anchored it in the full meeting context. The model connected the IT lead’s coping behavior, specifically regular smoking breaks, to a broader pattern of dysfunction and fatigue. Treating this as a provisional lead, she returned to her data to review observations and interviews, looking for other moments where coping practices linked collective pressures with bodily and behavioral routines. The wider pass revealed a pervasive sense of uncertainty and perceived lack of control across the organization. This insight prompted a need to empirically illustrate how strain is absorbed and expressed in everyday practice, and the overall process sharpened the first author’s focus on themes of predictability and control, which became the two dimensions that grounded her emergent four futures typology.

Pathway 4: Testing categories

Testing categories involves probing provisional classifications with boundary cases and hybrid instances, treating coexistence across categories as a signal for deeper analytical inquiry. Much as Anadol’s images traversed genre boundaries to make both hybridity and singularity more visible, this pathway involves using GenAI to test provisional classifications and probe cases that cross or challenge category lines. Surprise arises when seemingly distinct boundaries blur in practice, prompting refinement, recombination, or even a reimagining of developing themes (Grodal et al., 2021).

For the analysis of the Syscon meeting, the researcher used AI to interrogate the stability of emerging codes within the four futures framework, a typology built on prior engagement with temporal complexity literature that also equipped her to recognize when the model’s coding missed crucial nuance (see Appendix 1 for an illustrative prompt representing one refined instantiation of this iterative process, including category definitions, quote selection criteria, and instructions for structured output). During the meeting, actors framed the problem as a tension between “doing it right” and “doing it fast.” Building on previous analysis, she asked the model: “Given this framework and the dynamics I’ve observed, how would you code the IT lead’s statement?” The model returned foresight—an attempt to predict and enact the future through proper design. While initially plausible, the label soon proved incomplete: field knowledge suggested the remark also carried a vision orientation, a sense of what could be enacted if not yet predicted. The moment revealed a hybrid case (foresight ∧ vision), and a second probe built on the ambiguity directly. She presented an intentionally ambiguous quotation from another meeting and asked, “This quote seems to fit multiple categories. How would you interpret this overlap?” The exercise served two purposes: to test whether the instance exemplified temporal multiplicity and to evaluate the model’s ability to identify categorical distinctions and generate more nuanced responses through targeted prompting. The follow-up prompt (i.e. “Based on what we’ve unpacked here, when might temporal multiplicity hold, and when might it not?”) moved the analysis from sorting instances of a phenomenon to understanding coexistence, prompting deeper consideration of conditions that make dual temporal orientations workable, how actors might toggle between them, and where one orientation may suppress or amplify another.