1 Introduction

Carlo Ginzburg’s reflections on distance emerge from a simple, but consequential, asymmetry: that historians, unlike anthropologists, cannot generate new evidence (Ginzburg, 2012). They must work with what has survived, however fragmentary, biased, or contaminated it may be. Whether dealing with inquisitorial archives, visual representations, or literary texts, Ginzburg insists that historical knowledge is always mediated through stand-ins — documents, images, traces — that both point toward and substitute for absent realities. Historians are bound to the traces left behind by others, often produced under conditions of coercion, inequality, or violence (Campt, 2017; Luft, 2020; Luft & Subotić, 2025; Manoff, 2004; Skarpelis, 2020). This constraint is not merely a practical, but also an epistemological one: at its best, it forces a particular relationship with evidence that is attentive to gaps, distortions, and asymmetries of power that become constitutive of historical knowledge itself (Fuentes, 2016; Sabbagh-Khoury, 2022; Smith, 1974). The difficulty of this endeavor drives much of Ginzburg’s project: how can we, as scholars, approach realities that we cannot directly access?

Ginzburg’s work on distance, perspective, and evidence — most fully developed in Wooden Eyes and Clues (Ginzburg, 1989; 2001) — offers a crucial but underused framework, particularly for thinking about how knowledge is produced when direct access to social reality is impossible. Although developed as a work of historical scholarship, his reflections are highly relevant for contemporary sociology and, in fact, ever more salient as the discipline increasingly relies on computational methods and explores generative artificial intelligence as a form of research methodology. If we read Ginzburg’s intervention as a theory of knowledge under conditions of unavoidable mediation, it can make explicit sociology’s longstanding but often implicit reliance on evidential modes of reasoning.

This perspective is particularly valuable for evaluating new forms of data and representation produced by generative AI, including but not limited to simulations and synthetic outputs through Large Language Models. While these technologies hold considerable promise for social scientific research, they also pose significant epistemological challenges. LLMs appear to constrain the range of possible meanings (Barrie & Celina, 2026; Hao et al., 2026), they exhibit “cultural tendencies” when prompted in different languages (Zhou & Zhang, 2024; Lu et al., 2025) and generate implausible anachronism (Underwood et al., 2025). Since the growing prevalence of deepfakes and other forms of synthetic media calls into question the evidential status of recordings, images, and textual data that have long functioned as trusted sources of testimony (Chesney & Citron, 2019; Rini, 2020), sociologists will have to justify knowledge claims when the link between evidence and the realities it purports to represent becomes uncertain.

Distance by itself shapes inference — little does it matter whether it takes on temporal, cultural, emotional, or epistemic form. In Wooden Eyes (Ginzburg, 2001), for example, Ginzburg insists that it is not merely a limitation imposed on the historian, but a constitutive condition of knowledge. Distance and Perspective: Two Metaphors, chapter 7, situates this type of perspectivism as a cognitive tool that allows scholars to hold together irreducible viewpoints without collapsing into relativism — partial viewpoints multiply angles and perspectives under conditions of constraint, rather than imposing one position, or contending that anything goes. Echoing feminist scholarship (Anderson, 1995; Haraway, 1991; Harding, 1987; Hill Collins, 1991), perspective is treated as neither an epistemic defect nor a purely subjective stance. Rather, it is a historically variable form of knowledge organization.

What matters, for Ginzburg, is not the (impossible) elimination of this mediation layer but a disciplined approach to its forms, affordances, and limits. Perhaps counterintuitively, given his thick, historical, and qualitative approach (which is excitingly on the cusp of a renaissance in sociological inquiry), Ginzburg’s micro-historical and critical engagement with written sources has latent methodological utility for computational work: it provides a conceptual grammar for “research at a distance” that is core to a critical epistemology in an age of artificial intelligence. Qualitative scholarship is essential for computational forms of inquiry. His concern, developed in parallel to debates in philosophy of science and historiography, thereby acquires renewed urgency as simulation, synthetic data, and AI agents increasingly function not merely as tools of analysis but as epistemic proxies for social reality itself. We can sum up this set of concerns pending more explicit theoretical articulation under the moniker of in silico perspectivism.

2 Disciplining the Analogy: Positive, Negative, and Neutral Transfers

Before I continue, I should clarify two things: first, what I mean by artificial intelligence; and second, what kind of analogy I presume potentially exists, as the step from inquisitorial archives to synthetic data and digital doubles is a steep one. To reiterate a commonplace from the Science and Technology Studies literature, there is no such thing as “AI”. As STS scholars Noortje Marres and Lucy Suchman have asserted, the concept of AI is strategically vague and neither a “self-evident”, nor an “autonomous technology” (Marres, Katzenbach, Munk, & Jobin, 2025; Suchman, 2023, p. 3). Here, I borrow Suchman’s (2023, p. 2) definition of AI as “a label for currently dominant computational techniques and technologies that extract statistical correlations (designated as patterns) from large datasets, based on the adjustment of relevant parameters according to either internally or externally generated feedback”. These techniques and technologies are fundamentally social and historically embedded, and within this large field of possibilities, I will focus on simulation and using AI agents as stand-ins, or proxies, for studying social action.

Second, the question of analogy and inference. Which inferential practices can travel from archival analysis to the analysis of synthetic data, and which ones are transformed or suspended in the process? I am in no way claiming equivalence or a neat transfer. To clarify the scope and limits of the analogy between Ginzburg’s archives and contemporary synthetic data, it is useful to draw on Mary Hesse’s (1966) classic account of analogical reasoning in science. In Models and Analogies in Science, Hesse (1966) distinguishes between the positive, negative, and neutral dimensions of an analogy. Positive analogies mark those properties that legitimately transfer from model to target; negative analogies identify known disanalogies; and neutral analogies designate properties whose status remains indeterminate and therefore analytically productive. The force of an analogy here is not exhausted through equivalence but in specifying which features are shared, which are excluded, and which remain open to inquiry.

Viewed through Hesse, my analogies are explicitly partial: their positive dimension lies in their epistemic stance. Both inquisitorial archives and synthetic data confront researchers with mediated access to social reality where direct observation is impossible. In both cases, knowledge must be inferred from stand-ins produced under asymmetric conditions of power — whether coercive interrogation or algorithmic optimization. We can now see how (and when) certain inferential sensibilities travel across domains. The first is interpretive caution, or a trained suspicion toward documents that appear too clean or consistent. Ricoeur (1969) referred to this as surface coherence that signals construction. The second is an attention to anomalies such as gaps, contradictions and surprises that are not noise to be omitted, but clues to be followed; one of the key insights of the indiciary paradigm. And third, a sensitivity to institutional context, that what counts as evidence is always shaped by institutional and organizational constraints (Smith, 1974).

At the same time, the negative dimension of the analogy — what does not transfer — remains relevant. Inquisitorial records, however distorted or ideologically saturated, remain indexically tethered to historically situated actors and events. Their contradictions and semantic slippages can be read as traces of competing cultural logics that “leak out” despite fairly structured and regimented overall regimes of control. Synthetic data, of course, lacks this form of clear indexicality. Generated through recursive patterning rather than lived interaction, it is less likely to preserve unintended traces of subjectivity or experience. As a result, anomalies in synthetic outputs cannot be interpreted as expressions given off in Goffman’s (1959) sense, that is, the kind of “non-verbal, presumably unintentional kind” of data in an information game that individuals find hard to manipulate and that is so crucial for others to observe to get a full and deeper handle on a specific person’s motivations, beliefs, or suppressed lifeworlds. They point instead to model limits, training distributions, or optimization failures. The analogy thus holds at the level of epistemic discipline, not at the level of referential grounding. Allow me to demonstrate with a case study designed to render the epistemic distinction operational and to reorient how we might think about stand-ins beyond their most obvious technical framings.

3 Analog Stand-Ins, Inference, and Nazi Racial Science

One of the most demanding implications of Carlo Ginzburg’s indiciary paradigm, detailed in Clues, Myths, and the Historical Method (1989), is the obligation to reconstruct perspectives that are not merely distant but morally repugnant. Take records of the inquisition that exclude the perspective of those interrogated and so are perspectivally narrow: They encode the inquisitor’s questions far more clearly than the defendant’s worldview (power indeed settles semantic disputes). Yet it is precisely where this perspectival control falters, at moments during which contradictions, hesitations, or semantic slippages appear, that alternative cultural realities “leak out”, if we want to deploy Goffmanian terminology (Goffman, 1959). Ginzburg tells us that historians can indeed derive value from such contaminated sources that were produced under extreme asymmetries of power because their very coercive conditions generate the kinds of tension through which evidence emerges. Saturated with power, the documents have the potential to reveal knowledge when read obliquely and dialogically.

This tension between identification and distance is central to Ginzburg’s indiciary paradigm, which has direct implications for sociology today. Emotional identification with historical actors, especially victims, can paradoxically limit interpretation by rendering certain analytical moves morally impermissible. Of course, Ginzburg’s refusal to collapse epistemic analysis into moral judgment does not imply ethical indifference. Instead, he generates conditions that allow us to reconstruct uncomfortable perspectives without having to endorse them. Take Wooden Eyes (Ginzburg, 2001), in which he argues in favor of defamiliarization and suspending taken-for-granted assumptions and works through the frictions produced by unsettling perspectives. His project is not one of rehabilitating such perspectives, but one of reconstruction.

The methodological commitment that emerges here approaches a particularly uncomfortable intensity when the archival records are of epistemic projects explicitly sustaining domination. One of the cases I am thinking through with Ginzburg, Goffman, and Peirce is that of the Nazi party expulsion trial of the racial psychologist Ludwig Ferdinand Clauss (Skarpelis, 2025). Clauss, while not a darling of the National Socialist regime, was nonetheless respected and a key player in Nazi racial science. And yet, he would find himself in the early 1940s subject to a Nazi party expulsion trial for having employed and collaborated with a Jewish woman.

At first glance, the Clauss trial appears to revolve around a narrow bureaucratic question: Whether the psychologist should be expelled from the Nazi party for having employed Margarete Landé, a woman classified as Jewish under the Nuremberg Race Laws, whose status neither he, nor she, had declared to the state as was newly required under the law. Superficially, the trial seems to center around the accuracy of ethnoracial classification and descent, and administrative compliance. Read obliquely, however, the files reveal something far more consequential: a rather sophisticated conflict over what constitutes legitimate knowledge, proper epistemology, and scientific authority within Nazi racial science itself. Reflexive racists!

Following Ginzburg, we are told not to take the trial’s categories at face value, but to attend to moments where controlled discourse falters and where contradictions or argumentative detours allow an alternative cultural logic to “leak out” of an otherwise monologic record. We are not quite dealing with the same power imbalances of the Inquisition, and still, it is precisely when Clauss departs from rote defense (“I did not know about her Jewishness”) to advance a substantive epistemological argument that the files become evidentially productive and, frankly, fascinating.

Clauss’s central claim is this: that Landé’s ostensible, merely partial Jewishness uniquely qualifies her to conduct research on “Semites”. Drawing on his phenomenological training, Clauss argues that Landé’s “inner twoness”, which is how he characterizes her concomitant inhabitation of both Jewish and non-Jewish worlds, affords her a form of insight unavailable to “Aryan” researchers. Therefore, rather than disqualifying her from scientific work, Landé’s positionality becomes indispensable to the production of “Aryan” racial knowledge, specifically Nordic Ideology. This argument places not only Clauss but also the racial state overall in a deeply paradoxical position: on the one hand, he panders to the Nazi regime’s essentializing obsession with racial difference; on the other, he implicitly undermines the ideological claim that Jews are epistemically inferior or incapable of scientific contribution. The trial thus becomes a conflict not “merely” about a single Jewish woman, potentially the love interest of a famous Nazi racial scientist, but over the epistemic foundations of Nazi racial science itself.

This is a paradigmatic instance of what Ginzburg describes as evidence emerging from contradiction. The trial records do more than simply reflect Nazi ideology, and Clauss’s defense puts a fundamental tension on the table: if Jewish scholars are essential to producing authoritative racial knowledge, then the very project of racial science depends on the capacities it officially denies. The archive of trial records preserves this contradiction, allowing the historian to reconstruct an epistemological debate that the regime itself would not openly acknowledge, and that postwar historians have, for good reasons, mostly steered clear from.

Analytical distance comes in handy here; the researcher needs to compartmentalize emotional identification with the victims of the Nazi regime to understand how complex knowledge systems operated from within. We have an emotional analogue to Ginzburg’s witchcraft trials: we may identify with the victims, but need to temporarily bracket this and take seriously those we despise in order to understand the epistemic machinery that made racial science plausible and operational, and include often paradoxically appearing internal contestations.

Having introduced a measure of discomfort, I wish to advance the argument by developing an association that you may have already considered and immediately dismissed: that Clauss’s notion of “inner twoness” bears an unmistakable structural resemblance to W.E.B. DuBois’s concept of double consciousness (Du Bois, 1903). Both describe a form of reflexive awareness produced by inhabiting multiple social worlds; both treat this divided standpoint as epistemically generative. The reason why the analogy is (hopefully!) unsettling to you is that you remember that DuBois mobilizes double consciousness toward emancipatory critique. Clauss, instead, instrumentalizes the idea in the service of hierarchical domination (with a tiny element of participation, in the form of a single woman who denounced her family and her Jewishness in the service of an intellectual project that wanted her dead). Comparing DuBois to Clauss does not assert equivalence, but it reveals how analogous mental constructs can be embedded in radically divergent moral and political frameworks.

Ginzburg offers a crucial clarification: analogies are heuristic devices that reveal structural similarities without erasing historical specificity or asserting moral equivalency. Reading Clauss alongside DuBois does not redeem racial science (which would lead to absurd claims like “See, the Nazis are using Black epistemology!”), but it clarifies how claims about perspective, positionality, and knowledge can be severed from emancipatory ends and redeployed as tactics of oppression. We are made to confront the possibility that domination may occasionally rely on sophisticated epistemic arguments.

How could we possibly make it back to AI at this point? Let me attempt the following: long before simulation and synthetic stand-ins, Clauss treated perspective not as a claim to represent otherwise under-represented voices or awarding Jews any form of positive recognition, but as an epistemic resource detachable from the real-life woman who embodied it. What mattered was not Landé’s agency, but her positional usefulness. Her value consists in the dual status of her in- and exclusion; she succeeds as an early “stand-in” only because her perspective was deemed inaccessible to the dominant group. This was because the regime presumed that she was essentially different in a crucial way, and that this difference provided otherwise inaccessible insights.

While the trial does not entirely prefigure some of the current controversies around identity of AI agents and synthetic subjects (who is authorized to represent whom, under what conditions, and for what epistemic ends?), there is some continuity in that the unique perspectives an “inner twoness” affords can be wielded for projects of domination. Today as in the past, perspective gets detached from subjecthood, is extracted without reciprocity or the chance for resistance and reversal, and all this epistemic extraction takes place under conditions of duress and domination. Epistemic stand-ins operate within historically situated contexts of power and violence.¹ The trial thus functions as a historical instance of the same structural move: the severing of perspective from subjecthood while retaining its epistemic value. By reconstructing odious perspectives, we can explain how domination sustains itself through (occasional attempts at) epistemic sophistication, rather than remaining stuck in blatant irrationality.

In insisting on the reconstruction of such unpleasant perspectives, the case of Clauss’s trial underscores that the indiciary paradigm, despite its focus on an oftentimes playful detective approach and the element of surprise, demands that sociologists work through archives that encode domination, contradiction, and violence, without mistaking analytic reconstruction for endorsement. Understanding how truth claims are made, contested, and stabilized requires taking seriously, if only temporarily, the cognitive worlds that produced them. What remains open to debate in this beginning analogy between 20th-century archives and generative AI is to ask how we can recognize moments in which such stand-ins become epistemically productive. Ginzburg, after all, trains us to watch out for the surprise, to study moments of breakdown, during which regimes of representation lose some control over what they seek to formalize. What moments of surprise exist in computational systems that eliminate noise?

4 From Archives to Algorithms: New Conditions of Distance

Ginzburg’s work remains salient when sociological research is conducted at a technological distance. Large language models, simulations, and synthetic datasets introduce a new type of mediated distance between researcher and object. Unlike traditional archives, these systems do not merely preserve (the biased, power-laden, deeply socially and culturally imprinted) traces of past social worlds; they offer the promise of actively generating plausible stand-ins for them. In classical sociological approaches, stand-ins for reality remain grounded through a representational logic: a document, image, or datum is assumed to refer back to an object in the world, however imperfectly. Even when this relationship is contested, as in the case of problematic sources, at least the bias emerges straight from the source and remains indexically linked to lived experience. Generative artificial intelligence’s use of LLMs as proxies for social action in turn unsettles this assumption, and we are left with a nonreferential approach to meaning that comes with its own challenges, like hallucinations and other factually or logically wrong statements (Alvarado, 2024). While not new, and irreducible to the problem of truth and deception, synthetic data challenges us in ways that are not fully captured by thinking about perspective and distance. The question for sociology is thus not whether such data are biased — they invariably are — but what kind of inferences they allow (Larson, 2021).

Contemporary anxieties about “counterfeit people”, deepfakes, or synthetic personhood (Chesney & Citron, 2019; Dennett, 2023; Hacking, 2006) become more than simple problems of deception, especially when companies try to redress historical biases. Public controveries abound around simulated humans, especially where this generation of humans racializes in publicly awkward ways: recall Google Gemini’s “Black Nazi” gaffe (Titcomb, 2024), one of many instances that led to a July 2025 executive order issued by the White House titled “Preventing Woke AI in The Federal Government” and characteristically written in all caps, which mandated that henceforth the only acceptable large language models had to adhere to principles of “truth-seeking” and “ideological neutrality” (The President of the United States, 2025). Formulating the issue as one of “woke AI” frames the problem as one of ideological distortion imposed on an otherwise neutral technology that is naturally oriented toward accurate representations of the world and whose outputs only become distorted by human interventions in the form of guardrails, fine-tuning, or diversity-oriented adjustments. Bias becomes an external imposition, ideology a removable layer, and truth a stable baseline that can be restored through technical correction.

From a Science and Technology Studies as well as from a Ginzburgian perspective, this framing is epistemologically naïve. Sociologically, it also naturalizes a particular distribution of epistemic power. By portraying engineers, ethicists, or corporations as moral entrepreneurs who supposedly “meddle” with otherwise neutral outputs, the political critique obscures actual infrastructures of knowledge production — like training data, model architecture, and institutional priorities — that fundamentally shape what AI systems can generate. Sure, there are moral entrepreneurs who actively define the social problem of AI, as Mira Vale has shown (Vale, 2024), but these are not the imaginary arbiters envisioned in White House rhetoric, who are supposed to be interfering with an otherwise objective and universal “truth” of large language model outputs. Appeals to “unbiased AI” thus echo earlier and historically recurrent efforts to cleanse knowledge of rhetoric, interest, or standpoint. There is no untainted evidential baseline. The task of sociology, then, is not to purify simulation of bias, but to situate it within the broader ecology of knowledge production.

This type of public controversy, often instigated by actors invested in the overall field, may disguise as much as it reveals (Marres et al., 2025). Let us therefore move towards an example that Marres et al. (2025) would call a “type 2” kind of controversy, in which the controversy articulates the situation. For example, let us look at cases of AI-generated representations of historical figures and marginalized subjects, in which attempts to redress historical wrongs have led to “woke-washing”. In short, cases that “explicate and demonstrate an underlying — diffuse, more general — political situation” (Marres et al., 2025, p. 5). Critics have pointed out that projects such as Khan Academy’s “Harriet Tubman” chatbot, which invites students to interview a simulated version of a nineteenth-century abolitionist, modernize Tubman’s language and sensibility, smoothing over historical difference in the name of accessibility and thereby reproducing a form of epistemic violence by generating “surface-level impersonations” (Wallace & Peeler, 2024). Similar concerns animated public backlash against Meta’s short-lived “sassy black queer” AI assistant, widely denounced as an instance of digital blackface (Attiah, 2025).

Both public controversies act as a diagnostic in that they allow us to articulate politically salient situations. In both cases, offense did not stem from the artificiality of the output per se — few would object to the general principles of chatbots adopting invented personas — but from the way simulation echoes historically familiar patterns of appropriation, caricature, and power asymmetries. Synthetic data in these instances does not merely fail to represent historical reality; it risks reenacting the very hierarchies that shaped the archives whose gaps it purports to fill. What is objectionable is not the voices’ artificiality, but that their mode of production alters historically grown power relations to suit contemporary goals, all while presenting itself as an innocent mode of access or recovery. Simulation in these instances does not stand in for missing archives, it overwrites them. The problem is not artificiality but monologic simulation, where users are served outputs that leave no space for leakage or slippage and in which power has settled meaning in advance.

5 Inference Under Conditions of Epistemic Severance

All the while, the social sciences are cautiously excited about the potential of large language models and other generative systems as tools for “reconstructing” lost voices or simulating social interaction where empirical access is limited because of cost, ethics, or feasibility. The idea here is to have LLMs “serve as a (potentially informative) proxy for human behavior” (Van Loon, 2025). As stand-ins for social actors, AIs turn into entities that can be queried, experimented upon, and observed. This perspectival shift from models as tools to models as agents recasts simulation as a mode of inquiry. in that, they become epistemically expansive. Of course, we have a “fundamental paradox” here when we try and use LLMs to simulate subjects (Kozlowski & Evans, 2025), because the very thing that simulation is good at — being a predictive tool to model behavior that we have no ground truth for — fails the very moment we try and use a system that only works when it can rely on available ground truth. It thereby introduces a form of distance that differs qualitatively from the archival distance Ginzburg theorized. Whereas trial transcripts may be distorted by coercion and ideology, they remain indexically linked to historical actors. Synthetic data, on the other hand, preserve statistical regularities without maintaining representational correspondence.

In machine vision research, for example, artificially generated images have been used to improve recognition rates for underrepresented populations by augmenting biased training sets (Buolamwini & Gebru, 2018; Denton, Hanna, Amironesei, Smart, & Nicole, 2021; Scheuerman, Hanna, & Denton, 2021; Skarpelis, 2024; Wang et al., 2022; Zhao, Wang, & Russakovsky, 2021). The success of the model is measured not by representational fidelity but by downstream performance: do the images improve facial recognition rates for marginalized groups? These systems’ instrumental logic, of course, sits uneasily with sociological concerns about meaning, history, and power. A dataset may be “good” in the sense of optimizing a classifier while remaining epistemically opaque with regard to the social processes it encodes. A “good” image or dataset is no longer one that faithfully represents reality, but one that performs well within a specified task environment; a truly operational image (Paglen, 2014).

Synthetic data is not new, nor is it inherently deceptive. Statistical practice has long involved working with data that is partially constructed. We impute, reweight, and model counterfactuals. We can do so because these operations carry explicit assumptions about the relationship between the stand-in and the underlying social process. What distinguishes contemporary generative systems is that this relationship is no longer specified or testable; the synthetic simply replaces the real without an inferential bridge. The indiciary paradigm thus is faced with a new configuration in which clues no longer point backward to an event or actor, but sideways to an optimization function. How evidence is grounded is fundamentally altered: the validity of its grounding now comes from relational performance. In that sense, the work that LLMs are made to do in the simulation of human subjects might be best understood as a fairly superficial form of creating composites or ideal types — in short, an abstracted kind of fiction.

We can define fiction here as “arrangements of facts and inventions that we approach without the expectation that they accurately reflect all aspects of the knowledge or the target systems they invoke” (Hanlon, 2024). Consider Weber’s ideal types as an example, the explicitly heuristic devices that exaggerate certain features that, through this idealized abstraction, allow us to benchmark behaviors. Goddard (1973, p. 15) helpfully explains that we end up with a useful construct of a thing in the world “by stressing its typicality, that is, by stressing its distinctive elements from the point of view we are interested in the thing.” Applying well-defined concepts through abstraction (and some fiction) exposes clearly “fragments of reality” (ibidem). Fictions in and of themselves are not contrary to social scientific inquiry, but they have to be modeled as such and not used as substitutes for direct observation. Ethnographers concerned about the safety and anonymization of their sources have reflected on this in great detail. Arjomand’s (2022, p. 437) discussion of composites as a tool in ethnography, for example, helps us understand the promises and limits of such fictionalizing moves “without sacrificing rigorous attention to empirical data.” Among the main epistemological gains of Arjomand’s postpositivist analytical sociological approach are that we gain theoretical breadth as we sacrifice a small amount of “empirical exactitude” that we can and ought to be internally accountable for.

Of course, this accountability is currently missing from large language models, and the situation becomes even more complex when authorship itself is destabilized. Ginzburg’s dialogic model of reading presumes an encounter between historically situated subjects, even when power relations are unequal. With AI-generated texts, that presumption no longer holds. As recent literary theory has argued, readers can no longer maintain the standard expectation that an unknown text is human-generated (Bajohr, 2024; Henrickson, 2024). This shift forces a reconsideration of what dialogue means under conditions where there is no authorial intent, and whether it is at all possible.

Returning to epistemic severance, we might ask: if synthetic data has largely dissolved the semiotic link between datum and world, what kind of inference remains possible? As Rafael Alvarado put it, large language models are not trustworthy because “they do not meet the basic epistemic criterion of justified true belief” (Alvarado, 2024). But LLMs are untrustworthy in ways fundamentally distinct from, say, inquisitorial records. If classical archival inference combines abduction and contextual corroboration, in silico inference may be said to consist in abduction sans indexical grounding. The answer to what remains of inference might lie in a reconfiguration rather than an abandonment of evidential reasoning: abductive inference in the Peircean sense, rather than a reconstructive mode of inference. What is inferred is no longer past social worlds based on ground truths, but plausibility structures and optimization regimes. The indiciary paradigm thus persists, but as a way of diagnosing systems of knowledge production, of how regimes of plausibility, normativity, and power are encoded in computational systems, rather than reconstructing absent voices.

6 Surprise in the Machine

In connecting Ginzburg’s indiciary paradigm to in silico research and the creation of generative AI stand-ins for humans, I suggest that his approach offers a potential critical epistemology of distance that we might call in silico perspectivism. The archive and synthetic output are not epistemically equivalent. What travels are interpretive caution, attention to asymmetries of power, and the productive value of surprise. What does not travel is the indexical grounding that underwrites them in the archives. The instance of the racial psychologist Clauss is a paradigmatic hinge case, in that it reveals how epistemic sophistication can sustain (and disguise!) domination, and how perspective can be instrumentalized and severed from subjecthood. This helps clarify the stakes of contemporary AI debates by showing that epistemic extraction without reciprocity is not new, even if its technological form may be. And while synthetic outputs do not “leak” in the way that we would like them to, moments of surprise point sideways and allow for a media infrastructural analysis that can point us towards design choices and moral assumptions built into the system that we can then seek to understand better through other modes of inquiry. Distance is a condition of knowledge, not its obstacle.

Compared to critical AI scholarship, sociology’s purpose is not to debias AI, but to suggest different modes of methodological reflexivity when considering using AI agents as proxies, for example. We can explicitly model synthetic data as useful fiction and heuristic, or even as composite, in ways that invite collaboration between historical sociology, science and technology studies, and computational social science. What are the consequences of such methodological reflexivity, especially under conditions where the social realities we study are increasingly mediated by generative systems? If this new layer of mediation is indeed there, we need to move beyond studies of bias, truth, and deception, but also move beyond getting stuck at the diagnosis of epistemic severance. Ginzburg taught us that mediation through biased traces can be made methodologically productive: our task is to make visible the new regimes of knowledge.