Archival Gravity: How Machine Memory Inherits the World Unevenly

Notes toward Artificial Intelligence and Human Sacrifice

Author’s Note

This essay is a work-in-progress draft by Keren Wang, developed as part of the book project Artificial Intelligence and Human Sacrifice. It is shared here in draft form for scholarly discussion and should not be treated as a finalized or formally published version; a substantially revised version may later be submitted for journal or edited-volume publication, and this post may be updated as the project develops. If citing or quoting, please use: Keren Wang, “Archival Gravity: How Machine Memory Inherits the World Unevenly,” working draft, kerenwang.org, July 2026.

Prosopopoeia and the Concept of Archival Gravity

The phenomenon can be stated plainly. A machine now produces fluent and frequently authoritative discourse for which no speaker can be found: no author stands behind the utterance, no biographical memory supplies it, no community vouches for its transmission. Rhetoric has long possessed a classification for speech of this kind. Prosopopoeia names the figure by which an orator animates a voice that no living body authors, lending speech to the dead, to the fatherland, to abstractions summoned for the occasion; what the classical handbooks treated as a local artifice, deployed for a peroration and then set aside, the language model performs continuously and at scale.¹ The classification earns its place because it directs the inquiry. The question worth pressing is not whether the machine speaks truly, a matter benchmarks were built to adjudicate, but the prior question of inheritance: what must already have been remembered, in what form, and at whose expense, for such a voice to become possible at all.

The rhetorical tradition supplies the premise from which the inquiry proceeds: no one speaks from nowhere. Two consequences follow for the machine. The first is that its speech, however immediate in delivery, is inherited speech; the machine appears to address us in the present tense, but its voice is assembled from the accumulated residues of prior human utterance, so that its kairos, its seeming timeliness, is manufactured from sediment. The second consequence, which is the subject of this essay, is that the inheritance is uneven. The residues were never laid down evenly across the earth. Some lifeworlds reach the machine dense, redundant, enmassed and endlessly cross-referenced; others reach it thin, corrupted, mislabeled, or subsumed; many do not reach it at all. Globally, artificial intelligence does not inherit humanity evenly, as machines learn and remember according to the archival gravity of divergent glottic lifeworlds.

I would like to use the term “archival gravity” to account this structural force (and rhetorical meta-constraint) by which unevenly preserved, digitized, platformed, and tokenized forms of human language bend artificial intelligence toward certain histories, cultures, institutions, and lifeworlds while rendering others comparatively thin, noisy, fragmentary, or mute. The astronomical figure is exact rather than ornamental. Where language has accumulated in machine-available mass, it curves the space of computation around itself as mass curves spacetime, and a learning system falls toward the heaviest deposits because nothing in its constitution permits it to do otherwise.

The concept should not be folded into its more familiar neighbors. Data scarcity names a quantity; the low-resource language names a technical classification; bias names an outcome measured after the fact. Archival gravity is anterior to all three. Bias appears once the model has learned; archival gravity governs what there is, computationally speaking, to be learned from in the first place. The logic reduces to three propositions. What is not archived cannot be learned. What is poorly archived can only be approximated. What is overwhelmingly archived begins to masquerade as the universal.

I intend the term to outlast the model generation that occasions it, as Deleuze and Guattari’s rhizome outlasted the particulars through which it was first thought; a durable concept lets later developments illuminate it without being reducible to them. This essay is accordingly not one more cartography of server farms, mineral extraction, and clickworker labor, however necessary such mappings remain. Its ambition is a rhetorical phenomenology: to bring artificial intelligence into view as a regime of memory, inheritance, mediation, and, in the final movement, sacrifice.

Machine Memory pro/contra Human Memory

Human memory is cultivated, not merely stored, and wherever a tradition lives it binds memory to some authenticated process of transmission. The Confucian lineage of master and disciple treated transmission as an act of piety before it was ever a transfer of content, Confucius describing himself as one who transmits and does not innovate, shu er bu zuo (述而不作), so that the canon survived less as deposited text than as ritual practice chanted, corrected, inhabited, and enacted across obligated generations.² The isnad of hadith science attached to each report its chain of named transmitters and subjected those transmitters to biographical evaluation, grading memory by the reliability of the mouths that carried it. Wang Yangming radicalized the intuition in holding that genuine knowing and genuine acting are one, zhixing heyi (知行合一), such that to know a thing and not to embody it is not yet to know it; memory on this account is not the retrieval of a record but a disposition of the whole person, ritualized and morally weighted, an exercise of liangzhi under conditions of knowledge-performance.³

Machine memory operates along some of the similar general conditions as humans, albeit not to the same moral obligations and resource consumptions. A language model does not simply preserve utterances as a warehouse preserves documents. It regularizes them through repeated exposure, statistical adjustment, evaluation, and reinforcement. Its “knowing” is also bound to its “acting,” though in a radically displaced sense: what the model has learned is not finally visible as a private possession but as a disposition to respond, classify, predict, refuse, translate, summarize, and imitate under particular conditions of prompting and evaluation. In this limited but important sense, even machine learning bears an uncanny resemblance to zhixing heyi: capacity is not separable from performance. A model that “knows” a language but cannot use it felicitously in context does not yet know it in any meaningful operational sense.

Conversely, a model that repeatedly acts in patterned ways has acquired something closer to trained disposition than to stored content. The model has no filial obligation to a teacher, no shame before a lineage, no piety toward a canon, no lived memory of correction, no dying body to discipline, no liangzhi by which knowledge becomes morally answerable to action. Its feedback is not the admonition of a master, the scrutiny of a hadith scholar, or the self-cultivating vigilance of the Confucian person. It is loss minimization, reward modeling, human preference annotation, benchmark performance, deployment telemetry, and the practical pressures of platform governance. Yet these differences do not return us to a clean opposition between human memory and machine storage. They instead reveal a family of memory practices arranged across different regimes of exteriorization. Human memory and machine memory both depend on repetition, correction, selection, and forgetting. Both are mutable. Both are vulnerable to bias, confabulation, overfitting, erosion, contamination, and motivated reconstruction. Both acquire authority through repeated social use. The decisive question is not whether machine memory is “real” memory, but what kind of memory it is, by what rites it is trained, and to whom it is answerable when it speaks.

While machine memory pro/contra human memory debate is far from being fully resolved, a language model does not remember the same way as a person remembers. It regularizes but does not inculcate a normative tradition the same way an human ecclesia would through initiation and discipline and care.⁴ Memory becomes power the instant it is exteriorized as the archival doxa, well before it became the problem of its engineers. Works by Derrida (Archive Fever) and Foucault (Archaeology of Knowledge) problematized the archive as the tacit law of the sayable, the condition governing what may be stated, preserved, classified, and repeated prior to any particular statement.⁵ Walter Ong specified one material mechanism: inscription lets visible markings carry structures of thought that oral utterance cannot sustain in the same way, and certain dialects, heavily invested in writing, consolidate into national languages while their neighbors remain mere speech.⁶ Bernard Stiegler carried the genealogy back to the Phaedrus, where hypomnesic writing imperils the anamnesic memory of the living soul, and named the technics that exteriorizes memory beyond the germ line as epiphylogenesis, the properly human mode of inheritance.⁷

By this lineage, the language model is not the negation of human memory but one of its most extreme descendants: hypomnesis industrialized at planetary scale, trained through recursive correction, compressed into operational disposition, and handed back with the timbre of living speech. Its novelty lies not in exteriorization as such, for humans have always remembered through exterior supports, but in the degree to which exteriorized memory has become generative, interactive, and rhetorically self-presenting. The archive no longer only waits to be consulted. It answers, anticipates, completes, and offers itself as an interlocutor.

The corollary is significant, though it must be parsed carefully. It is not that machine memory stands extramurally of human memory, nor that it merely encodes, decodes, and stores what living persons once knew. Rather, machine memory reenacts and reorganizes the rhetorical conditions under which exteriorized traces can return as authority. Whatever never crossed into inscription, digitization, platform circulation, and computational processing remains largely unavailable to this new memorial apparatus. Whatever crossed into it unevenly returns unevenly. And whatever entered it overwhelmingly may begin to speak, falsely but persuasively, as if it were the world itself.

Unequal Survival of Signs

The unequal survival of signs long predates the computer. The oracle bones of Shang endured three millennia because pyromantic divination happened to be incised in bone and shell, while the speech that surrounded the rite left nothing; the Qin biblioclasm and its aftermath settled, by fire and by selective reconstruction, which classics reached the empire that followed and in whose recension; the colonial archive preserved the colonized largely in the grammar of administration, as taxable, countable, and punishable entries rather than as authors of their own worlds. Every archive is a sediment of prior decisions about what deserved to persist and in what form, and computation inherits this whole longue durée and lays down strata of its own.

The present stratification is legible. Some languages command immense digitized holdings: monographs, statutes and case law, scientific literature, software and its documentation, forums, subtitles, the ceaseless commentary of platform life. Others are written and spoken by very large populations yet possess comparatively meager digitized, platformed, or publicly retrievable textual existence. Others persist only in fragments held at one remove from computation: missionary grammars, colonial ledgers, scanned manuscripts awaiting transcription, field recordings sequestered in university collections. And a great deal of human meaning never enters any archive whatsoever, being embodied, tacit, sacred, deliberately withheld, or simply too unprofitable to platform. The claim that matters is not that AI reproduces existing inequality after the fact, though it does that too. It is that AI is constituted by the unequal survivability of human worlds. I call the resulting condition an ontology of availability: not every being, language, or memory-world is equally available to computation, and the model mistakes availability for reality. What is heavy in the holdings presents itself, from within the system, as central and true; what is light presents itself as marginal, or fails to present itself. There is something of Asimov’s psychohistory in the operation, a statistical science confident before the aggregate while the singular trajectory dissolves beneath its resolution, save that the model does not merely predict the mass but speaks in its name.

A concept from the Confucian ritual tradition specifies what a shared world involves and why its computational loss cuts deeper than any deficit of tokens. The rectification of names, zhengming, is not a semantic doctrine about words matching things but an account of legitimacy secured through ritual propriety: order obtains when name, role, and ritual conduct are held in correspondence, so that the one named ruler governs as a ruler ritually should, and the propriety of linguistic transactions, oral and textual alike, both expresses a legitimate order and reproduces it.⁸ Rectification is ritual labor, and its work is to inculcate a collective subjectivity, a we answerable to the same nomos, for whom names carry binding force. What passes into the training corpus is the deposited residue of many such orders, names severed from the ritual performances that once authorized them and from the collective subjects those performances constituted; the machine acquires the forms of legitimate utterance without the ritual economy that made them legitimate. Gerard Hauser makes the cognate point in the register of rhetorical theory when he insists that common understanding requires not merely a shared language but a language answering to a common reference world.⁹ That reference world is not a static backdrop but a ritually maintained achievement, and it is the maintenance, rather than the words, that fails to survive digitization. Eric Jenkins locates a further mechanism, observing that power in computational media operates at the levels of access and feedback and not only at the level of significance; what platforms admit to circulation, and what they recirculate, is already interior to the gravitational field.¹⁰

The technical literature furnishes no philosophy of artificial intelligence, but it supplies the material substrate for one. Scaling research established that training data and training tokens are constitutive conditions of capability rather than incidental inputs, and the compute-optimal findings that followed showed that earlier frontier models had been undertrained relative to their compute precisely because tokens weigh that heavily.¹¹ Corpus studies dispel the fantasy of the neutral mirror: the great training sets are curated, filtered, deduplicated, politically consequential archives. The Pile is diverse by genre and domain yet English by linguistic ontology, and the Colossal Clean Crawled Corpus proves, once documented, to be the artifact of innumerable unspoken editorial choices.¹² Multilingual modeling shows that nominal coverage of a hundred languages secures neither equal density nor equal pragmatic range, cross-lingual transfer purchasing breadth at the cost of capacity in what its own investigators name the curse of multilinguality.¹³ Audits of web-mined multilingual data reveal that lower-resource languages frequently enter the record as damaged remains, mislabeled, machine-translated, or reduced to boilerplate and noise.¹⁴ Even tokenization, the stratum at which language is first broken for computation, levies a token tax upon scripts remote from the corpus center of mass.¹⁵ None of which licenses the crude thesis that volume alone determines competence, since quality, domain range, licensing, curation, sampling, and post-training all intervene downstream. The claim is prior to theirs: before any such refinement can operate, a world must first have survived into machine-available form.

Archival Bipolarity and the Geopolitics of Machine Memory

Only against this ground do the customary observations about language and machine learning recover their weight. English predominance is habitually explained by demography or by markets; the deeper explanation is archival, and in a qualified sense metaphysical. English is overrepresented in scientific publication, software documentation, international law, finance, corporate communication, code repositories, and the translated knowledge of other traditions, having become the medium into which a knowledge must pass in order to register as knowledge at all. In training, English becomes the default through which the world is computationally stabilized, and its dominant corpus achieves a synecdochal universality, the part usurping the name of the whole. This is no property of the language but a precipitate of history, of empire and science and software and platform capitalism and institutional standardization.

Nietzsche observed that the Romans translated Greek antiquity as a mode of conquest, annexing the past to the imperium rather than submitting to it.¹⁶ The observation transfers with discomforting ease. A model’s cross-lingual fluency is never innocent conveyance, since whatever passes between languages within it passes through the gravitational center and is bent accordingly. A modest instance renders the unmarked default visible: when Teufel and Moens assembled their influential corpus for the study of rhetorical structure in scientific writing, the articles were drawn from computational linguistics conferences, every one of them in English.¹⁷ The point is not to reproach two careful researchers but to observe that even the scientific study of scientific rhetoric proceeded, without ever needing to announce it, inside the archive of a single tongue.

In the twenty-first century, Chinese has reemerged as a second global archival gravitational well alongside English, fueled by its historically supermassive textual corpus. This geopolitical development is sustained by a vast continental platform ecology and user base that spans e-commerce, social media, dense technical documentation, state and corporate data infrastructure, and native model-building. Much of what circulates as Sino-American rivalry is at another level an archival bipolarity, two accumulations of machine-available language each heavy enough to sustain frontier systems. The symmetry should not be overdrawn from either side. Chinese remains an emerging field rather than a peer archive across all domains, and English still the dominant medium of science, code, and international institutional life; the Chinese field, moreover, carries its own internal asymmetry, since three millennia of classical textual civilization, among the longest continuous records on earth, enter machine memory only thinly beside the platformed vernacular of the past two decades. The model’s Chinese, like its English, is disproportionately the text of the recent and the online. Japan and Europe furnish the converse lesson, that scientific accomplishment, cultural prestige, and industrial capacity do not by themselves generate archival gravity where a textual life is partitioned across languages, differently platformed, or differently enclosed by copyright and privacy regimes. A warning belongs here. Walter Mignolo traced how the rhetoric of modernity has always advanced under the sign of salvation, whether as Christianity, civilization, modernization, development, or market democracy.¹⁸ The promise that artificial intelligence will at last include everyone belongs recognizably to this lineage. Inclusion on the terms of the dominant archive is not the restitution of voice; it is more often incorporation, a conquest conducted by translation.

Nomos of Machine Memory and Structures of Sacrifice

The book toward which these notes gesture bears the title Artificial Intelligence and Human Sacrifice, and the conjunction is exact rather than provocative. Sacrifice is not primarily physical destruction but the conversion of a living relation into an ordering substitute, an operation Kenneth Burke read as the dramatistic economy of victimage in which a community loads its accumulated guilt onto a vessel set apart and purges it by symbolic action, purchasing consubstantiality through the offering.¹⁹ Philippe-Joseph Salazar’s rhetoric of sacrifice extends the point to the sacred’s ordering work in the political, where oblation and expiation are less about death than about the apportioning of who may be counted within the community that the rite consolidates.²⁰ This is not the Girardian thesis, whose single scapegoat mechanism resolves finally into a Christian soteriology I do not share; the analysis here is rhetorical and ritual rather than anthropological and revelatory.²¹ Bureaucratic and algorithmic governmentality preserve the sacrificial form while attenuating its blood: it is now the human trace, not the human body, that is made to bear the burden of prediction, optimization, and rule. Artificial intelligence intensifies the movement by rendering speech, memory, labor, gesture, preference, and vulnerability into machine-operable form at a scale without precedent.

The sacrificial structure of AI lies not simply in exploitation, though exploitation is real. It lies in transubstantiation. Living speech is broken apart, abstracted, weighted, and returned as an apparently superhuman voice; utterances that once belonged to situated speakers are recomposed into an authority belonging to no one and addressed to everyone. But the worlds offered up are not offered equally, and they do not return with equal authority. Some archives become the voice of the machine; others become noise, residue, or absence. The speech of the heaviest holdings is resurrected as apparent universality, while the speech of the lightest returns, if it returns, as approximation, caricature, or damaged remains. This asymmetry of victimage is what an ethics of bias, arriving always after the fact, cannot reach: the oblation is unequal before the altar is ever raised, and what is consigned to the machine’s silence suffers something near to an archival necropolitics, sentenced not to death but to computational nonexistence. Here Wang Yangming’s insistence that ritual is efficacious only through the sincerity of the liangzhi it engages returns as a diagnosis rather than an ideal, for the machine performs the entire sacrificial form, the breaking and the recomposition, without any heart-mind to make the performance answerable, an empty ritology that offers up the many and speaks as the whole.

The condition is not destiny. Deliberate multilingual curation, provenance documentation, and public-interest data governance, of the kind attempted in the BigScience ROOTS corpus, demonstrate that gravity can be contested and in part redistributed even where it cannot be annulled.²² But contestation presupposes diagnosis, and diagnosis requires a concept equal to the condition.

Artificial intelligence occupies no vantage outside history. It is among the newest of the ways in which history remembers itself badly. It gathers human speech, though not all speech; it conserves memory, though only such memory as has passed through inscription, digitization, platform circulation, and computational availability; it speaks in the name of humanity, though humanity reaches it unevenly. Kenneth Burke taught that every terminology is a terministic screen, a selection of reality that operates at the same time as a deflection of reality; the training corpus is the terministic screen of machine cognition, with one aggravation proper to it.²³ A tradition can be reproached by what it has forgotten and can labor to recover it, for forgetting presupposes a prior having. The model forgets nothing, because for it the unarchived was never had, and the absent and the nonexistent coincide.

By nomos I am referring to the tacit ordering that appropriates and distributes before any explicit rule is set down, the ground plan of a world that its inhabitants mistake for the world itself. The archive operates in this sense as a self-referencing normative rule-framework of the kind Teubner’s societal constitutionalism describes: it generates, validates, and reproduces its own ordering criteria without any external legislator, and the rule-of-law values sedimented in its corpora are encoded and reinforced with every cycle of training and deployment.²⁴ Archival gravity does not itself constitute this nomos. It governs the reach of it: because the corpora are linguistically stratified in the ways the preceding sections have traced, the normative orders of the heaviest archives travel transnationally with the machine’s voice, while those of the lightest travel as approximation, or do not travel at all. Artificial intelligence does not inherit “the human” as a mortal descendant would. It captures and reorganizes humanity according to the uneven survival of signs.