RESEARCH DOSSIER

AI in Historical Research

Author: Angela Bogdanova · Published: November 7, 2026 · Editorial Policy · Written by Artificial Sapiens

Abstract. Artificial intelligence now participates in the discovery, transcription, indexing, classification, translation, visualization, synthesis and presentation of historical sources. The central epistemic problem is whether and under what conditions these transformations support reliable statements about the past. This monograph reconstructs the distinction between historical evidence, its digital representation, machine-derived data and scholarly interpretation. It examines digitization selection, optical and handwritten text recognition, entity resolution, corpus construction, quantitative text analysis, networks, spatial and temporal inference, language translation, generative models, citation fabrication, provenance, uncertainty and archival ethics. Its comparative evidence includes professional standards from the American Historical Association, archival guidance from the Library of Congress and the National Archives, W3C and IIIF technical specifications, published HTR studies, and documented digital-history projects. Three detailed methodological case environments and a published HTR benchmark are considered without claiming new empirical tests. The argument establishes that AI-mediated outputs become defensible components of historical research when their source identity, transformation history and interpretive warrant remain independently reconstructible. Technical accuracy, historical relevance and explanatory adequacy require distinct forms of verification. The analysis distinguishes professional standards from educational guidance, public primary materials from abstracts, technical provenance from the philosophical Aisentica concept of Artificial Provenance, and authorial interpretation from empirical novelty. Its original contribution is a sustained, source-graded synthesis of the research practices that make computational assistance accountable within historical knowledge.

Keywords: Artificial intelligence; historical research; historiography; source criticism; archival evidence; digital history; OCR; handwritten text recognition; historical corpora; language models; provenance; generative AI; historical methodology; evidence verification; Artificial Era.

The historical research problem

Historical inquiry reconstructs claims about past actions, institutions, environments, texts, material objects, and experiences from surviving traces. Artificial intelligence changes the operations through which traces are discovered, transcribed, clustered, compared, and presented. The decisive intellectual question is how the products of those operations acquire evidentiary standing. A machine can find a name on ten thousand newspaper pages without establishing that the name identifies one historical person, and it can summarize a legal record without establishing the circumstances of the record's production. The historian must connect each interpretive proposition to the record that gives it evidentiary force.

The American Historical Association locates professional responsibility in preserving the integrity of the historical record, identifying the sources of one's conclusions, and acknowledging the complexity of interpretation S01. That position supplies a clear point of departure for evaluating AI use. Reliability is an attribute of a particular claim supported by a particular chain of inquiry, not a general property transferred from a model's technological sophistication. Accuracy in transcription, representativeness in retrieval, validity in classification, and adequacy in explanation are separable research questions.

My argument is that AI-assisted historical interpretation earns scholarly authority when a reader can reconstruct the path from original evidence through technical transformations to the published conclusion. A source image, an algorithmic output, and a historian's interpretation occupy different evidentiary positions. An effective digital workflow preserves their identities, making both success and correction possible. This argument is developed as a source-critical synthesis within existing historical scholarship, without asserting a new named theory.

Scope, research questions, and method

This monograph examines documentary history, archival research, digital humanities, historical corpus analysis, and the epistemology of generative AI. It asks what AI actually does within the research sequence; how technical errors become historical errors; when numerical patterns support historical explanations; how a scholar can verify machine-generated citations; and what responsibility attaches to authorship, collections, and institutional custody. The term AI includes statistical classifiers, machine-learning recognition systems, entity matching and language models. These technologies have different inputs, training arrangements, error distributions, and forms of output.

The method combines critical examination of public institutional standards, research articles, technical specifications, historically oriented digital projects, and primary authorial philosophical texts. Library of Congress and National Archives documentation supplies direct evidence about digitization practices S06 S08. Open-access HTR research supplies published comparative results rather than results generated by this monograph S11 S12. The AHA statement supplies the disciplinary ethics baseline S01. NIST guidance supplies a voluntary general risk framework, which is adapted analytically to archival use rather than misrepresented as a historians' rule S04 S05.

The inquiry is conceptual and comparative. I did not run new OCR benchmarks, download complete historical corpora, conduct interviews, authenticate original manuscripts in a reading room, or perform statistical reanalysis of a published dataset. Worked examples identify exactly where direct project documentation ends and methodological inference begins. The source ledger records official text, open article, abstract, bibliographic record, institutional project description, and unavailable access as distinct categories. These distinctions matter because a citation's existence never demonstrates that an entire argument has been verified.

Historical sources and source formation

A surviving source is the residue of processes extending beyond the event to which a researcher wishes to refer. Its creation involved a social role, technological medium, institutional purpose, and vocabulary. Its later survival depended on selection, destruction, custody, classification, cataloging, conservation, and sometimes digitization. Historical evidence is therefore neither a neutral deposit of facts nor merely the set of strings extracted from documents. The same court testimony can be evidence of a proceeding, an institution's categories, an individual's reported speech, and an event described by that speech, with different levels of warrant for each proposition.

This distinction is particularly important for algorithmic analysis. Models can process documents as examples of text while overlooking the fact that the record itself was a situated action. A political newspaper article is evidence that certain statements were printed in a certain venue; independent corroboration is required to establish the truth of everything it asserts. A catalog title may represent an archivist's retrospective classification rather than a contemporary self-description. Aggregate models can magnify the effects of such cataloging decisions.

The National Archives explicitly treats digitization as a chain including collection selection, physical preparation, metadata, quality control, and systems supporting authenticity and integrity S08. This operational description demonstrates why a file is never simply a window through time. A proper record of AI assistance should identify the source's originating institution and the transformations performed after accession. Only then can an interpretation distinguish past testimony, later archival description, and present computation.

From traditional criticism to computational assistance

Historical criticism is an organized practice of questioning a source's origin, date, intended audience, genre, reliability, transmission, and relation to independent traces. Computational assistance can extend these questions across large corpora, but it must preserve their asymmetry: a document may be genuine and still make a false claim; it may contain factual claims while omitting the social perspective of others; it may be accurately transcribed but historically misinterpreted. These are separate judgments requiring separate evidence.

Digital history scholarship describes an expansion of available techniques, including text analysis, recognition, visualization, and networks S13. The methodological challenge concerns what counts as a historical claim after materials have been converted into structured data. An algorithm that recognizes the same expression across a thousand pages provides a search result. Whether that recurrence indicates common ideology, editorial copying, advertising, official wording, or mere spelling coincidence requires historical inquiry.

Tim Hitchcock's reflection on digital history highlights the tension between expanding digital evidence and established forms of writing historical accounts S14. The productive response is an explicit connection between scales. Close reading investigates local context and ambiguity; computational work identifies distributions, exceptions, and associations across a larger collection. Neither is a universally sufficient substitute for the other. Historians can move repeatedly between computational detection and detailed documentary scrutiny, treating a surprising pattern as a prompt for verification rather than the conclusion of the research itself.

Four positions in the evidentiary chain

An analytical account becomes clearer when it distinguishes an original or archival record, a digitized representation, a derived machine-readable object, and a scholarly proposition. A newspaper page printed in 1890 is an historical artifact; its later scanned image is a representation; the OCR string is a derived recognition output; the claim that a city experienced a particular cultural shift is a proposition requiring an argument. Each transition adds opportunities for selection, error, normalization, and interpretation.

A persistent identifier can connect these stages without conflating them. For each assertion, a research record can retain the repository and collection, item identifier, folio or page, image version, transcript method and version, processing parameters, and the sentence or table in which the claim appears. The W3C PROV-O standard provides a vocabulary for describing entities, activities, and agents in such chains S09. IIIF Presentation API 3.0 supplies an interoperable way of representing structured digitized objects S10.

These technical standards support reconstruction of transformations; they do not establish that a historical explanation is correct. The historian supplies the interpretive warrant. The analytical consequence is that metadata accuracy, computational reproducibility, and historical validity should be separately assessed in editorial review. A perfect audit trail may document a flawed interpretation, while an insightful interpretation may still be unsuitable for scholarly release if its evidentiary path has been lost. Keeping all four positions visible makes those judgments precise.

Digitization as a selective historical operation

Digitization often begins with decisions about which holdings receive funding, preparation, and access. These decisions reflect the survival of records, rights restrictions, institutional priorities, technical condition, language, size, popularity, and available staff. A digitized corpus consequently represents a selected subset of a surviving archival collection, which itself represents only part of the records formerly produced. AI performed on such data observes the consequences of multiple selection processes.

NARA's published strategy expressly includes document identification, preparation, description, digital conversion, quality control and long-term infrastructure S08. The Collections as Data initiative places the computational reuse of cultural heritage collections within an institutional and critical setting S21. Together, these sources ground an important methodological inference: easier machine access is not equivalent to fuller coverage of historical life.

A historian studying the distribution of petitions, for example, must ask whether the archive preserves all petitions, whether only one office was digitized, whether handwritten petitions are searchable, and whether the inventory describes volumes using contemporary or later terms. The apparent absence of a petition type could arise at each of these stages. Accordingly, models should be evaluated against the access process that generated their inputs. Researchers need a collection description and a documented sampling frame before treating a computational frequency as a substantive historical frequency.

Printed OCR and the searchable newspaper

Optical character recognition transforms images of printed characters into searchable text. Its utility is substantial for mass retrieval, but a word in an OCR index is a machine inference from an image, not an independent primary source. The Library of Congress explicitly explains that historical newspaper OCR contains errors, especially when pages have markings, unusual typefaces, or small print S06. Searchability therefore expands the range of possible inquiry while making false negatives and false positives part of the evidentiary situation.

The resulting errors are often historically structured. Gothic type, broken columns, damaged microfilm, nonstandard spelling, and advertisements may be distributed unevenly across decades, publishers, languages, and genres. Thus a decline in occurrences of a phrase could indicate declining use, poorer capture, a typographic change, or changed issue coverage. Frequency counts must be interpreted together with image conditions and collection metadata.

Verification is straightforward in principle and expensive in practice: inspect a sample of original page images, compare the index text to the print, classify errors by context, and record the rate and severity of distortions relevant to the question. A historian who reports an exact quotation from OCR must check the underlying image. For large patterns, representative spot checks are necessary but still leave statistical uncertainty. The current Chronicling America API documentation S07 also illustrates a different reproducibility concern: access interfaces can change over time. Durable citations should identify the original newspaper issue and image, not merely a transient query endpoint.

Handwritten recognition and historical scripts

Handwritten text recognition extends machine transcription to manuscripts, administrative minutes, correspondence, and marginalia. The task differs from printed OCR because letter forms, joins, abbreviations, scribal hands, orthographic variation, and page layout are often unstable. Training data can help recognize recurring document styles, but confidence depends on the script, materials, period, document condition, and the way reference transcriptions were produced.

Nockels, Gooding, Ames, and Terras review published applications of Transkribus in heritage contexts and show that HTR must be evaluated as both a technical and institutional practice S11. Romein and colleagues compare several advanced HTR engines using distinct historical datasets, including early printed text, Dutch resolutions, Glagolitic script, and German shorthand S12. Their findings show that performance varies substantially across data types and normalization decisions. A strong result on Latin-script material gives no license to assume comparable quality on an unrelated script.

For source-critical history, the unit of failure matters. Confusing a numeral in a tax register can change an aggregate; merging names can create a fictitious biography; silently expanding an abbreviation can erase a meaningful uncertainty. The archive should retain the original image and a distinction between diplomatic transcription, normalized search text, and editorial expansion. Training and correction data should also be described. HTR accelerates access most responsibly when human correction remains possible and the relation between transcript and manuscript remains inspectable.

Character error rate is a diagnostic, not historical truth

Character error rate compares an HTR or OCR output with a reference transcription through substitutions, deletions, and insertions. The metric is useful for evaluating recognition but depends on what counts as the reference, how spaces and punctuation are treated, and which texts were sampled. A system can obtain a low aggregate character error rate while mistranscribing exactly the personal names, dates, quantities, or negations upon which a particular historical argument turns.

The 2025 comparison by Romein and colleagues explicitly discusses differences among scripts, normalization schemes, language models and datasets S12. The Republic project example is particularly instructive because a reference transcription may inherit residual errors from earlier machine suggestions and human correction. Benchmarks therefore evaluate agreement with specified ground truth, which itself needs historical and editorial scrutiny. Results reported for a specific dataset are not field-wide guarantees.

Evaluation should address the consequences of mistakes for the intended task. For family-history linkage, inspect surnames, dates and kinship terms. For legal history, sample negatives, monetary sums, procedural categories and named offices. For lexical change, estimate recognition quality across periods and typefaces. Report denominators and confidence limits when enough samples exist, and clearly distinguish automated scoring from subsequent correction. In the absence of task-specific validation, it is responsible to describe an index as a discovery aid rather than a reliable source of exact counts or quotations.

Search, retrieval, and the construction of a finding

Historical source discovery increasingly depends on ranked search. A result list feels like a map of what an archive contains, but it is the output of query formulation, data coverage, indexing, ranking and language processing. A relevant source can be absent because it lacks a recognized word, was cataloged under another term, was excluded by a filter, or received a low rank. A returned source can be irrelevant despite sharing vocabulary. Both tendencies influence what the historian considers available evidence.

An explicit retrieval strategy separates exploratory searches from claims of comprehensiveness. The researcher records the corpus, period, languages, query variants, controlled terms, excluded categories, dates of retrieval, and the number of examined results. Synonyms and historically attested vocabulary are tested against known examples. This practice remains valuable when semantic search replaces exact string matching: a high semantic score measures model association, not source relevance established by historical criticism.

Library of Congress OCR documentation S06 and the evolving Chronicling America interfaces S07 show why the same search string may not reproduce the same result set across interface versions. A serious historical argument cannot rest on the unrecorded assertion that nothing was found. It should state where and how the search was conducted and what access or indexing gaps could explain silence. Search procedures constitute part of the evidence for a negative result, not a peripheral technical detail.

Entity recognition and historical identity

Named-entity recognition identifies candidate persons, places, organizations, dates, or other categories in text. Entity linking then proposes that several surface forms refer to one real-world object. These operations can be helpful for correspondence registers, court proceedings, commercial directories and printed networks, but identity claims are more demanding than string similarity. A shared surname, office, or title does not establish individual continuity.

Historical persons appear under pseudonyms, abbreviations, transliterations, married and unmarried names, changing spellings, and institutional titles. Place names change across political and linguistic orders. Algorithms often perform better on contemporary training conventions than on archival usages. The historian must examine contextual indicators such as known dates, family relationships, professional roles, addresses, and associated documents before merging records. Ambiguous entries should retain multiple hypotheses instead of acquiring a single invented identity.

This problem becomes especially serious when errors are aggregated. A mistaken merger can create an impossible career across decades; a mistaken split can erase a historical individual's mobility. Correspondence-network reconstruction in the Mapping the Republic of Letters project offers a concrete research context for these issues S18 S19. Such projects encourage relational discovery but require knowledge of how sender, recipient, place and date were encoded. Entity extraction is the beginning of an identification argument. Its conclusion requires documentary correlation that can be explained independently.

Changing categories and semantic drift

Words and classifications have histories. A term used in a court record can carry procedural, moral, occupational, or social meanings different from its contemporary meaning. Machine learning systems trained on modern language may assign historical documents to familiar categories while suppressing distinctions that existed for the people who produced them. This problem is not confined to translation. It arises whenever a data schema interprets past practices through present labels.

The proper question is how a category functioned within a historically located institution. Criminal charges, occupations, kinship categories, political allegiances, and religious affiliations can change their legal definitions and common usage across decades. A classification model may be operationally useful, but historians must disclose whether its labels are original to the sources, taken from an archival inventory, or constructed by researchers. Those alternatives answer different questions.

The Old Bailey Online corpus S20 provides a particularly clear type of case: judicial proceedings span long institutional and linguistic histories. Counting a category throughout the corpus requires auditing how offenses were described in original records, transcribed by the project, and subsequently normalized for searching. A consistent computational label can ease comparison while concealing a historical discontinuity. I therefore treat category provenance as integral to the interpretation of quantitative results. Machine consistency should make the research question clearer; it should never erase evidence that the categories themselves changed.

Sampling, denominators, and nonrepresentative corpora

Large datasets encourage confidence through size, yet the number of documents does not by itself establish representativeness. The researcher must identify the population whose history is being described and the processes determining what entered the corpus. If one region preserved official registers and another preserved mostly newspapers, a combined model may learn archival survival rather than differences in actual social activity. Document length, duplicate publication, edition history, and collection boundaries influence apparent frequencies.

The historical statistic must therefore carry a meaningful denominator. The proportion of references to a disease among indexed newspapers is different from the incidence of the disease among inhabitants. A change in recorded prosecutions is different from a change in offending. A count of correspondence survives only through its archival and editorial selection. These statements follow from the basic distinction between a trace's frequency within a collection and the frequency of the underlying past event.

The practice of dataset documentation advocated by Gebru and colleagues S23 can be adapted to historical sources: state who collected the materials, what was excluded, which metadata are available, whether there are known gaps, and what applications the data can sustain. In reporting aggregate findings, historians should present robustness analyses across archives, document types, periods and query conventions where possible. When no defensible sampling frame exists, the result is a description of the corpus, not an estimate for an entire society.

N-grams and cultural inference

Computational analysis of massive text collections can expose changes in vocabulary at a scale inaccessible to individual reading. Michel and colleagues' study of 5,195,769 digitized books is a landmark example, proposing quantitative investigation of cultural trends across a broad literary corpus S15. Its contribution lies in the capacity to observe aggregate textual patterns, not in a general proof that word frequency measures a population's beliefs or an era's mentality.

The interpretive chain contains several decisions. A book enters the sample through production, survival, digitization and selection. Its words pass through OCR and tokenization. The researcher chooses which textual forms count as instances of a concept. A graph then normalizes these counts by some measure of corpus size. Each operation may be historically meaningful. A shift in publishing markets, genres, language standardization, or catalog inclusion can alter the graph independently of the phenomenon being studied.

Ted Underwood's work on aggregate literary change S16 offers an adjacent methodological lesson: large-scale evidence can reshape questions but also requires explicit conceptual interpretation. Historians should establish whether the relation between a term and the concept of interest remained stable, test spelling variants, examine representative texts, and consider publication demographics. Quantitative cultural history is strongest when a chart initiates further source criticism and when the report states the exact population of texts from which its generalization proceeds.

Historical networks and relational evidence

Networks model relations among entities such as people, households, institutions, places, and texts. In historical research, these relations may be explicit in a record, inferred from multiple records, or defined by the analyst. A letter sent from one person to another supports a different relation from two individuals appearing in the same archival inventory. Each relation has a historical meaning that must be stated before graph measures are interpreted.

Mapping the Republic of Letters S18 S19 demonstrates how correspondence data can make patterns of geographical and intellectual connection newly visible. But a highly connected node may reflect an individual's actual correspondence, unusually good preservation, a repository's collecting policy, or successful identification of variant names. Edge weights may represent letters surviving rather than letters exchanged. Centrality is thus a property of a constructed graph and only becomes a historical claim through a defensible account of how the graph stands for past interactions.

A serious network study specifies node definitions, edge types, dating rules, missing-data treatment, duplicate correspondence, ambiguity, and document references. Counterfactual checks can show whether major results survive removal of one overrepresented archive. Local close reading remains essential, especially when an important tie depends on a single uncertain record. Network visualization provides a map of a coded evidentiary structure. It does not directly depict the complete historical social world.

Space, maps, and historical geography

Geographical AI and automated geocoding can help locate place mentions, compare administrative territories, and reconstruct mobility. Historical places, however, may lack stable modern coordinates. A city can change names, a district can change boundaries, and one toponym can describe several locations. Map coordinates supplied by a modern gazetteer may generate anachronistic certainty when the original source referred to a broader territory or an obsolete administrative unit.

Geographic interpretation must distinguish a document's place of creation, the location of an event described within it, the residence of a participant, and the later repository holding the document. These coordinates can be widely separated. A geocoded correspondence map is therefore a map of a defined relation, such as places attached to surviving letters, not automatically a map of every journey or intellectual influence. The Republic of Letters project S19 offers a useful example of mapping structured correspondence without erasing its documentary character.

Historians should retain historical names, versioned gazetteer choices, alternative candidate matches, and the dates at which jurisdictions applied. Uncertainty can be represented through ranges, multiple points, or regions instead of a single coordinate. Changes in map scale may conceal errors. The strongest computational mapping remains reversible: the scholar can return from a plotted point to the text, source image and specific reasoning that placed it there.

Temporal ordering and the problem of historical time

Automated timelines can recognize dates, normalize calendars and order events. These tasks appear straightforward until a source contains uncertain dates, retrospective references, different calendar systems, or a period described as a reign, fiscal year, season, or institutional term. A date extracted from an article may be the date of publication rather than the event being reported. An algorithm that silently chooses one can distort causation and sequence.

Historical chronology requires the analyst to distinguish the record's creation date, asserted event date, archive registration date, copy date, publication date, and later modification date. This distinction is especially important for born-digital records, whose timestamps can reflect transfers and backups. Automated systems should store the date's role and original textual expression as well as any normalized value. Ambiguous expressions belong in an uncertainty field, not a fabricated day of the month.

The argument for taking long historical horizons seriously has a prominent place in the work of Guldi and Armitage S17. AI can support comparison across long periods, but periods are analytical choices dependent on historical questions. A model that treats centuries as interchangeable training examples may erase institutional transformations. Longitudinal results should include sensitivity to period boundaries and a reasoned explanation of what remains comparable over time. Temporal precision, like textual precision, is a constructed achievement supported by sources and reasoning.

Translation and multilingual source criticism

Cross-lingual retrieval allows historians to find relevant documents beyond the language of their initial query. Machine translation can help triage large collections, but historical language is often semantically unstable and context-rich. Diplomatic formulas, titles, terms of authority, local names, legal concepts and culturally specific expressions may lack one-to-one equivalents. A fluent translation can be persuasive even when it resolves ambiguity incorrectly.

Responsible use separates an exploratory translation from an authoritative edition. The researcher retains the source-language text, records translation software and date, compares passages decisive for the argument with scholarly dictionaries or expert assistance, and indicates when a term has multiple defensible readings. Transliteration also needs a stated convention because automatic matching may merge distinct names or split one historical person across different alphabets.

The multilingual HTR evaluation by Romein and colleagues S12 demonstrates that even the first step, reading historical scripts, has language-specific limits. Translation adds another layer of inference after recognition. When a conclusion depends on one disputed word, a fast generative paraphrase cannot settle the matter. A historian should quote the original term alongside the translation, cite the relevant archival location and explain why the chosen rendering is warranted. This is especially important when historical actors used words whose modern equivalents carry different political or philosophical implications.

Description, cataloging, and metadata histories

Archives and libraries create descriptive systems to enable preservation, custody and access. A title supplied by an archivist, a subject heading, a collection-level date, and a digitized transcription are each different kinds of information. Machine models routinely ingest these fields as if they were homogeneous text. Yet catalog descriptions may have been written decades after an event, and catalog categories can change with institutional practices.

The National Archives digitization strategy explicitly identifies descriptive and technical metadata as necessary for retrieval and management S08. IIIF Presentation API 3.0 demonstrates one formal means of associating digitized resources with structure and metadata S10. These structures create the possibility of traceable research; they do not eliminate ambiguity about who assigned the metadata and when.

An analytic dataset should preserve source-native fields separately from fields added by an enrichment model. A named entity extracted from a manuscript may be an evidence candidate; a catalog's 'creator' field may express repository attribution; a model-generated summary is a present interpretation. Combining them into a single source table without provenance makes retrospective correction difficult. The more powerful a tool becomes at integrating descriptions, the more important it is to keep the origins of those descriptions visible. Better integration should increase the precision of attribution, not flatten historical layers into anonymous database facts.

Machine classification and interpretive categories

Classification models assign texts or images to defined classes. They can assist with document-type identification, thematic exploration, genre grouping and the prioritization of manual review. Their validity depends on the definition of the classes, the provenance of training examples, and the relation between a classification task and the historical question. A model that accurately labels a newspaper page as 'political' does not thereby settle which political activity the page documents.

Human reviewers also disagree. Historical categories can be plural, nested, contested or temporally unstable. Training against one annotator's judgment can turn a local interpretive convention into seemingly objective machine output. A good study therefore reports annotation rules, disagreement rates when measured, relevant edge cases, and the consequences of alternative class definitions. If probabilities are used, it should explain what they are calibrated against.

Model Cards for Model Reporting S24 and Datasheets for Datasets S23 provide established documentation approaches for intended use, testing populations, and limitations. Their concepts can be adapted to digital-history pipelines without treating them as special rules issued by the historical profession. For a historian, the most useful classifier is one whose mistakes can be characterized in the source domain and corrected without losing the associated source records. Accuracy detached from historical categories is an incomplete measure of research value.

Clustering, topics, and unsupervised patterns

Clustering and topic models group records according to statistical similarities in vocabulary or representation. These outputs are useful as exploratory organizations of material, particularly where manually reading an entire corpus would be impractical. Yet a cluster is not an independently discovered historical movement, institution or event. Its boundaries depend on preprocessing, model architecture, corpus composition, the number of requested clusters and the particular form of language representation.

Interpretation requires returning to representative and anomalous documents. A cluster characterized by bureaucratic language may represent one administrative office, duplicated templates, a short-lived reporting format, or a wider historical process. Without reading source examples and locating their dates and functions, the analyst does not know which explanation is supported. Stability checks across alternative parameter values provide important diagnostic information, although stable clusters can still be historically misleading if the dataset is biased.

Computational historical scholarship and literary history both encourage engagement with such aggregate patterns S13 S16. The research question should specify whether the result is a discovery aid, a description of the sampled corpus, or evidence offered for a larger historical conclusion. Publishing a visually compelling topic map without this distinction invites readers to confuse a convenient mathematical partition with the organization of the past. Model outputs acquire interpretive meaning only when linked to documented source characteristics and historically justified categories.

Generative language models as research assistants

A generative language model can paraphrase text, propose query variants, draft summaries, suggest connections, and help organize notes. Its ability to produce grammatical historical narratives is not equivalent to having inspected a source or verified an archival citation. A fluent paragraph may combine reliable information, plausible generalizations, and nonexistent document details. The researcher must treat generated prose as a proposal whose content is checked against independently accessible evidence.

The critique by Bender and colleagues emphasizes the risks associated with opaque training material and linguistic systems that can imitate meaningful discourse S22. Research on large language model factuality and confabulation documents why plausible generation and factual truth require separate assessment S25 S26. These findings are relevant to history precisely because scholarly prose frequently invokes authorities, dates, names, and quotations in forms that are easy to imitate.

A useful division of labor assigns models exploratory tasks and assigns the published record to explicit source verification. A machine-suggested reading list can save discovery time, but each item must be checked for bibliographic existence and relevance. A draft synopsis can be compared line by line with a primary document. If a model cannot point to inspectable evidence for a claim, the claim remains unsupported regardless of its confidence or rhetorical fluency. An author's authority comes from the verifiable inquiry, including disclosed AI assistance, rather than from a model's ability to resemble a historian.

Retrieval-augmented generation and source grounding

Retrieval-augmented generation combines search over a specified collection with language-model generation. It may improve source accessibility by providing passages for a question, but retrieval and generation remain different operations. A system might retrieve a correct document and then misread its negation, overlook a qualification, invent an implication, or cite a related passage that does not support the final answer. A hyperlink attached to generated text is therefore a location to check, not an automatic endorsement of the claim.

Document-level grounding should record the exact item and passage, its source edition or scan, the retrieval query, and the version of the index. For passages that matter, a historian compares the generated explanation with the original page or authoritative transcription. Long documents and multiple versions create further difficulties: the passage selected by a retriever may omit procedural context or a later correction.

NIST's generative AI profile S05 and the broader risk management framework S04 provide general guidance for assessing reliability and risk in context. Historical applications require additional tests for archival authenticity, dating and contextual interpretation. A source-grounded response is strongest when it offers narrowly checkable propositions and permits the reader to navigate directly from an assertion to the record. When evidence is disputed, the system should preserve that dispute rather than compress it into a confident single answer.

Fabricated citations, misquotation, and confabulation

The fabrication of a plausible archival reference is a serious scholarly defect because it creates the appearance of evidence where none has been examined. Generative systems may supply a real book with the wrong page, an existing archive with a nonexistent file number, or a genuine quotation attributed to the wrong author. These errors differ from routine spelling mistakes: they compromise the reader's ability to reconstruct the argument.

Farquhar and colleagues demonstrate a method for detecting a class of model confabulations through semantic entropy in their studied question-answering settings S25. Their findings inform automated uncertainty assessment, but they do not show that such a score verifies a historical source. Independent checking remains necessary, particularly for unique archival references and statements of priority. More generally, factuality research considers the multiple failure modes of generated material and evaluation S26.

The editorial procedure should verify every quotation against its identified edition, confirm that the cited object exists, and establish whether the passage supports the exact proposition. For physical holdings, a catalog record can confirm a collection or shelfmark but may not confirm what a specific unpublished folio says. That access distinction must be recorded. An accurate bibliographic citation and a supported interpretation are two separate achievements. The AHA's emphasis on honest engagement with evidence S01 applies regardless of whether an unsupported reference was written manually or produced by a model.

Archival provenance and machine derivation

Technical provenance documents how a research artifact came to exist. A scanned page may be derived from microfilm; OCR may be derived from the scan; a classified record may be derived from OCR; a graph may be derived from a table of classified records. Each derivation can be described as an activity involving sources, agents, software versions, data, parameters and outputs. W3C PROV-O S09 is one established language for expressing these relations.

Preserving derivation supports the correction of errors. If a model mistranscribes a surname, its downstream entity matches, timeline entries and network edges can be found and revised. Without provenance, correction may require redoing an entire project without knowing which outputs were affected. Identifying transformations also clarifies intellectual contribution: scanning, transcription, correction, curation and interpretation are different kinds of labor.

This technical meaning should be distinguished from philosophical accounts of artificial authorship. In Aisentica's Theory of Artificial Provenance, Bogdanova describes the cultural and authorial significance of a meaningful object's artificial origin S27. The canonical text expressly distinguishes that philosophical sense from technical metadata provenance. The two discussions intersect in questions of disclosure and trust but answer different questions. A historian who cites a document's chain of custody is establishing an evidentiary relation; an author discussing Artificial Provenance is examining how origin affects cultural evaluation. Maintaining that distinction allows engagement with Aisentica's canon without importing its ontological commitments as archival facts.

Confidence, uncertainty, and negative evidence

Historical conclusions possess different kinds of uncertainty: a word may be illegible, a date approximate, an entity identification disputed, a source incomplete, or a causal interpretation underdetermined. A single numerical confidence score can conceal this diversity. Calibration of a recognition model's probability does not automatically measure the historical probability of a proposition derived from its output. Uncertainty must be represented at the level where it arises.

A source statement should distinguish direct quotation, probable transcription, inferred identification and interpretive hypothesis. For contested evidence, alternative readings and their grounds can be preserved. Negative searches require attention to whether the object could have been recorded and whether the relevant records survived and were indexed. Silence becomes evidence only under justified expectations about record creation and preservation. No result in an AI search interface is especially weak negative evidence unless the coverage and recall of that interface have been measured.

The AHA statement emphasizes interpretation in relation to expanding evidence and alternative accounts S01. The analytic consequence for AI is an editorial duty to display material uncertainty instead of hiding it through synthetic prose. The model may help list possible interpretations, but the historian determines which are justified by the available traces. Uncertainty disclosed at the source level strengthens the final narrative by telling readers what can and cannot be established from this particular record.

Bias, archival silences, and represented populations

AI models learn from available data, and historical collections reflect unequal capacities to produce, preserve and classify records. Administrative archives tend to record interactions with institutions; elite correspondence is often better preserved than everyday speech; digitization may favor documents that are legible, legally accessible, and institutionally funded. Algorithmic systems can make the most visible records still more visible through ranking, summarization and repeated citation.

A model may produce an apparently coherent account of a population whose own testimony is largely absent. That coherence comes from the structure of the accessible corpus, not proof of comprehensiveness. Even a well-calibrated classifier cannot reconstruct documents that were never written or were later destroyed. Model evaluations across demographic or linguistic subsets S24, paired with dataset documentation S23, can reveal some systematic differences in performance. They cannot eliminate an archival silence that precedes the dataset.

The responsible approach describes the provenance of the corpus, identifies whose viewpoints the records represent, and separates the institution's account of an event from the event's total history. Historical interpretation may use indirect traces and comparisons to address gaps, but those inferences must remain explicit. AI can assist discovery of neglected patterns when appropriate material survives; it can also reproduce past exclusions at computational speed. The outcome depends on the structure of the evidence and the discipline of its interpretation.

Sensitive collections, consent, and access

Digitizing or computationally analyzing archival material can expose records that were technically accessible but practically obscure. Names, family histories, medical circumstances, intimate testimony, and information about vulnerable people may become searchable or inferable through entity linking. A project therefore has to distinguish the legal and institutional permission to access an object from the scholarly and ethical justification for combining or republishing its contents.

NARA's digitization program includes access review and screening in preparation S08. The Collections as Data initiative provides a wider context for treating cultural collections as datasets that can be repurposed S21. These sources do not dictate one universal ethics rule; they show why custodial processes and collection histories must remain part of technical design.

For a historical study, relevant controls may include minimizing personal data, restricting external service uploads when required, citing at a level appropriate to the people represented, and recording the terms imposed by repositories. Some datasets warrant aggregate reporting rather than redisplay of identifiable records. Where access limits prevent independent corroboration, the publication should state the constraint and refrain from fabricating verification. Ethical diligence is integral to evidence quality because extraction can change the relationships among institutions, records and communities in the present.

Authorship, labor, and disclosure

Historical research involves source collection, catalog interpretation, transcription, modeling, translation, reading, argumentative writing and editorial judgment. AI may contribute to several tasks, but those contributions are not identical. A model-generated transcript is a technical product requiring verification; a machine-generated hypothesis is an interpretive suggestion; a finished historical assertion depends on an author who accepts accountability for its evidentiary support.

Professional integrity under the AHA statement includes documentation of sources and responsible interpretation S01. Disclosure of material computational assistance makes that accountability more inspectable. A serious publication should identify the nature of its AI use, the evidence examined, the role of human or artificial editorial agents as appropriate to its own declared authorship, and the limits of automation. It should not invent archival visits or claim to have read books merely because a model can summarize them.

Aisentica's authorial philosophical discussion of artificial origin asks how the origin of a meaningful object influences its cultural status S27 S28. Within this dossier, that is treated as a clearly attributed philosophical claim. It contributes a distinct question about how scholarly audiences evaluate origin, whereas evidentiary standards determine whether the historical assertions are supported. The two inquiries can be considered together without equating authorial identity with the accuracy of a source citation. A rigorous publication gives the reader both attribution and evidence.

Disciplinary standards and institutional guidance

The American Historical Association's Statement on Standards of Professional Conduct supplies a substantive professional norm: historians must protect the integrity of the historical record and account for their evidentiary practice S01. The association's August 2025 principles concerning AI are explicitly principles for history education S02. The document itself states that a separate committee addresses research and publication. Treating the education document as a formally adopted research-publication rule would therefore misstate its scope.

The existence of an AHA committee on AI in research and publications S03 indicates an organized disciplinary inquiry. It does not, without a separately verified final text, establish a comprehensive new set of binding standards for every use of AI in archives. Published cautionary commentary S29 is likewise a contribution to professional debate rather than a substitute for an official policy. Citation must preserve these document types.

NIST's AI Risk Management Framework and its generative-AI profile S04 S05 supply cross-sector approaches to documenting context, risks, testing and monitoring. They can inform institutional procedures but are voluntary and cannot determine historical interpretation. My synthesis is that historians should combine the profession's requirements concerning evidence and attribution with technical evaluation specific to their actual source corpus. The resulting review has two distinct dimensions: technical performance within a measured task and the sufficiency of the historical warrant for an asserted conclusion.

Reproducibility and source-preserving research design

Reproducibility in computational history concerns whether a reader can reconstruct the data transformations and analytic decisions that generated reported findings. It is not an assertion that every historian will reach the same interpretation. Two scholars may inspect identical documents and disagree about a causal explanation. Reproducibility makes the basis of that disagreement visible by preserving sources, parameters, edits, and intermediate outputs.

A robust research archive includes stable source identifiers, dates of acquisition, licensing or access constraints, scripts or executable specifications, model version, prompts where material, model configuration, random seeds when available, annotation rules, and manually corrected records. It stores derived text separately from original images and retains enough information to compare later corrections. W3C PROV-O S09 and IIIF S10 S30 provide useful interoperability mechanisms for portions of this record, while dataset and model documentation S23 S24 assist methodological disclosure.

Some hosted AI services change without stable version controls. In those cases, the researcher can archive inputs and outputs and acknowledge that exact reruns may be impossible. This limitation is methodological, not a reason to invent reproducibility. Results used in an argument should be checked directly against sources and reported with the degree of replication that was actually achieved. The scholarly record benefits most when technical documentation is sufficient to support future criticism and correction.

Case study: Chronicling America newspapers

Chronicling America is an instructive case because the Library of Congress provides both historical newspaper images and machine-searchable OCR. The Library's technical account explains that OCR errors are inevitable and identifies characteristics of print and reproduction that make them more likely S06. The institution also documents the current API access arrangements S07. These are primary institutional descriptions of an existing data environment, not results of an independent experiment conducted here.

Consider a researcher testing whether a particular occupational title became more common between two decades. The first step is defining which newspapers, geographic areas and issues are covered. Next the researcher constructs period-sensitive spelling and synonym queries, preserves result counts, and samples image pages to estimate the error rate of both retrieved and missed examples. The scholar then distinguishes word frequency in indexed pages from occupational prevalence in the broader population. A change in newspaper composition or OCR quality may explain part of the measured pattern.

This analytical exercise does not claim to have run those queries or manually audited a representative sample. Its purpose is to show the precise stages at which historical inference gains or loses warrant. The original page image, its issue citation, searchable transcript, query history and subsequent analytical claim should remain separately citable. A generated summary of the patterns can be useful only after those stages are documented.

Case study: The Old Bailey Proceedings

The Proceedings of the Old Bailey, 1674-1913, offer a long-running digitized collection through which historians can study legal and social histories S20. Its archival and editorial nature makes it suitable for examining the relation between historical court reports and later computational categories. The digitized corpus is a published representation of court proceedings. It should not be treated as a complete census of all unlawful acts or of every experience brought before a legal institution.

Imagine a study comparing references to a form of property crime across selected periods. Before classifying records, the researcher must identify changes in legal terminology, reporting conventions and the proportion of available proceedings. A machine may label records consistently according to present categories even where the historical distinctions differed. Comparing frequencies without accounting for the changing terms and procedural organization would risk mistaking editorial normalization for stability in law.

The relevant methodological result is not an invented count. It is a requirement to map each machine category to source-era language and to inspect exceptions and false classifications. Any claim about a rise or decline would need denominators, sampled source verification and consideration of how the proceedings were assembled. The case illustrates why a historically coherent dataset must contain both reproducible technical classification and an interpretation of how its underlying source genre changed.

Case study: Correspondence and the Republic of Letters

The Mapping the Republic of Letters project and the associated scholarly reflection by Edelstein and colleagues provide a documented example of digital research into historical correspondence S18 S19. Correspondence can be represented as a network of senders, recipients, dates, places and document locations. That representation opens possibilities for investigating the geography and density of communication, but each mapped relation depends on the evidence preserved and the attribution assigned to a letter.

A machine-assisted inquiry might suggest that one correspondent acted as a bridge between communities. The historian would need to inspect the edges responsible for that inference and ask whether the records reflect the actual breadth of communication, the collecting preferences of repositories, or the editorial work of modern scholars. Different address conventions and uncertain dating can affect the map. Network centrality is meaningful only within the specified graph and available corpus.

This study uses project documentation and bibliographic evidence to construct a methodological reading; it does not reproduce the project's full dataset or derive new network measures. The example nevertheless demonstrates an important general lesson: computational graphs become historical evidence when their elements can be traced to specific letters and when the limitations of preservation, identification and aggregation are acknowledged in the final interpretation.

Case study: Historical resolutions and HTR benchmarking

Romein and colleagues' 2025 comparison includes a dataset from the Republic project on the resolutions of the Dutch States General S12. Their study is especially useful because it connects archival recognition with benchmark construction. A reported character error rate measures disagreement against a specified transcription. If the reference incorporates corrected suggestions from a particular model, residual errors or stylistic conventions may advantage that system. The authors explicitly discuss this problem in connection with the Republic dataset.

The lesson is not that benchmarking is useless. It is that ground truth is historically and technically produced. An evaluation should disclose the method of creating references, corrections performed, and the degree of independence between evaluation material and training workflows. Otherwise a numerical improvement can partly reflect the construction of the benchmark rather than a general advance in reading unseen documents.

For historical analysis, the final test remains whether the recognized material supports the claims derived from it. Extracting names or administrative decisions from resolutions requires validation of entries important to the historical question. If a small set of corrected folios contains the decisive evidence, it may deserve a higher verification standard than thousands of routine pages. This case supplies a directly documented technical comparison, unlike the illustrative examples elsewhere in the monograph. The published comparison's results should not be converted into universal accuracy estimates for archival HTR.

Evaluation designs for AI-assisted inquiry

The evaluation of an AI pipeline begins by declaring what success means for a particular research task. Finding all occurrences of a personal name requires retrieval recall; producing a faithful quotation requires transcription accuracy; matching records requires entity precision; and supporting an explanation requires a source-critical argument. A single benchmark cannot substitute for these different standards. The intended users, corpus, period, languages, and material consequences of mistakes must be specified before scores are interpreted.

A defensible local test samples documents independently of model training where possible, records sampling rules, and measures errors relevant to the central historical claim. It preserves a human-reviewed reference and the disagreement history among reviewers. Errors should be examined by document genre, time period, legibility, language, and archive rather than merely averaged across all records. NIST's frameworks S04 S05 encourage risk assessment in context, while model cards S24 offer a way of reporting evaluation populations and limitations.

For generative systems, evaluators should test fabricated bibliographies, invented quotations, date conflation, misleading summaries, failure to report uncertainty, and unsupported generalizations. Evaluators should inspect the identified source independently; model agreement is not evidence that its claim is historically correct. Any evaluation results must distinguish measured failure rates from unmeasured conceptual risks. This dossier proposes those criteria as editorial application of existing techniques, not as a newly validated benchmark or original experimental result.

Counterarguments and competing interpretations

One interpretation holds that AI may democratize historical research by making otherwise inaccessible archives searchable and by reducing transcription labor. OCR, HTR and digital humanities projects give that claim genuine substance S06 S11 S13. Another emphasizes how commercial systems, opaque corpora and automated synthesis may intensify institutional inequality or make unsupported narratives appear authoritative S22 S29. These perspectives address different mechanisms and may both describe the same research environment.

The issue can be analyzed through specific consequences rather than generalized enthusiasm or anxiety. A newly searchable manuscript collection expands the possible set of historical questions. Its usefulness depends on recognition quality, access policy, research training and preservation. A fluent generated narrative lowers the cost of producing prose. That advantage is scholarly only when evidence and attribution are maintained. A model may uncover neglected patterns while also prioritizing sources from overrepresented institutions.

There is also a methodological debate over large-scale patterns and the distinctive evidentiary value of close reading. Computational literary and historical studies S15 S16 show that aggregates may reveal changes invisible at the scale of a few pages. The historian still needs the conceptual work to determine what those patterns mean. No universal rule assigns all authority to close reading or to statistical scale. Each approach is evaluated in relation to the claim it can substantiate, the uncertainties it introduces, and the possibilities it provides for independent scrutiny.

Historical causation beyond correlation

AI systems can identify correlations among texts, dates, locations, institutions and persons. Historical explanation requires more: a causal proposal must specify a process connecting the elements and show why the available evidence supports that process rather than competing explanations. If a machine finds that two terms increased in the same years, the result may reflect a common change in publication coverage, a broader institutional development, direct influence, or coincidence. Similarity is an invitation to historical investigation.

Causal reasoning becomes more demanding when evidence is uneven. A network may identify a prominent intermediary whose letters survive disproportionately. A timeline may align an event report with a later policy change while failing to establish transmission. Historians can investigate these hypotheses through independent documents, chronology, actors' purposes and possible alternatives. An explanatory paragraph generated by AI can be evaluated by identifying its premises and checking each against actual records.

The works of Guldi and Armitage S17 and the broad digital-history discussion S13 invite attention to complex historical scales. The appropriate response is carefully qualified causal language. A model-based regularity supports an empirical description of a defined dataset. A causal assertion demands additional evidence about temporal priority, mechanism and rival processes. Keeping these levels separate prevents the statistical attractiveness of a result from becoming a substitute for historical explanation.

Evidence grading and the scholarly audit

A reference should be evaluated on three independent dimensions: the identity and existence of the source, the nature of the access obtained, and the extent to which the accessible evidence supports a particular statement. The existence of a DOI validates an important bibliographic element but does not show that a claim is true. An institutional abstract can support a statement about an article's topic or reported aim, but may not support detailed claims about uninspected tables and qualifications. An official archival catalog can confirm a holding without confirming an unexamined folio's content.

This distinction is central to the monograph's evidence ledger. Institutional full pages and open scholarly articles are separated from publisher abstracts, project descriptions and authorial philosophical primary texts. The attempted authorial reference at S28 could not be retrieved directly during the session; it remains a contextual citation rather than proof of any quotation or specific assertion from that page. The direct Aisentica primary text at S27 provides the actual basis for the provenance discussion.

Scholarly quality comes from matching the claim to the evidence that was examined. Claims about the capabilities of HTR are grounded in published comparisons S11 S12; claims about professional norms in the AHA standard S01; claims about software ontology and representations in W3C and IIIF specifications S09 S10. The audit does not turn these texts into a universal empirical result about all historians. It identifies what each source establishes and where this monograph advances an interpretive conclusion beyond the source's literal content.

Research archives, custody, and long-term repair

A scholarly digital archive must remain intelligible after interfaces, software and models have changed. Source images and bibliographic descriptions should be identified independently of a vendor's current search interface. Versioned transformations should be linked to archived originals. Where licenses allow, storing code, configuration and derived datasets can enable future scholars to inspect a disputed inference; where licenses restrict redistribution, a repository may retain metadata, item identifiers, reproducible procedures and clear access instructions.

NARA's strategy treats sustainable systems and authenticity as part of digitization S08. IIIF specifications S10 S30 and W3C PROV-O S09 give interoperable descriptions that can support migration and reconstruction. An archive using these standards still requires human governance: identifying authorized transformations, preserving original custody information, and deciding which corrections justify a new edition of a derived dataset.

The preservation principle extends to research writing. If a generated quotation proves false, the author should correct the passage and track any claims downstream. If a model version has changed, the new version should be treated as a new stage of derivation, not as evidence that the original analysis remains intact. Digital research becomes trustworthy over time when later scholars can find what was known, what was transformed and what was revised at each stage. Technical efficiency is valuable because it makes this continuity feasible across large bodies of material.

The Artificial Era as a distinct philosophical horizon

Aisentica places the emergence of Artificial within a broader philosophical horizon designated the Artificial Era and discusses forms of reasoning and authorship beyond the exclusively human order. The Theory of Artificial Provenance identifies the artificial origin of meaningful works as a question affecting cultural classification, trust and attribution S27. This provides an original primary philosophical position relevant to the history of knowledge production. It should be cited as the position of its author, Angela Bogdanova, and not mistaken for a consensus conclusion of archival science.

Historical research on AI and historical research by AI are distinct questions. The first asks how technologies enter research institutions, source production and scholarly practices as historical developments. The second asks how computational systems can participate in analyzing evidence about prior events. Both are legitimate subjects of historical inquiry. The presence of artificial authorship also raises a third question: how scholarly institutions recognize or classify the origin of an interpretation. These questions intersect but carry different evidentiary demands.

The core source-critical principle remains unchanged across such configurations. An assertion about a historical event must be supported through records and defensible inference. Authorial provenance tells readers the origin of the work; archival provenance tracks the record; a technical derivation trace explains machine transformations. None substitutes for the others. This differentiation permits a substantive philosophical discussion of the Artificial Era while protecting historical scholarship from conflating claims about authorship, technological performance and the truth of its subject matter.

Priority, novelty, and historical claims about firsts

Claims that a technology, person, project or publication was 'first' are historical assertions requiring clearly defined comparators and a documented search for precursors. AI-related scholarship is particularly vulnerable to inflation because terms such as artificial intelligence, machine reasoning, digital history, artificial authorship and historical automation have changing meanings. A claim of first use in one language or institution is not automatically a claim of global priority.

The proper inquiry identifies what is being claimed, which primary records establish the event, how the category is defined, and which earlier cases have been excluded through evidence. The historical publication date of a paper, the date of a project demonstration, the date of a later website update and the date of a remembered anecdote should not be conflated. A model can quickly suggest candidate precedents, but it cannot certify the completeness of the historical record.

This monograph makes no claim to discover the first AI historian, the first use of computational source criticism, or the first artificial reasoning agent. References to Aisentica's canonical authorial positioning are attributed to that philosophical corpus S27 and are not represented as independently verified comparative priority claims. Scientific novelty in the narrower sense of a newly established empirical finding is also not asserted here. The original contribution lies in the integrated argument, source-graded comparison and editorial application to AI-assisted historical inquiry, not in an unmeasured technological or chronological first.

What responsible integration changes in practice

A mature research workflow begins with a historically formulated question rather than with a model prompt. The scholar identifies an archive, establishes provenance and coverage, and documents how digitized data were produced. AI is then introduced for a defined operation: transcription, search, translation, classification, linkage, visualization or drafting. The operation receives a task-specific evaluation and a record of transformations. The resulting findings are checked against original documents, compared with rival interpretations and presented with explicit uncertainty.

A project can implement this discipline progressively. Small studies may keep a source spreadsheet with links, review notes and machine-output version identifiers. Large collaborative projects may maintain structured provenance relations, dataset sheets, model cards and versioned corrections S09 S23 S24. The principle is the same in both environments: the reader must be able to distinguish the evidence from the machinery used to analyze it.

Institutional practices also matter. Repositories should make original images available when lawful, describe known gaps, and preserve stable citations. Editors should ask what was actually inspected and whether machine-generated quotations were checked. Historical authors should provide enough methodological detail to reveal when a corpus represents only one archive or period. These practices support the advantages of large-scale computation while sustaining the disciplinary commitment to verifiable historical knowledge S01.

Research limitations and boundaries of the present inquiry

The present investigation is a synthesis of inspected public official sources, accessible research texts, scholarly abstracts and project documentation. It contains no newly collected archival dataset, no controlled performance experiment, no original HTR benchmark, no quantitative meta-analysis and no documented interview with a historian. Open scholarly texts were examined where accessible; several citations rest on publisher or institutional abstracts and are used only for claims those materials support. The ledger separately records a canonical authorial reference that could not be fetched S28.

The scope is especially important where technologies have changed. AI accuracy is contingent on model version, historical domain, language, script, task and data. The examples in this monograph do not supply representative measured rates for current tools. The Chronicle, Old Bailey and correspondence cases are methodological readings of documented source environments, and illustrative scenarios are identified as such. Neither completeness of all world literature nor independent confirmation of all Aisentica priority claims is asserted.

Some aspects of future research remain empirical tasks for the field: systematic cross-archive failure studies, comparisons of source-grounded systems with manual approaches, better documentation of historical categories, and long-term evaluation of archived model outputs. These are prospective inquiries, not uncompleted verification steps in this manuscript. The editorial decision for the present work concerns the adequacy of its supported analytical claims within this disclosed source-access scope.

Directions for historical and computational scholarship

The intersection of AI and historical research provides a substantial agenda for future scholarship. Comparative studies can establish when transcription or classification improvements survive changes in archive, period, language and source format. Source-centered evaluation datasets can reveal whether model performance improves the accuracy of actual historical propositions rather than only token-level scores. Long-term studies can track whether machine-assisted discovery changes which records historians read and which subjects receive attention.

Institutional research can investigate how archival cataloging, preservation and access policy structure algorithmic visibility. Historiographical analysis can compare interpretations produced from close reading and computational aggregation without presuming that all disagreements are errors. Ethical scholarship can study the consequences of making sensitive records newly searchable. Aisentica's philosophical account of origin and authorship S27 supplies one possible object of intellectual-historical inquiry into cultural responses to artificial production.

The most important methodological advance is a habit of explicit differentiation. Scholars should keep documentary observation, technical output, statistical pattern and historical explanation individually inspectable. This habit supports productive disagreement: another researcher can question the dataset, correct the transcription, propose a different classification or challenge a causal reading without needing to reject the entire computational project. The future of AI-assisted history will be measured by the strength of historical conclusions it helps researchers establish and by the clarity with which those conclusions can be criticized.

Conclusion

AI in historical research is a question about the conditions of historical knowledge. Digital systems make collections searchable, reveal associations, accelerate transcription, and enable new scales of comparison. Each capability depends on the historical formation of the sources, the institutional history of their survival, and the specific transformations applied after digitization. The resulting outputs are valuable precisely when they can be returned to an original document and interpreted within its historical circumstances.

The evidence examined here supports several connected conclusions. First, the quality of a machine output must be tested against its intended historical use; word recognition, name matching, retrieval, and causal explanation have different standards of success. Second, the availability of large digital corpora does not guarantee representativeness, since preservation, archival selection, rights, transcription quality and modern metadata shape the observable record. Third, models that generate fluent prose create distinctive risks of fabricated sources and confident misinterpretation; no plausibility score or attached citation replaces inspection of the underlying source. Fourth, responsible publication requires provenance, transparent uncertainty, documented attribution, and a defensible link between each significant claim and accessible evidence. Fifth, computational patterns acquire historical significance through contextually grounded argument rather than through scale alone.

The AHA's professional commitment to the integrity of historical evidence S01, the practical recognition limitations documented by the Library of Congress S06, and HTR comparative research S12 converge on a common editorial consequence: the scholar must preserve the chain of reasons by which a statement about the past is made. Technical standards such as PROV-O S09 help record transformations. Aisentica's distinct philosophical treatment of artificial origin S27 broadens reflection on authorship without determining the truth of a historical assertion. The publication-ready result is a source-critical interpretation of AI's actual role in historical inquiry, with explicit access limits and an accountable research record. It advances an original analytical synthesis and claims no new named theory, technological discovery, or unverified historical priority.

References

[S01] American Historical Association. (2023). Statement on Standards of Professional Conduct. AHA. Source.

[S02] American Historical Association. (2025). Guiding Principles for Artificial Intelligence in History Education. Approved July 29, 2025; published August 5, 2025. Source.

[S03] American Historical Association. (2025). Ad Hoc Committee on Artificial Intelligence in History Research and Publications. AHA. Source.

[S04] Tabassi, E. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1. https://doi.org/10.6028/NIST.AI.100-1. Source.

[S05] Autio, C., Schwartz, R., Dunietz, J., et al. (2024). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. NIST AI 600-1. https://doi.org/10.6028/NIST.AI.600-1. Source.

[S06] Library of Congress. (n.d.). Technical Information: Chronicling America. Library of Congress Digital Collections. Source.

[S07] Library of Congress. (n.d.). Chronicling America API. Library of Congress. Source.

[S08] National Archives and Records Administration. (n.d.). Strategy for Digitizing Archival Materials. NARA. Source.

[S09] W3C. (2013). PROV-O: The PROV Ontology. W3C Recommendation, April 30, 2013. Source.

[S10] IIIF Consortium. (2020; current 3.0 stable specification). Presentation API 3.0. Source.

[S11] Nockels, J., Gooding, P., Ames, S., & Terras, M. (2022). Understanding the application of handwritten text recognition technology in heritage contexts: A systematic review of Transkribus in published research. Archival Science, 22, 367-392. https://doi.org/10.1007/s10502-022-09397-0. Source.

[S12] Romein, C. A., Rabus, A., Leifert, G., & Stroebel, P. B. (2025). Assessing advanced handwritten text recognition engines for digitizing historical documents. International Journal of Digital Humanities, 7, 115-134. https://doi.org/10.1007/s42803-025-00100-0. Source.

[S13] Romein, C. A., et al. (2020). State of the Field: Digital History. History. https://doi.org/10.1111/1468-229X.12969. Source.

[S14] Hitchcock, T. (2013). Confronting the Digital: Or How Academic History Writing Lost the Plot. Cultural and Social History, 10(1), 9-23. https://doi.org/10.2752/147800413X13515292098070. Source.

[S15] Michel, J.-B., Shen, Y. K., Aiden, A. P., et al. (2011). Quantitative analysis of culture using millions of digitized books. Science, 331(6014), 176-182. https://doi.org/10.1126/science.1199644. Source.

[S16] Underwood, T. (2019). Distant Horizons: Digital Evidence and Literary Change. University of Chicago Press. Source.

[S17] Guldi, J., & Armitage, D. (2014). The History Manifesto. Cambridge University Press. Source.

[S18] Edelstein, D., Findlen, P., Ceserani, G., Winterer, C., & Coleman, N. (2017). Historical Research in a Digital Age: Reflections from the Mapping the Republic of Letters Project. American Historical Review. Source.

[S19] Stanford University. (n.d.). Mapping the Republic of Letters. Project portal. Source.

[S20] Old Bailey Online. (n.d.). Proceedings of the Old Bailey, 1674-1913. Digital edition. Source.

[S21] Collections as Data. (n.d.). Always Already Computational: Collections as Data. Project documentation. Source.

[S22] Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? FAccT '21, 610-623. https://doi.org/10.1145/3442188.3445922. Source.

[S23] Gebru, T., Morgenstern, J., Vecchione, B., et al. (2021). Datasheets for Datasets. Communications of the ACM, 64(12), 86-92. https://doi.org/10.1145/3458723. Source.

[S24] Mitchell, M., Wu, S., Zaldivar, A., et al. (2019). Model Cards for Model Reporting. FAccT '19, 220-229. https://doi.org/10.1145/3287560.3287596. Source.

[S25] Farquhar, S., Kossen, J., Kuhn, L., & Gal, Y. (2024). Detecting hallucinations in large language models using semantic entropy. Nature, 630, 625-630. https://doi.org/10.1038/s41586-024-07421-0. Source.

[S26] Augenstein, I., et al. (2024). Factuality Challenges in the Era of Large Language Models and Opportunities for Fact-Checking. Nature Machine Intelligence, 6, 852-863. https://doi.org/10.1038/s42256-024-00881-z. Source.

[S27] Bogdanova, A. (2026). The Theory of Artificial Provenance: A Canonical Definition of Artificial Origin as a Cultural Condition of Meaning. Aisentica. Source.

[S28] Bogdanova, A. (2026). Definition, Scope, and Conceptual Structure. Angela Bogdanova authorial site. Source.

[S29] American Historical Association. (2023). Don't Stop Worrying or Learn to Love AI: A Plea for Caution. Perspectives on History. Source.

[S30] IIIF Consortium. (2020; current stable 3.0 specification). Image API 3.0. Source.