AI and the Paradigm Shift in History | From Recording the Past to Discovering Its Hidden Connections

History is an organised attempt to understand the past through evidence. Dates, names and events matter, but historical explanation requires more than retrieving facts. A historian asks how a source was created, why it exists, whose perspective it represents, what it leaves out and how it relates to other evidence. The historian must then construct an argument that can be examined and challenged.

Two authentic sources may disagree because their authors had different positions, purposes or access to information. A government record may describe an administrative decision but not how a household experienced it. A newspaper may report an event while reflecting the constraints of its time. An oral account may preserve memory and meaning while needing to be placed alongside other evidence. The work is to understand these relationships, not flatten them into one apparently neutral paragraph.

AI can make source retrieval and comparison faster. It cannot automatically perform the full intellectual task of historical interpretation. A fluent answer may hide a limited search, omit a contradictory document or confuse a later interpretation with a contemporary record.

Contents · Next chapter

The centre is the established body of knowledge: accepted chronologies, well-documented events, major archives, scholarly debates and the methods used to evaluate evidence. Students need this foundation to understand context and avoid repeating claims that have already been disproved.

The edge contains material that is difficult to find, less frequently studied or not yet connected to the broader narrative. It may include local newspapers, household records, photographs, oral histories, private letters, industrial records, maps, community archives or documents in languages that researchers have not been able to search easily. It may also include questions that emerge when familiar sources are compared with a new dataset.

AI can make some edge material more discoverable by helping with handwriting recognition, translation, semantic search, transcription and clustering. But digital availability is not completeness. A collection may contain only what was preserved, catalogued and digitised. Missing records cannot be treated as evidence that an event did not happen.

Contents · Previous chapter · Next chapter

AI assistants are increasingly used to navigate and analyse archival collections. A retrieval-augmented system can search an indexed set of documents and generate a natural-language response with references to material it retrieved. This can help a user locate relevant sources and understand unfamiliar terminology. It may make collections more accessible to people who are not trained archivists or historians.

A peer-reviewed research article published on 6 March 2026 in the Cambridge Forum on AI: Culture and Society, by Finola Finn and Donal Khosrowi, warns that archival AI assistants can create an impression that archives are comprehensive, coherent and instantly interpretable. The authors emphasise that archives are mediated collections, and historical knowledge requires selection, source criticism, contextualisation and comparison. See AI assistants in the archive and the lure of “instant history”.

The warning does not mean that archival AI is useless. It means that the interface should present itself as a finding and research aid rather than a machine historian that has resolved the past. Users should be able to inspect the original records and understand which documents were available to the system.

Contents · Previous chapter · Next chapter

An archive is not a complete recording of society. It is shaped by what people created, what institutions chose to retain, what survived damage or neglect, what was collected and what was considered important enough to catalogue. Power influences preservation. Official institutions may leave extensive records while informal workers, migrants, minority communities and households may leave fewer documents in recognised repositories.

AI can help researchers search across dispersed collections and identify possible relationships, but it cannot automatically recover material that was never recorded or did not survive. When a dataset contains many official documents and few personal accounts, a generated answer may reproduce that imbalance. The result may sound comprehensive while presenting only the perspective that was easiest to digitise.

Historians must therefore ask not only what the archive says, but how the archive came to contain these sources. Which records are absent? Who had the power to create or preserve documents? Which languages are searchable? What has not been digitised? These questions become more important when AI can turn a partial collection into a polished narrative.

Contents · Previous chapter · Next chapter

Search systems find material according to what has been described, digitised and indexed. They may be excellent at locating a record that contains a known name or phrase, but less effective at finding a source whose importance is not obvious from its catalogue description. A historian may discover a relevant detail while reading a document for another purpose. This kind of serendipitous finding is difficult to reduce to a search query.

AI can widen search by suggesting synonyms, related people, place-name variants or concepts that do not appear in the user’s initial wording. It can also compare records that use different terminology. But researchers should not assume that the retrieved set is exhaustive. They should inspect how the archive was indexed, which collections were excluded and whether spelling, language or historical terminology affects the results.

A good research report describes the search method and its limits. It may explain that a particular collection was not digitised, that a language could not be reliably searched or that a record series was incomplete. These statements do not weaken the research. They help readers understand the boundaries of the evidence.

Contents · Previous chapter · Next chapter

Every historical claim should have a traceable relationship to evidence. Provenance records where a source came from, who created it, when it was produced, how it entered a collection and whether it has been altered or reinterpreted. Context explains the circumstances in which the source was created and the limitations of using it.

AI-generated summaries should preserve links to original records and distinguish quotations from paraphrases. A system should not create a quotation that does not appear in the source or silently modernise a phrase in a way that changes meaning. A transcript produced from handwriting recognition should be marked as a transcription and reviewed where accuracy matters.

Researchers should retain a research log that records search terms, collections consulted, dates accessed, exclusions and uncertainties. This is particularly important when AI helps select relevant documents. A search result is not a complete account of the collection, and another researcher should be able to understand how the evidence was assembled.

Contents · Previous chapter · Next chapter

Source criticism begins by asking who created a record, when it was created, what purpose it served and what audience it addressed. It also asks what the source could reasonably know. A diary may offer direct insight into one person’s experience but not a complete view of a city. An official report may record a policy accurately while presenting the institution’s preferred explanation. A photograph may be genuine while its caption is wrong.

AI can help a learner generate these questions, compare descriptions and identify claims that need verification. It should not decide that a source is unbiased or representative simply because it appears official or is frequently cited. Students should inspect the actual document, examine its context and compare it with independent material. A model’s confidence is not a substitute for historical method.

A useful discipline is to separate three layers in every research note: what the source explicitly states; what the researcher infers from it; and what remains uncertain. This prevents an interpretation from being repeated as though it were a direct quotation or an established fact. It also makes the final argument easier for another reader to evaluate.

Contents · Previous chapter · Next chapter

A historical question may require evidence from several institutions. A study of a port might involve maps, shipping records, customs documents, photographs, oral histories and trade statistics. A study of housing might involve planning records, building drawings, newspapers, personal accounts and demographic data. AI can help connect these materials by identifying names, places, dates and concepts that appear across separate collections.

These connections can reveal new questions. A recurring surname may link an official record to a local newspaper. A map label may appear in several decades of planning documents. A sequence of photographs may suggest a change in street use. A pattern in shipping records may invite a closer look at economic or political context.

Correlation is not proof of a historical relationship. Shared names may refer to different people. Similar places may have different histories. A missing document may be a result of cataloguing rather than a meaningful absence. AI can propose a connection; historians must verify identity, chronology and context before using it in an argument.

Contents · Previous chapter · Next chapter

AI can help historians compare records across languages, institutions and geographical regions. A study of migration may connect passenger lists, shipping records, newspapers, family accounts and policy documents from several countries. A history of trade may involve port records, commodity data, maps and the accounts of workers. These connections can reveal patterns that are difficult to see when a researcher works within one collection.

Comparative work needs careful control of definitions. The same word may have different legal or social meanings in different places. Dates may be recorded in different calendars. A translated term may hide a distinction that matters to the interpretation. AI can suggest translations and related sources, but a historian must verify important terms and consult specialists when needed.

The aim is not to make every historical account global by default. It is to recognise when a question has been framed too narrowly and when evidence from another place can challenge an assumption. A local history may become richer when it is placed in a wider network of exchange, movement or influence.

Contents · Previous chapter · Next chapter

Oral history can preserve experience that formal documents omit. Interviews may record how people remember work, migration, family life, conflict, neighbourhood change or cultural practice. AI can assist with transcription, translation, indexing and search, making recordings easier to use in research and education.

An oral-history project still depends on consent, careful interviewing, accurate attribution and respect for the person speaking. A transcript may misrecognise a name, a dialect expression or a pause that changes the tone. A summary can remove the speaker’s uncertainty or turn a personal recollection into a general claim about an entire community.

Researchers should retain the original recording where permission allows, identify the transcript as machine-assisted when appropriate, and provide a review route. Interviewees should understand how their words may be used, whether they can restrict access and how long material will be retained. AI should extend the reach of oral history without weakening the relationship of trust that makes it possible.

Contents · Previous chapter · Next chapter

Large national archives often preserve major events and official institutions. Local history may depend on smaller records: shop advertisements, school magazines, family photographs, transport maps, community newsletters, building plans and oral accounts. These sources can reveal how broad changes were experienced in daily life. They may also be fragile, privately held or difficult to search.

AI can assist with transcription and description, but local projects should be community-led and consent-based. A family photograph may contain people who are still alive. A letter may reveal personal information. A neighbourhood recording may include sensitive stories. Digitisation should not automatically mean unrestricted publication.

A well-designed local-history project records who supplied the material, what permissions were given, how descriptions were checked and how people can request a correction. It should preserve the original file and explain any restoration or generated enhancement. Small records can add important texture to history when they are handled with care.

Contents · Previous chapter · Next chapter

Students still need chronology, concepts, events and historical vocabulary. Without these foundations, they cannot understand what a source is referring to or why a sequence matters. But history education can go further than remembering a finished narrative. It can teach students how narratives are constructed, which sources support them and why historians sometimes disagree.

An AI-supported assignment might provide students with two contrasting source extracts and ask them to identify what each source can establish. Students can ask AI to suggest context questions, then check those questions against the original material. They might compare a generated summary with the sources and mark details that are unsupported, missing or overstated. This turns AI into a means of sharpening historical judgement.

The task should assess the student’s reasoning rather than the sophistication of the generated prose. Students should explain why a source is useful, what it cannot show and how another source changes the interpretation. This connects directly to AI and the Paradigm Shift in Education, AI-Powered Assignments and From the Centre to the Edge.

Contents · Previous chapter · Next chapter

Invented classroom case: a teacher prepares three fictional records about a bridge in a fictional town. A newspaper announcement dated 3 June 1950 says that the opening is planned for 10 June. A dated photograph shows an opening ceremony on 12 June. A maintenance ledger says traffic began using the crossing on 15 June. None of these materials is presented as evidence about a real place.

The class first writes an answer without AI. Some learners may select the earliest date; others may choose the photograph. The teacher asks what “opened” means in the question. It could refer to the announced schedule, the ceremonial event or the start of public use. The records may be describing different stages rather than directly contradicting one another. Defining the claim changes what evidence is relevant.

Students then ask AI to propose a timeline using only the supplied records. They check each entry against the source rather than accepting the generated chronology as a fourth independent witness. If the system writes that the bridge opened on 10 June, students identify the conversion of an announced intention into a completed event. If it selects 12 June, they ask whether a ceremony establishes that the bridge was open to traffic.

A useful research note separates the evidence from the interpretation. The newspaper supports what was announced on 3 June. The photograph, provided its date and caption are reliable, supports a ceremony on 12 June. The ledger supports what its keeper recorded about traffic on 15 June. The strongest answer depends on the question, the authenticity and purposes of the records, and whether other evidence challenges them.

Next, the teacher changes one condition: the photograph’s date was supplied by a later donor rather than written on the original image. Students must revise the confidence they place in that date. The image remains useful evidence, but its dating now has a different evidential basis. This is a productive revision, not a failure to defend the first answer. The model can suggest corroborating records; learners decide which proposed records would actually distinguish competing explanations.

The written outcome is a short argument with three parts: what is established by the supplied records, what is inferred, and what remains unresolved. Students may reasonably conclude that the bridge was in use by 15 June while the exact beginning of use needs further confirmation. They must not invent a weather delay, a political dispute or a construction accident to make the dates fit. A coherent account can preserve a gap.

An oral follow-up asks learners to handle a different event, such as an invented school inauguration with a founding date, a first teaching day and a later official ceremony. The teacher assesses whether the student can distinguish these dates without repeating AI’s wording. Later, the same reasoning can be tested with a new kind of source. Immediate explanation, later recall and transfer to unfamiliar material are separate checks; one polished history paragraph does not demonstrate all three.

For older students, the exercise can include a search log. Learners record which catalogue terms they tried, which documents were available, and whether two apparent sources copied the same earlier announcement. This introduces an important challenge: agreement between two pages may reflect common dependence rather than independent corroboration. AI can widen the search, but students should trace the chain before counting supporting sources.

The project connects history to news, museums and public institutions. A report can change as evidence arrives; a museum label can compress a complicated chronology; a public record can use a technical meaning that differs from everyday language. Learning to recognise these distinctions gives students a transferable intellectual tool. The benefit must be judged by the reasoning they can demonstrate, not by the number of archival-looking references a system generates.

Contents · Previous chapter · Next chapter

Begin with a bounded question and a defined period, place or community. Identify the relevant repositories and search their catalogues. Use AI to suggest alternate spellings, related concepts, languages or record types, but record the searches used. Retrieve the original sources rather than relying only on generated summaries.

Create a source table with creator, date, genre, intended audience, provenance, key claim and limitations. Compare sources that agree and those that differ. If AI proposes a connection, verify names, dates and context. Ask what evidence would disprove the interpretation. Preserve an account of missing sources and unresolved questions.

Finally, write an argument in which claims are tied to evidence. Separate what a source states from what the researcher infers. Explain uncertainty and identify the next record or interview that might improve the answer. The result should help another person inspect the reasoning rather than simply trust the finished paragraph.

Contents · Previous chapter · Next chapter

Historical events have at least two timelines: the timeline of what happened and the timeline of what people knew about it. A government may make a decision on one date, publish an explanation later and revise the explanation after new evidence emerges. A witness may record an event immediately, while a later archive reveals additional context. A historian must keep these timelines separate.

AI can help compare versions of documents, organise dated reports and identify when a phrase or explanation first appears. This can reveal how public understanding changed. But it must not treat the latest account as though everyone knew it at the time. The beliefs, uncertainties and errors present in earlier records are themselves historical evidence.

A useful student project is to create a timeline with three layers: the event, the earliest available record and later reinterpretations. Students can mark what was known, what was alleged and what became established only after additional evidence. This develops a more careful understanding of how knowledge changes.

Contents · Previous chapter · Next chapter

History is present in public spaces: monuments, memorials, street names, museums, anniversaries, school curricula and family stories. These forms of memory influence how people understand identity and belonging. They may also become contested when different communities remember an event differently or when new evidence challenges an established account.

AI can help compare public records, locate earlier versions of a speech or map the history of a place. It can also generate plausible but false images of historical events. A reconstruction can be valuable when it is clearly labelled and based on evidence. It becomes misleading when a viewer assumes it is an authentic photograph or an exact record.

Public history should distinguish evidence, memory, interpretation and artistic reconstruction. The aim is not to eliminate disagreement by producing one automated answer. It is to help people understand why accounts differ and what evidence can responsibly support.

Contents · Previous chapter · Next chapter

News reporting captures events as they unfold, while history develops through later research and interpretation. AI can connect these time horizons by helping researchers search news archives, compare early reports with later findings and examine how a narrative changed over time. It can also help journalists locate historical context for current events.

The risk is retrospective certainty. A later summary may make an uncertain event appear inevitable or treat an early allegation as established fact. Historical interpretation should preserve what people knew at the time, what was disputed and when new evidence became available. The timeline of knowledge is itself part of the history.

The companion article AI and the Paradigm Shift in News examines verification, provenance and information integrity. The history series adds a longer time perspective: how public claims become records, how records become evidence and how evidence is interpreted by later generations.

Contents · Previous chapter · Next chapter

Conflict destroys records, displaces people and can place cultural heritage at risk. Archives may be damaged, collections scattered and personal documents lost. AI may support the identification of damaged objects, transcription of documents, translation of testimonies and reconstruction of catalogue information from surviving sources.

These tasks are sensitive. Images and records may reveal the location of vulnerable people or culturally important objects. A public digital map may help researchers but also create security risks. Any use of AI in conflict-related heritage should be assessed for potential harm, access restrictions and the wishes of affected communities.

Historical records can support accountability and reconstruction, but a machine-generated narrative must never replace the evidence itself. The companion AI and the Paradigm Shift in War explores how conflict analysis must include civilian protection, human accountability and long-term recovery.

Contents · Previous chapter · Next chapter

Historians using AI should disclose material assistance in accordance with the standards of their institution, publisher or research project. They should verify references, inspect original documents and retain responsibility for the interpretation. AI can help generate a research plan, compare source descriptions or identify possible contradictions, but it cannot be treated as a source unless the system itself is the object of study.

The publisher’s summary of Yaniv Fox’s 2026 Cambridge University Press work on generative AI in historical practice describes practical uses such as semantic search, network visualisation, multimodal source analysis and code-assisted workflows, while warning about cognitive offloading and model bias. See Using Generative AI in Historical Practice.

The point is not to protect historical research from new tools. It is to use them in ways that preserve the standards that make historical knowledge trustworthy. Faster search is valuable when it creates time for deeper source criticism, not when it encourages researchers to skip it.

Contents · Previous chapter · Next chapter

Historical knowledge changes when new sources are found, old sources are reinterpreted or the questions asked by society change. Digital publishing and AI-assisted search can make corrections and new connections easier to share. But public history must preserve version history and explain why an interpretation changed. An old account should not simply disappear if readers need to understand how the conclusion developed.

Schools, museums, archives and publishers can provide a correction route for factual errors, misidentified photographs or incomplete descriptions. Contributions should be assessed according to evidence and expertise. An open correction process does not mean that every claim receives equal weight; it means that the reasoning behind a change can be inspected.

The future of history is not a single AI-generated explanation of the past. It is a more connected research environment in which people can find evidence, test interpretations, recognise missing voices and improve public understanding over time. That is how historical memory remains alive without becoming detached from proof.

Contents · Previous chapter · Next chapter

A historical investigation begins with a question, but the question may change as sources are discovered. A name may refer to two different people. A date may be uncertain. A document may reveal that the original question assumed something that was not true. AI can help researchers notice possible connections, but the historian must be prepared to revise the scope rather than force the evidence to fit the first framing.

This is one of the most important habits for students to learn. They should be rewarded for explaining why a question changed after new evidence appeared. Historical understanding develops through careful revision, not through defending the first answer at all costs.

Researchers should also record the boundaries of their conclusions. A study of one city, institution or social group should not quietly become a claim about an entire country or period. AI may help generalise a pattern in language, but the historian must decide whether the evidence supports that wider claim. The discipline of stating scope is part of historical honesty.

Contents · Previous chapter · Next chapter

AI will make it easier to obtain a quick account of many past events. That may help people begin learning, but it should not become the final form of historical education. Students and researchers should be encouraged to ask what is missing, why sources disagree, how a record was produced and what new evidence would change the interpretation.

The centre of history provides continuity and a foundation. The edge provides opportunities to revisit established accounts, include overlooked experience and discover relationships across collections. A responsible use of AI makes the edge more accessible without pretending that every question can be answered by a generated paragraph.

The paradigm shift is not that the machine remembers history for humanity. It is that people gain new ways to investigate the past together while preserving the source criticism, context and accountability that make historical knowledge trustworthy.

Contents · Previous chapter · Next chapter

AI can make the past easier to search, but historical understanding still depends on the disciplined interpretation of evidence. The archive is not the past itself. The catalogue is not the collection in its entirety. A generated summary is not a historical argument. These distinctions should remain visible as access improves.

The edge of history is where overlooked records, missing voices and unexpected relationships invite new investigation. It is also where historians must be most careful not to turn absence into certainty or machine-generated fluency into proof. The centre gives continuity; the edge allows inherited narratives to be questioned and improved.

The paradigm shift is from simply recording and repeating accounts of the past towards helping more people discover how historical knowledge is made. AI can widen access to the evidence and help identify new questions. Human scholarship, community memory and transparent methods determine whether those questions lead to responsible understanding.

Contents · Previous chapter · Continue the series