AI at Polmanarkivet

Our principles, our limits, and the questions we're still sitting with. A framework for intentional AI use in archives, genealogy, and cultural heritage.

Statement of AI use

Polmanarkivet uses AI as a tool, always under human direction. We do so because this work cannot be done at scale without it. Professional translation of primary sources across six centuries of European archives is beyond what a self-funded archive can sustain. We believe accessible cultural heritage matters, and AI is part of how we make it possible.

AI use comes with real risks and we take them seriously. Our use is governed by four principles: human responsibility, accuracy, transparency, and stewardship.

This document sets out where we use AI, where we don't, and the reasoning behind those choices. It also offers reflective questions for cultural heritage practitioners navigating AI use without institutional support.

About Polmanarkivet

Polmanarkivet researches, translates, collects, preserves, and tells the stories of the Polman, Påhlman, and von Pohlmann families — documented across six centuries. Established in 2023, it is self-funded and operates without institutional governance or infrastructure. The work would typically require a team of specialists:

Research — original historical and genealogical research across archives, primary sources, and secondary literature
Translation and transcription — primary sources in Early Modern Swedish, German, and Latin that would otherwise be inaccessible
Collection and curation — physical acquisitions and digital objects spanning six centuries
Cataloguing and preservation — archival description, metadata, and collection management
Storytelling — narrative articles, biographical lexicons, and published research that makes the history accessible

Much of this history exists only in primary sources — letters, manuscripts, and court records in Early Modern Swedish, German, and Latin, in handwriting that requires specialist training to read. Most have never been translated. They belong to descendants, researchers, and communities who cannot access them. Professional translation at the scale this mission requires is beyond what a self-funded archive can sustain. That is a loss — and accessible cultural heritage is a democratic question. AI is part of how we address it.[[1]][[2]]

Polmanarkivet uses AI across its work: in research and synthesis, translation and transcription, cataloguing, and editorial work. This policy sets out how we use it and the principles that guide that use. It is written for our own readers and for specialist archivists, heritage professionals, genealogists, and independent researchers navigating AI use without institutional support or oversight.

But AI use comes with responsibility. Archives are entrusted with the curation of validated knowledge, evidence, and memory.[[3]] The ethical risks of AI in archival and heritage work are specific: hallucination, provenance loss, bias amplified by training data, and erosion of the standards that make archival work trustworthy. This policy exists because those risks are real. If you use AI, engaging with them honestly is the only responsible way to do it.

Where we use AI, and where we don't

The table below sets out where we use AI and where we don't — and what keeps each use accountable. For the reasoning behind these commitments, see Our principles.

Research and evidence

Use	Do we use AI?	How we keep it accountable
Finding secondary sources	Mostly no	Only for general Swedish historical context, never for the specific family history, which falls outside what models reliably know. AI augments our own sourcing. It never replaces it. Every source it surfaces is checked to confirm it exists and then read directly.
Summarising and synthesising secondary literature; identifying themes and gaps	Sometimes	AI supplements our own research. Output is a starting point only. We return to the underlying sources, read them directly, and verify before anything informs published work. Sources suggested by AI are checked to confirm they exist before use.
Estimating dates from documented evidence	Sometimes	AI helps narrow plausible date ranges from assembled evidence — birth order, marriage records, known lifespans. All results are reviewed by a human and published as estimates, never as established fact.
Determining relationships, building family trees, or making genealogical connections	No	We do not use AI to establish relationships between people, confirm family connections, or build family trees. This is among the highest-risk uses in genealogical research and falls outside what AI can reliably do.
Artifact dating and material culture description	Sometimes	Used to identify style periods, date ranges, or specialist terminology for objects, heraldry, and visual material. This is an area where AI is particularly unreliable, and we treat its suggestions with corresponding caution. Results are presented as contextual information, never as established fact. Where a source describes something as undated or unknown, that designation is preserved.

Translation and transcription

Use	Do we use AI?	How we keep it accountable
Translating published secondary sources (Swedish, German, Russian, and others) into English	Sometimes	DeepL is our primary tool for this use. LLMs are used occasionally for comparison or where a translation needs further checking. Output is checked against the source and against other secondary literature on the same subject.[[4]] See: Accuracy: Care for the record
Transcribing and translating primary sources — letters, manuscripts, court books, and estate inventories in Early Modern Swedish, German, Latin, and historical script	Yes	Confidence is built through convergence: transcription is cross-checked across Transkribus and an LLM; translation is re-run across multiple models; every output is checked against what we know of the family history and its wider context, with particular attention to names, place names, and titles. When multiple tools converge, confidence increases. When they disagree, confidence decreases and we investigate further. Where a professional translation exists, we compare against it. We seek specialist consultation where resources allow. Where confidence is insufficient, we do not publish the translation directly — we summarise or paraphrase the content instead. Where confidence is lower but sufficient, we flag it in the article. The specific tool and model are named in the disclosure of any article that relies on these translations. We invite corrections from expert readers.[[5]] See: Accuracy: Care for the record
Textual reconstruction of missing or damaged sources	No	We do not use AI to speculate, predict, or fill in missing, faded, or physically damaged words in primary source documents. Gaps and unreadable sections are left as they are in the record.
Silent modernisation of historical text	No	We do not use AI to silently correct, modernise, or standardise the spelling, grammar, or syntax of historical transcriptions. The original form is preserved because it is part of the source's integrity.

Content creation and editorial

Use	Do we use AI?	How we keep it accountable
Generative drafting of article prose	Yes	A human directs the process throughout, including the prompts, and takes full responsibility for the result. These prompts direct AI to work only from our assembled sources and not to invent facts, fill gaps with outside knowledge, or add imagined detail. Every factual claim is verified against its source. This workflow applies to some articles, not all. Every article produced with AI assistance carries a disclosure statement.[[6]] See: Human responsibility: Care for judgment · Transparency: Care for the reader
Contributor-authored articles	No	Articles written and submitted by contributors are their own work. We do not apply AI to their writing at any stage. Their approach to AI in their own practice is theirs to determine. See: Stewardship: Care for what's entrusted
Generating or altering images presented as historical or archival evidence	No	Every image we publish as historical or genealogical content is real and sourced. We do not use AI to generate or alter images and present them as genuine. We do not alter the originals we hold or publish.
Direct communications — responses to members, contributors, and readers	No	Members, contributors, and readers receive responses written by a human.
Social media and newsletter content	Sometimes	AI drafts from material and discoveries we supply. A human edits, approves, and publishes.
Impersonating historical figures — AI "personas" or chatbots that speak as the people in our collection	No	We will not create AI that puts invented words into the mouths of real historical people. It fabricates speech that was never spoken and treats the dead as a novelty — both incompatible with the dignity and authenticity this archive exists to protect.
Altering historical images and documents — colourising images, enhancing or reconstructing archival materials	No	We do not alter historical images or documents held in or published as part of the archive — including photographs, manuscripts, and primary source scans. The condition of an original is itself part of the record. Digital adjustments for design or layout purposes, where an image is not being presented as archival evidence, are handled case by case.

Archive infrastructure

Use	Do we use AI?	How we keep it accountable
Cataloguing and metadata — extracting and formatting structured records from sources: suggesting descriptive fields, links between records, keywords, and classifications	Yes	AI operates within fixed schemas and controlled vocabularies — it extracts and formats, but does not interpret or infer. Anything ambiguous is flagged for human review rather than guessed at. Every suggestion is reviewed and confirmed before entering the catalogue. Records are reviewed again before export to public display.
Object description — narrative description of items we hold or where existing descriptions are minimal	Sometimes	Description is grounded in historical context and specialist sources. AI assists but does not interpret — a human writes and approves the final description.
Automated cross-linking and content discovery	No	Relationships and connections between catalogue entries, family members, historical events, and objects are made by a human, not generated automatically.

Our principles

These principles draw on guidance across genealogy, archival science, museum practice, and heritage management — from organisations including CRAIGEN, the American Historical Association, the Association of Research Libraries, and the University of Virginia Library, among others listed at the end of this document. They also reflect our own practice and the feedback of professionals across these fields who engaged with earlier versions of this work. For more on our methodology, see Methodology.

Human responsibility: care for judgment

AI provides augmented intelligence; it does not replace human intelligence.

Christopher Dede asks whether AI can be like the owl on Athena's shoulder — augmenting human wisdom without supplanting it. The owl sits on your shoulder, not the other way around.[[7]] At Polmanarkivet, that is the operating principle. AI does not author, curate, or decide. It responds to human direction. Every output is reviewed by the person responsible for it. Guidance across these sectors names the same standard.[[8]][[9]]

You need expertise to judge AI output, and AI cannot give you that expertise.

Responsible use means having the prior knowledge to evaluate what AI produces. AI can generate work that looks authoritative, but assessing whether it is accurate requires knowledge the tool cannot provide.[[10]] The expertise has to come first. This also means using AI where it genuinely serves the work — not by default and not for every task. Adopting AI is not an obligation, and knowing whether a specific tool serves a specific purpose is itself a form of expertise.[[11]]

As AI makes answers more abundant, human judgment becomes more valuable, not less.

AI will not replace the people who do this work. The ability to weigh and verify AI output — to distinguish what is correct from what only looks correct — is a skill that becomes more important as AI becomes more capable.[[12]] That judgment cannot be automated.

How we apply this: Everything that carries the Polmanarkivet name has been read and can be defended by the person who published it. AI assists across the work, but judgment about what is accurate, relevant, and appropriate to publish stays human throughout.

Accuracy: care for the record

In archival work, an error doesn't stay private — it enters the record, gets cited, and spreads.

Accuracy, context, and provenance are the standards against which our work is measured, AI-assisted or otherwise. Errors damage not only scholarly credibility but the accuracy of the public record. All final work must be "created, interpreted, and validated by experts to ensure accuracy, context, and provenance."[[13]] They are inseparable. A correctly transcribed fact means little without the context that explains it, and cannot be trusted without the provenance that traces it.[[14]][[15]] In genealogy specifically, errors multiply. A mistake published here gets copied into family trees, and those copies are impossible for us to correct.

AI hallucinates, substitutes, and fills gaps with false certainty — and the errors are hardest to catch where you are most reliant on AI.

AI introduces specific risks to accuracy. Large language models (LLMs) hallucinate. These risks are most pronounced in the material we work with. The models are trained primarily on contemporary, English-language text and handle older forms of Swedish and German poorly.[[16]] Much of what Polmanarkivet holds has never been digitised or published, which means AI has no prior knowledge to draw on when it encounters these sources.

The subtler risk is harder to catch. AI tends to substitute a familiar word for an unfamiliar one, and the substitution is convincing. That word is often the most important one: a place name, a family name, or a term whose meaning depends on its historical context.[[17]] Where the record is genuinely uncertain, AI produces a confident answer, turning a real gap into false certainty. Both risks share the same problem: you can only catch an error if you already know enough to recognise it. The less familiar the material, the more you lean on AI, and the less able you are to question what it gives you.

AI reproduces the gaps and silences already present in the historical record.

The bias in AI output reflects the bias already present in the historical record. Someone decided whose histories were worth keeping, and someone else decided what was described or digitised. AI trained on that record reproduces its gaps. Checking a claim against a source cannot correct for this, because the source contains the same absences.[[18]][[19]] The problem extends to description. Collections from marginalised communities have often been under-described or never described, making them invisible to researchers and AI alike. Resources go to collections already in demand, which means the least-known collections stay unknown.[[20]]

How we apply this: Every published claim traces to a source we have consulted directly. Footnotes and citations are how we show that. Our research is grounded in literature review, primary sources, and specialist consultation. Where AI contributed to a piece of analysis or text, we document its role. Where errors are found, we correct them promptly, note the correction visibly, and document what changed.

Transparency: care for the reader

Transparency is owed to readers — and a polarised climate around AI makes it more necessary, not less.

In archival and heritage work, how work is made is as important as what it says. The current climate around AI creates a social cost to disclosure — many practitioners who use AI thoughtfully say nothing about it. As Emilie Hardman has observed, the hostility to AI in archival fields reflects "a pattern that archivists and librarians have experienced repeatedly: technology introduced as a substitute for labor, tools procured without adequate governance, and efficiency narratives used to justify disinvestment."[[21]]

These concerns are real and widely held. In the AI for Access survey (2026), 59% of American archival workers described themselves as highly concerned about the ethical implications of AI, and when asked to say why, the concern they named most often was environmental and resource cost.[[22]] Among the public, a 2026 survey of over 2,000 US adults found that 81% want to know when cultural heritage institutions use AI, either every time or, at minimum, for interpretive content.[[23]] But silence does not resolve the questions AI use raises — it makes them harder to address.

What matters in disclosure is not which tool was used, but how it was used and who stands behind the work.

Our approach to disclosure follows the Elsevier declaration, which requires that authors describe how AI was used and confirm that they "reviewed and edited the content as needed and take full responsibility."[[24]] We name the specific tool and model where AI was used to translate or transcribe a source. For research synthesis and drafting, we name the kind of tool rather than the product. We do not disclose routine assistance such as idea generation or formatting citations into our publishing system.

How we apply this: Articles that used AI contain a disclosure statement. It describes how AI was used, names the specific tool and model where translation or transcription is involved, and confirms that the author reviewed, verified, and takes full responsibility for the content. That statement links to this policy for readers who want a full account of how we use AI. Where confidence in a translation is lower, we flag it in the article. Where we cannot verify a translation sufficiently, we do not publish it directly — we summarise or paraphrase the content instead. Readers who want to examine the evidence for themselves can do so through the citations, source references, disclosure statements, and documented corrections that accompany every piece of published work.

Stewardship: care for what's entrusted

Polmanarkivet holds its content, its contributors' work, and the material entrusted to it as a steward. Stewardship means taking responsibility for the care, integrity, and appropriate use of material that does not wholly belong to us — and exercising what authority we do hold in service of access and the wishes of those who entrusted it.

Archives exist to facilitate access to knowledge, not to restrict it.

Dave Hansen warns that frameworks built around institutional control risk remaking archives in the image of rights-holders rather than facilitators of knowledge.[[25]] That is not what authority means at Polmanarkivet. As Europeana's data sovereignty principle holds, institutions should keep authority over how their content is used so that use stays faithful to the values under which material was entrusted to them.[[26]]

The people who contribute to and entrust material to this archive have a right to know how it is used — including whether AI is part of that.

Consent at Polmanarkivet covers two distinct groups. Contributors and writers — people who create and publish work with us — may hold a range of views on AI. Some do not want their work associated with it or processed by it. Their reputation is tied to what they publish here, and we respect their preferences about how their work is handled. Donors and those who entrust material to the archive have a different but equally important relationship with us. They have a right to know how their material is used, and we do not process it through AI systems without their knowledge.

When archival material loses its chain of context, it loses its integrity as evidence.

Every piece of historical material has a provenance — where it came from, who held it, how it arrived, what it means in context. That chain is what makes it trustworthy. When material is absorbed into an AI training model, that chain breaks. The information persists, influencing outputs, but its origin is gone — what Leo Lo calls "memory without origin."[[27]] It cannot be pointed to, verified, corrected, or removed.[[28]]

The tools we use were built on data scraped without consent. We cannot undo that, but we can acknowledge it and let it shape our choices.

This is an unresolved tension we sit with honestly. We cannot fix what has already happened through platform selection alone. What we can do is opt out of AI training on our own content, choose platforms that support that, and remain attentive as policies change.

AI has an environmental cost. We do not pretend otherwise.

Training AI models requires significant energy and water. This is an industry-wide problem that individual practice cannot resolve. We use AI deliberately rather than habitually, and match the model to the task rather than defaulting to the largest available. We are exploring local processing as an alternative to cloud-based tools where the work allows. We support communities opposing data centre expansion on environmental grounds. These are small contributions to a large problem.

How we apply this: We opt out of AI training on our published content, block training crawlers, and use a managed robots.txt to signal this. Contributors can request page-level exclusions, considered case by case. Contributor work is not processed through AI without their knowledge. Material donated or entrusted to the archive is handled the same way. Working material processed through cloud-based tools is subject to those platforms' own data terms. We use platforms that do not train on input data where possible. We are exploring local processing as an alternative to cloud-based tools where the work allows — for reasons of both sensitivity and environmental cost. AI chats are retained as a working record. For published work, the record that matters to future researchers is the disclosure statement, the citations, and any flagged uncertainty in the text.

For practitioners

The guidelines for AI use that informed this policy were built for institutions. Most independent practitioners — archivists, genealogists, and researchers working alone — don't have access to governance committees, ethics boards, or formal AI policies to lean on. These questions are offered as a thinking tool for anyone navigating AI use on their own.

Deciding whether to use AI

Who benefits from this use?[[29]] If the answer is efficiency or workload, pause. If it's your audience, your research community, or the people whose history you're documenting, that's a different kind of answer.
What are the costs of not using AI? Consider what the work requires without AI — not just time, but whether certain work happens at all.
What are the costs of using AI? Name the risks honestly: hallucination, provenance loss, environmental impact, and the erosion of your own thinking.
Does this serve the purpose of the archive and the people that purpose exists for?
Is this use of AI enhancing your own thinking, or replacing it?[[30]]

These questions are not a test with a correct answer. Practitioners work somewhere on a spectrum: from those who choose not to use AI at all to those who use it extensively and thoughtfully. Every position on that spectrum is legitimate if it is arrived at honestly.

Putting the principles into practice

If you've decided to engage with AI in your work, the following questions help you apply each principle to your own practice.

Human responsibility

Can you stand behind this output? Everything published under your name should be something you can defend.
If something you produced with AI contained an error, could you explain exactly where your judgment was present? Could you explain where it wasn't?
Is AI sharpening your judgment in this work, or gradually replacing the need to exercise it?[[31]][[32]]
Are you using AI to do the same work faster, or to do better work?[[33]]

Accuracy

Can you trace every claim back to a source you have actually consulted? Are any claims relying on AI output you haven't fully verified? A claim without a traceable source cannot be trusted or built upon.
Where in your work is AI most likely to produce something that looks right but isn't? Do you have a way of catching that? The risk isn't uniform — knowing where it's highest in your work is where to start.
How would you recognise a hallucination that looked convincing? The obvious errors are not the problem.
Whose histories and perspectives are well represented in what AI knows — and whose aren't? How does that shape what it produces for your specific work?

Transparency

When you disclose that AI was involved, what exactly are you taking responsibility for? Are you prepared to stand behind all of it? Disclosure is not just a label — it's a commitment.
Is your disclosure specific enough that a reader could understand how AI shaped what they're reading? Or is it just a flag? There's a difference between "AI was used" and "AI drafted the initial synthesis that this argument is built on."
If you're saying less about your AI use than you could, what's driving that? Is it a reason you'd be comfortable defending publicly?[[34]] The social cost of disclosure is real. So is the cost of silence.

Stewardship

The people who have entrusted material to you — do they know how AI is part of how you work with it? Have you made it easy for them to find out? Knowing and being able to find out are different things.
Can you account for how AI has shaped your work? Not just what it produced, but how it influenced your thinking, your framing, your conclusions? What AI produced is traceable, but what it did to your thinking is harder to see.
The tools you use were built on data scraped without consent. Does that sit comfortably with you? If not, how does that discomfort shape your choices?
Are your AI decisions facilitating or restricting access to knowledge?[[35]] Ask whether each decision genuinely serves the people the archive exists for, rather than reflecting unease about AI itself.

A note on this policy

This policy is a living document, reviewed annually. These technologies are evolving rapidly, and we will update it as our practice develops. This policy reflects a specific context: an independent, digital archive without institutional infrastructure, operating across genealogy, archival science, museum practice, and heritage management. Ethical frameworks and guidance specific to these sectors remain underdeveloped — including in a Swedish context, as far as we have found. It is offered as a contribution to that conversation, and we welcome you to adapt or build on this work.

Approved by Jake Påhlman Peterson, Director, Polmanarkivet. Last updated: 6 June, 2026.

This policy was produced with AI assistance. Research, synthesis, and drafting were carried out with an LLM, directed and rewritten throughout by the author, who reviewed and takes full responsibility for the published content.

Glossary

Definitions to technical terms used in this policy

Artificial intelligence (AI)
A field of computer science focused on building systems that can perform tasks normally requiring human intelligence: reasoning, learning, language understanding, and decision-making. In this policy, AI refers to a range of tools, from general-purpose language models to more specialised transcription and translation tools.

Augmented intelligence
An approach to AI that emphasises its role in extending and supporting human capability rather than replacing it. Where artificial intelligence implies substitution, augmented intelligence implies partnership — the human remains in direction.

Bias (in AI)
The tendency of AI systems to reflect the gaps, imbalances, and silences already present in the data they were trained on. In historical and archival work, this means AI reproduces whose histories were recorded, preserved, and digitised — and whose were not.

Data sovereignty
The principle that an institution retains authority over how its content and data are used by others, including AI systems. For archives and heritage institutions, data sovereignty is a means of ensuring that use of their material stays faithful to the values under which it was entrusted to them.

Hallucination
When an AI model produces output that sounds plausible but is factually incorrect or entirely fabricated, including invented sources, false quotations, and confident answers where no reliable answer exists.

Large language model (LLM)
A type of AI trained on large quantities of text to understand and generate human language. LLMs identify statistical patterns in their training data rather than reasoning from knowledge — which is why they can produce fluent, authoritative-sounding text that is nonetheless wrong.

Provenance
The documented origin and chain of custody of a source or piece of information: where it came from, who held it, and how it arrived. In archival work, provenance is what makes material trustworthy as evidence. A claim or record without traceable provenance cannot be verified or built upon.

Training data
The material used to build an AI model — the texts, images, and other content the model learns from. The scope, composition, and origins of training data directly shape what an AI model knows, what it gets wrong, and whose histories and perspectives it reflects.

Resources

Frameworks, policies, and guides

Frameworks, policies, and guidelines specifically addressing AI use in heritage, archival, museum, genealogical, or research contexts.

Alliance of Independent Authors (ALLi) — AI for Authors: Ethical and Practical Guidelines. selfpublishingadvice.org
American Alliance of Museums (AAM) — Artificial Intelligence in Museums: Discussing Ethics and Protocols (2025). aam-us.org
American Historical Association (AHA) — Guiding Principles for Artificial Intelligence in History Education (2025). historians.org
Archives and Records Association UK & Ireland (ARA) — AI Preparedness Guidelines for Archivists (FLAME Project, Colavizza & Jaillant, 2026). archives.org.uk
Association of College and Research Libraries (ACRL) — AI Competencies for Academic Library Workers (2025). ala.org/acrl
Association of Research Libraries (ARL) — Research Libraries Guiding Principles for Artificial Intelligence (2024). arl.org
Coalition for Responsible AI in Genealogy (CRAIGEN) — Guiding Principles for Responsible AI in Genealogy. craigen.org
Council of Europe / CDCPP — Guidelines on AI and Cultural Policy (2025). coe.int
Elsevier / Digital Applications in Archaeology and Cultural Heritage — Declaration of Generative AI Use Policy. sciencedirect.com
Europeana Foundation — AI4Culture and 2026 Business Plan. europeana.eu
History Associates Incorporated (HAI) — Statement on the Responsible Use of AI. historyassociates.com
InterPARES Trust AI — Multi-national research programme on AI and archival science, producing studies on trustworthy records, ethical codes, metadata, and AI-assisted access (2021–2026). interparestrustai.org
Museums and AI Network — AI: A Museum Planning Toolkit (Murphy & Villaespesa, 2020). themuseumsai.network
National Archives and Records Administration (NARA) — AI Strategic Framework and Use Case Inventory (2024). archives.gov
Riksarkivet / AIRA — AI lab, HTRflow, and open source tools documentation (2025). riksarkivet.se
Smithsonian Institution — Statement on the Use of Artificial Intelligence. si.edu
Sweden — National AI Strategy (February 2026). government.se
University of Virginia Library — Archival AI Protocol (Lo, L., 2026). library.virginia.edu

AI tools for historical research

Tools used in archival, genealogical, and historical research contexts.

Transcription and handwriting recognition

Archive Studio and Endeavour — transcription tools for historical documents, created by Mark Humphries. generativehistory.substack.com
Gemini — Google's LLM; particularly noted for handwriting transcription of historical documents. gemini.google.com
Leo — AI transcription tool for historical research. tryleo.ai
Transkribus — specialist HTR platform for historical handwritten documents. transkribus.com

Specialist historical research tools

Genealogy Eyes — custom GPT for dating historical images through deep analysis of clothing, objects, and period details. chatgpt.com/g/g-gmIAn5mh6-genealogy-eyes
The Historian's Friend — custom GPT built by Benjamin Breen for historical text transcription, translation, and analysis; cites relevant academic sources. chatgpt.com/g/g-u5cMhl7RY-the-historian-s-friend

Provenance, authenticity, and content verification

Archival Producers Alliance (APA) — GenAI Initiative: guidelines for ethical AI use in documentary and audiovisual archival work. archivalproducers.org
C2PA (Coalition for Content Provenance and Authenticity) — Open technical standard for establishing the origins and edit history of digital content. c2pa.org
Project Origin — Infrastructure for tracking provenance and authenticity of multimedia content. originproject.info
Trust in Archives Initiative (TAI) — AI Toolkit for cultural heritage institutions managing risks around AI-generated content. trustarchives.org
WITNESS — Advocacy and frameworks for AI transparency protecting vulnerable communities and preventing deepfakes from undermining authentic evidence. witness.org

Methodology

How this policy was developed

This policy was developed through a four-stage process.

The first stage was a review of the relevant literature: sector-specific guidance on AI use, ethical frameworks, and peer-reviewed research on AI in historical research, transcription, translation, and collections management.

The second stage was frameworks mapping. Policies and guidelines were reviewed from across genealogy, archival science, museum practice, and heritage management — including from the American Historical Association, the Association of Research Libraries, CRAIGEN, the Smithsonian Institution, and the University of Virginia Library, among others. The goal was to identify where principles aligned across sectors. The four principles in this policy reflect that alignment.

The third stage was accountability development: honest reflection on Polmanarkivet's own practice, which produced the table of uses and the specific safeguards attached to each.

The fourth stage was community consultation. Earlier versions of this policy were shared across professional communities in genealogy, archival science, museum practice, and digital humanities. Feedback from working archivists and museum professionals shaped the final document in specific ways, including the genealogy propagation risk in Accuracy, the expanded transparency commitments, and the paradata commitment in Stewardship.

[[1]]: Riksantikvarieämbetet. Nationell strategi för digitalt kulturarv. 2024. raa.se. The strategy identifies access as the primary purpose of digital cultural heritage work and names the translation of analogue material as a key enabler of that access.

[[2]]: UNESCO. Recommendation on the Ethics of Artificial Intelligence. 2021, updated 2023. unesco.org. The Recommendation specifically promotes AI use in the preservation, enrichment, and accessibility of cultural heritage.

[[3]]: Dempsey, Lorcan. "Generative AI and Libraries: Seven Contexts." November 2023. lorcandempsey.net

[[4]]: Yan, Jianhao, Pingchuan Yan, Yulong Chen, Jing Li, Xianchao Zhu, and Yue Zhang. "Benchmarking LLMs Against Human Translators: A Comprehensive Evaluation Across Languages, Domains, and Expertise Levels." IEEE Transactions on Big Data 12 (June 2026): 801–813. doi.org/10.1109/TBDATA.2025.3644594. Found LLMs perform at junior-to-mid professional translator level across language pairs.

[[5]]: Humphries, Mark, Lianne C. Leddy, Quinn Downton, Meredith Legace, John McConnell, Isabella Murray, and Elizabeth Spence. "Unlocking the Archives: Using Large Language Models to Transcribe Handwritten Historical Documents." Historical Methods: A Journal of Quantitative and Interdisciplinary History 58, no. 3, 2025. doi.org/10.1080/01615440.2025.2500309. In tests on 18th and 19th century handwritten documents, LLMs achieved near-human accuracy when outputs were cross-checked across models.

[[6]]: Lubar, Steven. "AI for Local History Organizations." History News, American Association for State and Local History, Technical Leaflet #308, vol. 79, no. 4, 2024. aaslh.org

[[7]]: Dede, Christopher, quoted in Mineo, Liz. "Is AI dulling our minds?" Harvard Gazette, November 13, 2025. news.harvard.edu

[[8]]: Harvard University AI Task Force. Statement on Generative AI. 2024. provost.harvard.edu

[[9]]: Association of Research Libraries. ARL Guiding Principles for Artificial Intelligence in Libraries and Archives.2025. arl.org

[[10]]: American Historical Association. Guiding Principles for Artificial Intelligence in History Education. 2025. historians.org. Source of what this policy terms the "expertise paradox": the expertise needed to evaluate AI output is the same expertise AI use risks eroding.

[[11]]: Association of College & Research Libraries. AI Competencies for Academic Library Workers. October 2025. ala.org/acrl

[[12]]: Mineo, Liz. "Is AI dulling our minds?" Harvard Gazette, November 13, 2025. news.harvard.edu

[[13]]: History Associates Incorporated. HAI Statement on the Responsible Use of AI in Historical, Archival, and Research Work. 2025. historyassociates.com

[[14]]: Huvila, Isto, et al. "Paradata as a Tool for Understanding and Communicating the Provenance of AI-Assisted Archival Work." Archivaria, 2024.

[[15]]: Riksantikvarieämbetet. Nationell strategi för digitalt kulturarv. 2024.

[[16]]: Sweden Government. Sweden's National Strategy for Artificial Intelligence. February 2026. government.se. The strategy explicitly notes that "Sweden also needs better access to Swedish language models" to improve quality and cultural adaptation — acknowledging the inadequacy of existing models for Swedish-language use.

[[17]]: Breen, Benjamin. "On the Dangers of AI-Assisted Research." Res Obscura, 2023. resobscura.substack.com

[[18]]: Dikow, Rebecca, et al. "Developing responsible AI practices at the Smithsonian Institution." Research Ideas and Outcomes, 2023. doi.org/10.3897/rio.9.e113334

[[19]]: Baker, Maher Asaad. "AI Ethics in Historical Research: A Framework for Bias Mitigation, Transparency, and Accountability." SSRN, October 2025. doi.org/10.2139/ssrn.5565499

[[20]]: Hardman, Emilie. "Bringing Hidden Histories to Light: An Archivist Reflects on AI, Archives, and the Future of Digital Stewardship." JSTOR, March 2025. about.jstor.org. Resources flow to collections already in demand; under-described collections from marginalised communities remain invisible to researchers and AI alike.

[[21]]: Hardman, Emilie. "Like It or Not, AI Has Arrived in Archives." Katina Magazine, March 2026. katinamagazine.org

[[22]]: AI is for Access: An Investigation of AI Adoption — Final Report. CLIR/DLF Born-Digital Access Working Group, March 2026. osf.io/gzd6k. In a survey of US archival workers, 308 respondents were asked how significant ethical concerns were in whether they use AI; 59% described themselves as highly concerned. Asked to elaborate, nearly a third cited environmental, climate, and resource costs.

[[23]]: American Alliance of Museums and Wilkening Consulting. AI in Museums and Community Trust: A 2025 Annual Survey of Museum-Goers Data Story. March 30, 2026. aam-us.org. In a survey of 2,045 US adults fielded January 2026, 45% said they want to know every time a cultural heritage institution uses AI, and a further 36% said they want to know when AI is used for interpretive content specifically — 81% in total.

[[24]]: Elsevier / Digital Applications in Archaeology and Cultural Heritage. Declaration of Generative AI Use Policy.sciencedirect.com

[[25]]: Hansen, Dave. "Library and Archives 101: AI and the False Promise of Control." March 2026. authorsalliance.substack.com

[[26]]: Europeana Foundation. Europeana 2026 Business Plan and AI4Culture Project. europeana.eu

[[27]]: Lo, L. "Memory Without Origin: Provenance, Consent, and Trust in the Age of Generative AI." University of Virginia Library, February 2026. library.virginia.edu

[[28]]: Lo, Leo S. The University of Virginia Archival AI Protocol. Version 1.1, January 27, 2026. University of Virginia Library. library.virginia.edu

[[29]]: This question is informed by the value co-creation framework developed by Derda, Izabela, & Predescu, Domnica in "Towards human-centric AI in museums: practitioners' perspectives and technology acceptance of visitor-centered AI for value (co-)creation." Museum Management and Curatorship 40:4, 532–554. 2025. doi.org/10.1080/09647775.2025.2467703. Their analysis of AI adoption in museums returns consistently to the question of who benefits from a given use.

[[30]]: Scheinfeldt, Tom. "Generative Artificial Intelligence and Archives: Two Years On." foundhistory.org, 2025. foundhistory.org

[[31]]: Vallor, Shannon. Technology and the Virtues: A Philosophical Guide to a Future Worth Wanting. Oxford University Press, 2016.

[[32]]: Vallor, Shannon. The AI Mirror: How to Reclaim Our Humanity in an Age of Machine Thinking. Oxford University Press, 2024.

[[33]]: Dede, in Mineo, "Is AI dulling our minds?"

[[34]]: Hardman, "Like It or Not, AI Has Arrived in Archives."

[[35]]: Hansen, "Library and Archives 101: AI and the False Promise of Control."