Public History and Historical Communication
Knowledge communication along the writing and production process
This exercise was translated automatically from the German original and can contain errors. Consult the original version in case of doubt.
Overview and Didactic Goal
This exercise trains knowledge communication in public history as translation work in a process: from a scholarly article (argument, evidence, uncertainties) to formats for an audience. You use generative AI not as an authority, but as a tool for drafting, variation, editing and proofreading, with a consistent reference to the sources, with traceability and with clear limits.
At a Glance
| Item | Value |
|---|---|
| Duration | about 4 hours |
| Level | Intermediate |
| Core path | 1 → 2 → 3 → 4 → 5A → 6 |
| Optional deepening | 5B (podcast script), 5C (poster text blocks) |
| Hand-in (minimum) | A blog post of 700 to 1000 words. AI log: 8 entries and a note of reflection of 300 to 500 words. |
The core path leads you to a blog post and to two short formats in about 4 hours. Split the time over two sessions. Each optional deepening costs about 45 minutes more.
Instead of the blog post you can hand in a podcast script of 6 to 8 minutes. Replace step 3 with step 5B for that.
The blog post stays a hand-in of text, so the exercise names a word count for it. For the AI log you count entries.
What You Need
Subject knowledge
- A basic understanding of historical research methods
- Basic knowledge of work with generative AI, above all prompting
If prompting is new to you, we recommend that you work through the exercise Prompt Engineering first.
Capabilities of the AI system
The system must be able to do the three following things:
- Read a file that you upload, that is your base article and your source package.
- Work only with the material that you give it. You check this constraint in step 4.
- Deliver an answer in a format that you prescribe, for example as a table with claim IDs.
Other tools and access
- A scholarly article as the base text. Choose a freely accessible article from your field.
- A file for your AI log.
You can do every exercise with AI systems from different makers. What a system can do depends on the product and on your account. Settle the five following points before you start.
Cost
A free account is enough for short text tasks. For some capabilities you need a paid account. This mostly concerns the search on the web, the reading of long files and the built-in reasoning mode. Check first whether your university provides access.
Tools
Check which tools your account unlocks. Can the system search the web? Can it read a file that you upload? Can it call external services? Each exercise names the capabilities that it requires.
The section Check and classify capabilities explains five newer capabilities. It also describes what each capability means for your work with sources.
Memory
Some systems carry over statements from earlier sessions or from a project context. Check the settings for memory, projects and chat history. A new session therefore guarantees no empty context.
Account: institutional or private
Check which account you work with. The rules differ in three points.
| Point | Account of your institution | Private account |
|---|---|---|
| Contract | Your university concludes it and shares the liability. | You conclude it yourself and carry the liability alone. |
| Training | A contract often excludes the use of your inputs. | Many providers use inputs for training. |
| Permitted material | The university rules which data you can enter. | You have no such rule and carry the risk yourself. |
Ask the IT office or the library of your university when you cannot find the rules. Record the answer in your AI log.
Do not upload material that you have no right to pass on. An upload leaves your computer and you cannot take it back.
These materials stay outside the AI system:
- Personal data of living persons, for example from interviews, from private papers or from holdings under a protection period.
- Digitized items whose terms of use forbid a disclosure.
- Unpublished texts of other persons, for example manuscripts or reviews.
- Licensed full texts, when the license of your library forbids a disclosure.
In case of doubt, work with a short excerpt or with an anonymized version. You can also use a local model, because the material then stays on your computer.
Law: data protection and disclosure
Two areas of law concern your work directly. This section gives an orientation and no legal advice.
Data protection. In the EU the General Data Protection Regulation applies. In Switzerland the revised Data Protection Act applies (Verordnung (EU) 2016/679 Zum Schutz Natürlicher Personen Bei Der Verarbeitung Personenbezogener Daten (Datenschutz-Grundverordnung) 2016, Bundesgesetz Über Den Datenschutz (DSG) 2020). Both laws protect personal data of living persons. The text of the law itself is binding. Article 6 GDPR requires a legal basis for every processing operation, and an upload is a processing operation.
The role of the provider decides the classification. A provider can act as a processor and process your data only on your behalf. Article 4 point 10 GDPR expressly excludes a processor from the third parties. This role needs a contract under Article 28 GDPR. Without such a contract, a provider uses your inputs for its own purposes. An upload is therefore not always a disclosure to a third party, but without a contract you must assume that it is. Check the contract that applies to your account.
Research is subject to safeguards, but it gets no exemption. Article 89(1) GDPR requires appropriate safeguards, for example data minimisation and pseudonymisation. Article 89(2) GDPR permits the legislator to derogate from single rights of the data subject. The information of the data subjects can also fall away, namely under Article 14(5)(b) GDPR. That applies only when the information proves impossible or would involve a disproportionate effort. Research therefore does not release you from the information duty in general. Settle your individual case with the data protection office of your university.
The European Data Protection Board presented a guideline on research in 2026 (European Data Protection Board 2026). This guideline was in public consultation until June 2026. It is therefore not applicable law, but an aid to interpretation. For deceased persons, the protection periods and the rules of use of the archive apply.
Disclosure. The AI Act of the EU governs the transparency of generated content (Verordnung (EU) 2024/1689 Zur Festlegung Harmonisierter Vorschriften Für Künstliche Intelligenz (Verordnung Über Künstliche Intelligenz) 2024). Article 50 applies from 2 August 2026. It separates two addressees: the provider of an AI system and the deployer. You are a deployer as soon as you use an AI system professionally. For purely private use you are not a deployer, under Article 3 point 4.
As a deployer you carry two separate duties from Article 50(4).
- Deep fake. If an AI system generates or manipulates an image, an audio recording or a video, and the content is a deep fake, then you disclose that. Under Article 3 point 60, a deep fake resembles existing persons, places or events and falsely appears authentic. A generated image without this resemblance to reality does not fall under this duty.
- Text on public matters. If you publish generated text to inform the public on a matter of public interest, then you disclose that. The duty falls away when a person reviewed the text and holds the editorial responsibility for the publication.
Article 50(4) therefore requires no disclosure for every generated image and for every generated text. Article 50(2) addresses the provider instead. The provider must mark generated content in a machine-readable format. This duty does not lie with you, and it does not replace your own disclosure.
The legal disclosure and the academic disclosure are two separate duties. Your university requires a declaration even when the law requires no disclosure. The exercise Citing shows how you declare your AI use.
Record your answers to these five points. They belong to the description of your work environment.
Learning Objectives
- Break a context of communication into target groups, medium, purpose and conditions (a communication briefing).
- Derive a core message (a message map) and a matrix of evidence from a scholarly article, without inventing new facts.
- Build drafts in several public history formats (blog, audio, social media, poster) and revise them to fit each format.
- Check answers systematically (claim by claim, source veto, marked uncertainties) and document the changes traceably in the AI log.
- Prepare an interview kit (soundbites, questions and answers, a no-go list, a statement of transparency).
The base text is a scholarly article from your field, if possible in open access. For example:
Methodological Frame: AI as a Critically Considered Communication Infrastructure
Terms for this exercise:
- Source veto: every narrative sharpening is allowed, as long as it does not contradict the statements that the article supports.
- Claim-by-claim check: you break statements into claims that you can check. You link each claim to its evidence, or you mark it as speculation.
Search engines and AI systems structure visibility in a problematic way. (Noble 2018) AI-generated content for public history carries a particular risk, because it reaches a broad audience without subject expertise. (Bender et al. 2021)
Work with a source package: the title and DOI, the abstract, the central results in your own notes, a few key points with their evidence (pages or sections), and figures or tables if you have them. Give the AI only material that you can control yourself, and let the AI work only from that material.
Plan a short element of transparency at the end, for example “How we worked” or “AI support”. State what you used the AI for (structure, variants of language, proofreading) and what you did not use it for (facts, decisions of interpretation).
This exercise leads to products for an audience. Beside the academic disclosure, the AI Act of the EU therefore applies (Verordnung (EU) 2024/1689 Zur Festlegung Harmonisierter Vorschriften Für Künstliche Intelligenz (Verordnung Über Künstliche Intelligenz) 2024). Article 50(4) applies from 2 August 2026 and requires a disclosure from you as a deployer in two cases.
- A deep fake in an image, an audio recording or a video. Disclose it when generated or manipulated image, audio or video content is a deep fake. A deep fake resembles existing persons, places or events and falsely appears authentic. Check the formats 5A to 5C on this point one by one.
- Text on public matters. Disclose it when you publish generated text to inform the public on a matter of public interest. This duty falls away when you reviewed the text and hold the editorial responsibility.
Article 50(4) therefore requires no marking for every generated image. A generated image without a resemblance to existing persons or events does not fall under it. The separate duty to mark generated content in a machine-readable format lies with the provider under Article 50(2), and not with you.
The exception for reviewed text does not lift your academic disclosure. Disclose your AI use even when the regulation requires no marking. The exercise Citing shows how you meet both duties.
An AI log is a short record of your AI use. In each entry you record five things: which material you entered, what you asked, what the AI answered, how you checked the answer, and how you decided.
The AI log works like a research journal. The way to the result stays traceable, for your readers and for you. You need the log later for the method section and for the disclosure of your AI use.
The German version of this course calls this record the KI-Protokoll. The English literature also calls it an audit trail. That term comes from accounting. There, an auditor follows a trail of records and traces each figure back to its origin. Your log does the same for your work steps. In these exercises we use only the term AI log.
Write the entry directly after the work step, and not at the end of the exercise. After one hour you no longer know the exact wording of your prompt.
Where you keep the log
You need no special tool. A text file or a document in your word processor is enough. A sheet in a spreadsheet program also works.
- Create one file per exercise.
- Give the file a clear name, for example
ai-log_source-criticism.md. - Write each entry below the previous one in the same file.
- Save the file together with your hand-in.
How long an entry is
An entry is short. It runs to five or ten lines and costs you one or two minutes. One entry covers one work step, that is one prompt and the answer that belongs to it.
Different rules apply to the prompt and to the answer:
- Prompt: verbatim. Copy the prompt exactly as you entered it. If the prompt is very long, copy the core passage and shorten with
[…]. - Answer: summarized. Summarize the answer in two or three sentences. Copy a passage verbatim only when you take that passage into your text, or when you want to document an error of the AI.
Each exercise states how much it requires. Some exercises name a number of entries, and others name a length in words. If an exercise names nothing, write one entry for each work step with AI.
Template for an entry
Copy this table again for each entry.
| Field | Content |
|---|---|
| Date | |
| Step (exercise, number) | |
| Goal (1 sentence) | |
| Material (input) | |
| Prompt (verbatim) | |
| Answer (short version) | |
| Verification steps | |
| Decision | |
| Reason (1 sentence) |
If you write in Markdown, copy this source text:
| Field | Content |
| ----------------------- | ------- |
| Date | |
| Step (exercise, number) | |
| Goal (1 sentence) | |
| Material (input) | |
| Prompt (verbatim) | |
| Answer (short version) | |
| Verification steps | |
| Decision | |
| Reason (1 sentence) | |A list with the same fields is equally valid. Choose the form that you fill in faster.
A filled-in entry as an example
The example comes from the exercise Source Criticism. A student there checks a memorandum of 7 July 1949.
| Field | Content |
|---|---|
| Date | 14 March 2026, 10:15 |
| Step (exercise, number) | Source Criticism, step 2 (external source criticism) |
| Goal (1 sentence) | I check whether the AI derives the formal data of the memorandum correctly from the text. |
| Material (input) | Memorandum Bern, 7.7.1949, Petitpierre/Hansen (Dodis 5020), full text from the edition copied into the prompt. |
| Prompt (verbatim) | “Analyse the following memorandum only from the viewpoint of external source criticism. Name the author, the addressee, the date, the text type and the form of transmission. Mark every statement that is not in the text expressly as an assumption. […]” |
| Answer (short version) | The AI names Petitpierre as the author, Hansen as the interlocutor and 7.7.1949 as the date. It calls the text a “minute of a Federal Council meeting”. It adds without marking that the memorandum lies “in the Federal Archives under the reference E 2001”. |
| Verification steps | I compared all statements with the edition at https://dodis.ch/5020. Author, date and interlocutor agree with the text. The text type does not agree: the text records a conversation, and it does not minute a meeting. The reference appears nowhere in the material. |
| Decision | Partly adopted. |
| Reason (1 sentence) | I adopt the author, the date and the interlocutor, I correct the text type to “memorandum”, and I reject the reference as unsupported. |
The entry shows the normal case: one part of the answer is usable, one part is wrong, and one part has no support.
Note the third category. The student does not claim that the reference is invented. She records only what she can check: the reference is not in the material. Whether the holding exists is a different question. You can call a statement wrong only after you checked it. Until then it has no support. This distinction is the core of source criticism, and it also applies towards the AI.
Structure of the Exercise
- Situation analysis: context, target group, purpose and constraints (the briefing)
- Planning: core message, evidence and narrative route (the message map)
- Draft: a first version (blog or podcast) with a clean logic of evidence
- Revision: check of facts and evidence, clarity, tone, ethics
- Remix and adaptation: short forms (social media, poster) and preparation for an interview
- Documentation: the AI log and the reflection
1. Analysis of the Target Group and Communication Briefing
Goal
Treat communication as a decision that depends on the context: for whom, where, why, and in which medium?
Task
Choose a concrete context of communication, for example a museum blog, a local newspaper, a science podcast, the social media account of an archive, or an exhibition panel. Build a communication briefing with:
- the target group or groups: prior knowledge, interests, possible triggers and misunderstandings
- the purpose: to inform, to contextualize, to open a debate, to give knowledge for action, and so on
- the medium and format: length, tone, images or audio, interaction
- the constraints: languages, accessibility, terminology, legal and ethical limits
- the criteria of success: what should the audience afterwards know, be able to do, or see differently?
AI Workflow (Prompt Sketch)
You are an assistant for audience research. Produce 2–3 plausible personas of a target group for the following context (without inventing facts about real people).
For each persona: prior knowledge, motivations, possible misunderstandings, suitable examples and metaphors, and no-gos.
Context: ...
Topic and article in 3 sentences: ...Critical Comparison
- A persona is not reality. Add to your assumptions and check them, for example against real information on the target group, museum concepts or media guidelines.
- Which simplifications would be risky in epistemic terms? Note the red lines.
2. Message Map and Narrative Route
Goal
Build a core architecture of the communication that stays stable across the formats.
Task
Build:
- A one-liner (1 sentence): what is the core message, and why does it matter for this target group?
- 3 to 4 supporting points (1 to 2 sentences each) with a reference to the evidence
- A matrix of evidence: which supporting point rests on which data or passage?
- The context: what must you explain, so that nothing gets distorted?
- The uncertainties: what is contested, limited, or only plausible and not proven?
AI Workflow (Prompt Sketch)
Role: you are a structurer. Work only with the source package and the briefing.
Produce a message map (a one-liner, 3–4 supporting points and a 'so what' for the target group).
Mark each point with a claim ID and point to the paragraphs of context that it needs.
Material: ...Critical Comparison
- Where does “presentism” threaten (the present projected onto the past), or a moral simplification?
- Where do you need terms and definitions instead of storytelling?
3. Draft I: Blog Post
Goal
Build a base model in text that serves as the reference for the further adaptations.
Task
Write a blog post of about 700 to 1000 words. The blog post holds:
- an opening that fits the target group and the purpose,
- a clear structure of sections with subheadings,
- an explicit logic of evidence (claim IDs, or footnotes and references to the evidence),
- a short section “What we know and what we do not know”,
- a call to action for further reading, for example a link to the article, to a glossary or to the sources.
AI Workflow (Prompt Sketch)
Role: you are a blog editor. Write a blog post for the following context.
Constraints: 900 words, clear subheadings, no new facts.
Use claim IDs in the text in square brackets, for example [C3].
Material: the briefing, the source package, the message map and the table of claims.Critical Comparison
- Check each paragraph: which claims does it make? Do they carry evidence?
- Check the tone and the register: “it sounds plausible” is no criterion of quality.
- Check the fit: which terms need an explanation?
4. Revision: Fact Check, Clarity, Ethics
Goal
Turn a plausible draft into a robust product of communication.
Task (Checklist)
- Claim by claim: every statement carries evidence, is marked as a framing, or goes.
- Clarity: reduce or define the technical terms that you do not need; check the examples.
- Ethics and responsibility: sensitive groups and persons, violence, stigmatization, data policy, image rights.
- Transparency: add a short note on the method, the sources and the AI support.
AI Workflow (Prompt Sketch)
Role: you are a fact checker.
1) List all the facts and statements that the draft claims.
2) Assign them to claim IDs, or mark them as 'not in the source package'.
3) Suggest more precise wordings that express the uncertainties correctly.
Input: the blog draft and the source package.Critical Comparison
- The AI can point to a gap. The verification stays your task: you check every piece of evidence against the source, and you answer for the result.
- Document the changes in the AI log (why did you change it? which passage supports it?).
5. Remix: Adaptations of the Format
The following tasks are deliberately adaptations: you reuse the message map and the baseline text instead of “inventing” anew in each medium.
5B Podcast Script (6 to 8 Minutes)
Goal: carry the argument and the context into an audio format (voice, rhythm, examples).
Task:
- 1 script with the segments hook, context, 2 to 3 key points, framing and conclusion
- speakable language with short sentences and active verbs
- show notes with 5 bullet points, 1 reference to a source and 1 glossary of 3 terms
AI workflow
Write a podcast script (7 minutes) for [target group and context].
Structure: hook (15s) – context (60s) – key points (3x90s) – uncertainties (45s) – conclusion (30s).
Use only claims from the source package, and mark the claim IDs in the margin of the script.5C Poster Text Blocks (A1 or A0, Exhibition or Conference)
Goal: a visual hierarchy and short texts that work without prior knowledge.
Task:
- 1 headline (at most 10 words) and 1 subheadline (at most 20 words)
- 3 text blocks of at most 60 words each: “What is this about?”, “What is new?”, “Why does it matter?”
- 1 caption for an image or a graphic (at most 30 words) with a reference to the source
- 1 text for a QR code (at most 15 words) with the target of the link
AI workflow
Produce poster text blocks with a strict word limit and without new facts.
Watch for: clear terms, no implicit judgements, and one claim ID per block.
Input: the message map and the matrix of evidence.6. Interview Kit
Goal
Speak about your research with confidence, precision and low risk, including a control of the “no-gos”.
Task
Build an interview kit of 1 to 2 pages. The kit holds:
- 3 soundbites of 12 to 20 seconds each, formulated in substance,
- 5 core questions with short answers of 3 to 5 sentences each,
- 2 critical questions (for example “Is that not …?”) with strategies to bridge them,
- a no-go list with statements that you do not make, because the evidence, the complexity or the ethics speak against them,
- a sentence of transparency on the method and on the state of the data (1 to 2 sentences).
AI Workflow (Prompt Sketch)
Role: you are a media coach. Produce an interview kit for radio or a newspaper.
Rules: no new facts, the soundbites must render the claims correctly, and they must not hide the uncertainties.
Input: the message map, the source package, and the blog or podcast draft.Critical Comparison
- Do the soundbites match the claim IDs, or do they “slip” into exaggeration?
- Are the strategies to bridge honest in substance (no evasion without a framing)?
Learning Outcome
At the end you hold these artifacts:
- A communication briefing (1 page)
- A message map and a matrix of evidence (as a table or in bullet points)
- A blog post (about 700 to 1000 words) or a podcast script (about 6 to 8 minutes)
- Two short formats from: social media, poster text blocks, interview kit
- An AI log with at least 8 entries and a note of reflection (about 300 to 500 words)
One methodological insight comes on top: you can give reasons for where the AI improves a communication, and where you carry the responsibility for the evidence and the interpretation yourself.
Further Resources
- AI Pedagogy Project: What is AI? – Conceptual introduction for non-technical audiences.
- Critical AI Literacy: Curated Resources – Ethics, power, labour and politics of AI.
- Drucker (2011) – Critiques the conventions of data visualization; introduces the concept of “capta”, the idea that data are constructed and not objectively given.
Bibliography
Citation
@inreference{mähr2025,
author = {Mähr, Moritz},
title = {Public {History} and {Historical} {Communication}},
booktitle = {Critical AI Literacy for Historians},
date = {2025-12-29},
url = {https://maehr.github.io/critical-ai-literacy-for-historians/en/exercises/public-history.html},
langid = {en}
}