Academic Writing
Role-based exercise set along the writing 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 the critical use of generative AI as a writing assistant in several clearly delimited roles. The roles follow a typical academic writing process: clarify the problem → develop ideas → structure → draft → revise → finalize.
The goal is not to “automate” the production of text. The goal is to raise the quality of your decisions: better research questions, more consistent arguments, more precise language. At the same time the authorship and the evidence stay traceable.
At a Glance
| Item | Value |
|---|---|
| Duration | about 45 to 60 minutes |
| Level | Beginner |
| Core path | 1 → 2 → 3 → 4 → 6 → 8 |
| Optional deepening | 5 (blockages), 7 (style and translation), 9 (readability test) |
| Hand-in (minimum) | One revised section of text. AI log: 6 entries. |
The core path leads you to a revised text in 45 to 60 minutes. Each optional deepening costs about 15 minutes more. Write one entry in the AI log for each step of the core path.
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 two following things:
- Work only with the material that you give it. You check this constraint in every step.
- Deliver an answer in a format that you prescribe, for example as an outline.
This exercise needs neither a web search nor a file upload. A free account is enough.
Other tools and access
- A fragment of your own topic, for example from a seminar paper. A section of half a page to one page is enough.
- Your own bullet points or notes on the topic, at least 10 points.
- 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
- Treat answers as hypotheses or drafts, and check them systematically.
- Design prompts so that the AI does not “invent”, but works with the material that you supply.
- Draft an argument structure (thesis, evidence, conclusion) and test it for consistency.
- Have paragraphs written from your own bullet points, and then revise them in subject and terminology.
- Run the revision in two steps: first the logic and the clarity, then the language and the style.
- Document your own AI use transparently in the AI log.
Methodological Frame: AI as Writing Infrastructure
This exercise treats AI as a heuristic writing tool. It supports wording, structuring and checking, but it replaces neither the work with sources nor the decisions of interpretation. (Bender et al. 2021; Campbell 2025; Oberbichler and Petz 2025) A clear division of labour helps:
- Historical authorship: the research question, the choice of evidence, the interpretation, the normative positions.
- Contribution of the AI (instrumental): generate variants, suggest structures, test readability, diagnose errors.
- Verification procedure (mandatory): every AI statement must be (a) traceable to your input material, or (b) marked as a hypothesis and verified externally. (The Turing Way Community 2022)
For more reliable work, use two principles of prompting consistently. 1. Delimit the workspace (“work only with my bullet points; mark uncertainties”). 2. Ask for a format (lists, argument maps), so that you can check systematically and adopt selectively.
Use AI only as a tool of transformation and verification for your material (notes, excerpts from sources, summaries of literature, drafts of an outline). If the AI names additional facts, literature or sources, that counts as a hypothesis for a search that you must verify externally.
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
The roles form phases. You can skip single steps. The core path in the section “At a glance” leads you through one complete round up to the revision.
- Comprehension checker (clarify the task, identify blind spots)
- Brainstorming partner and devil’s advocate (open the space of ideas and find counter-arguments)
- Structurer (the architecture of the argument)
- Text from bullet points (drafting under fact discipline)
- Blockage solver (local writing problems, transitions)
- Logic checker (coherence, evidence, rules of inference)
- Copy editor and translator (style, grammar, register; several languages as an option)
- Abstract writer (compression, focus, “so what?”)
- Comprehension checker (second run) (readability from the view of the target group)
1. Comprehension Checker
Goal
Make the writing task and your own intent so precise that the later AI steps do not “evade” and do not stabilize a false premise.
Task
- Give the AI only this: the topic in 2 to 3 sentences, the context (course, genre), the target group, the length and the format of the hand-in.
- Have the AI ask questions and mark uncertainties.
Prompt scaffold
You are a comprehension checker. Ask me 8–12 precise questions to clarify my writing task and my goal of argument.
Mark:
(1) the missing information,
(2) the implicit assumptions,
(3) the terms that need an operational definition.
No suggestions for content, only questions and diagnosis.
Context: …
Topic and thesis (preliminary): …
Genre and length: …
Target group: …Critical Comparison
- Which questions force you to clarify terms (for example “modernization”, “neutrality”, “governance”)?
- Which questions would you prioritize differently, and why?
Result:
- 1 revised task (3 to 5 sentences)
- 3 defined key terms
2. Brainstorming Partner and Devil’s Advocate
Goal
Generate ideas without inventing evidence, and think about counter-arguments and alternative readings at once.
Task A: Brainstorming (oriented on ideas, neutral on evidence)
Role: brainstorming partner. Generate 12 possible lines of argument or hypotheses on my topic.
Conditions:
- Formulate each line as a claim that I can check (no facts).
- For each line, state which kind of evidence I would need (source types, literature types).
- Avoid concrete literature titles or archive references; if necessary, formulate a search heuristic.
My topic: …
My material (types only, no sensitive content): …Task B: Devil’s Advocate (Confrontational, But Productive)
Role: devil's advocate. Choose the 3 most plausible of my hypotheses and argue against each one.
For each counter-argument:
- Name the vulnerable assumption.
- Suggest a "stress test" (which finding would overturn my argument?).
- Formulate an alternative reading (max. 2 sentences).
Use no external facts.
Hypotheses: …Critical Comparison
- Which hypotheses are falsifiable, in the sense of clear conditions of refutation?
- Where does presentism threaten (a category of the present projected onto the past)?
Result:
- 1 hypothesis or thesis with priority
- 2 central counter-arguments
- 1 planned stress test
3. Structurer
Goal
Draft a chain of argument that you can check (thesis → sub-theses → evidence → conclusion), before you generate any text.
Task
Insert your own word count in the following prompt. The example names 3000 words.
Role: structurer.
Build an outline (max. 8 main points) for a text of about 3000 words.
For each point:
- the goal of the section (1 sentence)
- the central claim (1 sentence)
- the evidence that it needs (in keywords; "from my notes or sources")
- the transition to the next point (1 sentence)
Constraints:
- No new facts.
- If evidence is missing, mark it as [EVIDENCE GAP] instead of adding something.
Material (bullet points, notes): …Quality Check
- Does each claim carry a trail of evidence, at least as a placeholder?
- Are there sections that only give an “overview” but carry no function in the argument?
Result:
- 1 outline
- 2 to 3 marked gaps in the evidence
4. Text from Bullet Points
Goal
Have paragraphs written from your own material, without giving up control over the content.
Task
Choose one point of the outline (for example “main argument 1”) and supply 8 to 12 bullet points (evidence, observations, quotations or paraphrases, terms).
Role: text generator from bullet points.
Write a paragraph (120–180 words) in academic English.
Conditions:
- Use only information from my bullet points.
- No new examples, data, names or sources.
- If a sentence would need additional information, mark it as [UNCERTAIN].
- Keep my terminology; ask me when something is unclear.
Bullet points: …Critical Comparison
- Delete every sentence that you cannot trace directly to your bullet points.
- Check the terminology: where does the AI smooth a term impermissibly (for example “state”, “society”, “actor”)?
Result:
- 1 paragraph
- 1 second version of the paragraph, in which you sharpened 3 to 5 sentences yourself
5. Blockage Solver
Goal
Solve local problems (transitions, opening sentences, definitions of terms), without a rewrite of the text.
Micro-Exercises (Choose 2)
A) Build a transition
Role: blockage solver.
Formulate 5 variants of a transition sentence between paragraph A and paragraph B.
Variant 1 neutral, 2 stronger in argument, 3 stronger in explanation, 4 shorter, 5 with a marker of contrast.
Paragraph A: …
Paragraph B: …B) Operationalize a definition
Replace the term “neutrality” in the following prompt with your own term.
Formulate 3 possible working definitions of the term "neutrality" in the context of my topic.
For each definition:
- the delimitation from 1 neighbouring term
- 1 observable indicator (how do I recognize "neutrality" in sources?)
No external facts.
Context: …C) Sentence diagnosis
Mark in my paragraph:
- the central sentence (the topic sentence),
- the sentences with an unclear referent ("this", "thereby", "one"),
- the sentences that in fact hold two claims.
Suggest minimal rearrangements (no rewriting of the content).
Text: …Result:
- 1 revised transition or 1 revised definition
- 1 short reason why you chose this variant
6. Logic Checker
Goal
Make the logic of the argument visible: does the evidence support the claims? Are the rules of inference traceable? Where is a premise missing?
Task
Role: logic checker.
Analyse my text (max. 300 words) as an argument map:
1) the main thesis
2) the sub-theses
3) the explicit evidence (from the text)
4) the implicit premises (that I apparently presuppose)
5) the possible jumps in logic or contradictions
Then give 5 concrete recommendations for a revision, ordered by effect.
Text: …Critical Comparison
- Which “implicit premises” do you accept, and which ones need evidence or a new wording?
- Where is a claim of cause in fact only a sequence in time?
Result:
- 1 argument map
- 2 targeted revisions in the text, with a note on what improved logically
7. Copy Editor and Translator
Goal
Improve the language and the register without a change to the substance of the argument. As an option, carry the terminology consistently into a second language.
Task A: Copy Editing (Minimally Invasive)
Role: copy editor.
Improve the grammar, the style and the precision, but change no statement of content.
Work in a minimally invasive way:
- Replace a vague word (for example "important", "shows") only when the replacement is unambiguous.
- Keep the technical terms; when you are unsure, mark [TERM?].
At the end, give a list of the 5 most frequent patterns of error that you corrected.
Text: …Task B: Translation (Optional)
Role: translator.
Translate the text into another language (for example French, German or Spanish).
Conditions:
- Keep the register (academic) and the hedges (for example "could", "suggests").
- Keep the proper names and the quotations unchanged.
- Give a small glossary table (5–10 pairs of terms) for the key terms.
Text: …
Target language: …Result:
- 1 corrected version
- optional: 1 translation
- optional: 1 glossary
8. Abstract Writer
Goal
Condense the text to its load-bearing statement: the research problem, the approach, the finding or argument, and the relevance.
Task
Role: abstract writer.
Write an abstract of 120–160 words.
It must hold:
- the research problem (1 sentence)
- the approach and the material (1 sentence)
- the core argument or finding (2–3 sentences)
- the contribution and the relevance ("so what?") (1 sentence)
No new facts; when something is unclear, mark it as [UNCLEAR].
Text base: …Quality Check
- Does the abstract cover the actual point of the argument, or only the topic area?
- Does the method that you claim match what you really do (work with sources against a discussion of literature)?
Result:
- 1 abstract
- 1 meta comment in 2 sentences: what did the AI emphasize wrongly, and what did it leave out?
9. Comprehension Checker (Second Run)
Goal
Test the readability and the room for interpretation from the view of the target group.
Task
Role: comprehension checker.
Read my text as a critical reader in a seminar.
1) What is the thesis in one sentence?
2) Which 3 points are unclear or ambiguous (each with the passage quoted)?
3) Which passage looks unsupported?
4) Which two questions would you put to me in your feedback?
Text: …Result:
- 3 sharpened passages in the text, each revised in 1 to 2 sentences
Learning Outcome
At the end you hold a small package of artifacts that you can reuse in your academic work:
- an outline (with the gaps in evidence)
- 1 to 3 revised paragraphs (with traceable revisions)
- an argument map (from the logic check)
- an abstract (and, as an option, a translation and a glossary)
- an AI log with at least 6 entries
Further Resources
- AI Pedagogy Project: LLM Tutorial – Limits, hallucinations, verification and critical use of LLMs.
- The Turing Way – Handbook for reproducible, ethical and collaborative research.
Bibliography
Citation
@inreference{mähr2025,
author = {Mähr, Moritz},
title = {Academic {Writing}},
booktitle = {Critical AI Literacy for Historians},
date = {2025-12-24},
url = {https://maehr.github.io/critical-ai-literacy-for-historians/en/exercises/writing.html},
langid = {en}
}