# AI disclosure: what you must declare in 2026

> A German court has held that a single undeclared use of generative AI can cross the line into academic misconduct. What belongs in the declaration, what belongs in an AI log, and where the grey areas are.

Acurio · 2026-07-26 · AI disclosure, Academic integrity, Declaration of authorship, Misconduct

Two students at the University of Kassel had signed a statement that they had written their work without unauthorised assistance. One was a term paper in a public management master's, the other a bachelor's thesis in computer science. The university classified both cases as serious academic misconduct: fail, plus exclusion from the resit. That ended both degrees.

On 25 February 2026 the seventh chamber of the Administrative Court of Kassel dismissed both claims (case nos. 7 K 2134/24.KS and 7 K 2515/25.KS). The decisive line in the reasoning: where unauthorised outside help is concerned, it does not matter whether it came from a human or from an AI. The threshold to a non-independent examination performance is crossed "already with a single undeclared use of generative AI".

**In short:** The Administrative Court of Kassel held on 25 February 2026 that a university may treat undeclared AI use in an assessed piece of work as serious academic misconduct. The basis is the signed declaration of authorship, not a ban on AI. The judgments are not final; appeal to the Hessian Higher Administrative Court remains possible.

You probably use AI in some form: for the outline, for rephrasing, for finding literature. Most institutions permit that. What they do not permit is concealing it. This article shows you where the line runs, what belongs in the declaration, and which four grey areas most students misjudge.

## What the ruling actually means

First the framing, so you do not misread the situation.

| What holds | What does not hold |
|---|---|
| A university **may** treat undeclared AI use as serious misconduct | AI use is not banned as such |
| Even a **single** undeclared use can be enough | The judgment is **not final** (appeal possible) |
| The signed declaration is the lever | There is still **no uniform national rule** |

The point is not that AI is forbidden. The point is that your signature under the declaration of authorship is a statement about how you worked. If that statement is untrue, that is the allegation, not the AI use itself.

**In practical terms:** the declaration costs you five minutes and removes the entire basis for the charge. Concealment saves five minutes and risks your degree.

The second half of protecting yourself is the referencing. If you used AI for research, you can run your bibliography against catalogues and databases in about ten seconds: [free, no signup](https://app.acurio.ch/quick-check?lang=en&utm_source=blog&utm_medium=body_cta&utm_campaign=eigenstaendigkeitserklaerung-ki).

## The four grey areas people misjudge

### Grey area 1: "I only rephrased"

The most common misjudgement. Rephrasing feels like correction, but technically it is text generation: you put a text in, a model writes a new one.

The line does not run along the product name but along the function:

| Tool | What it does | Declare? |
|---|---|---|
| Word spell check | flags errors | usually no |
| LanguageTool (grammar) | suggests corrections | mostly no |
| DeepL (translation) | produces new text | mostly yes |
| DeepL Write / ChatGPT (rephrasing) | produces new text | yes |
| ChatGPT (outline, research) | produces content | yes |

When in doubt: if the tool produces text you keep, it belongs in the declaration.

### Grey area 2: "I only used it for research"

Research is the area most institutions explicitly permit. It still belongs in the declaration, and for a second reason: the sources it produces are unverified.

A study by Rossiter analysed 2,300 AI-generated references. 71 per cent were faulty. That is not an edge case, that is the majority.

Sarah, a fifth-semester sociology student, had found eight studies for her seminar paper through a chatbot and declared them properly. That took the misconduct charge off the table. What remained: two of the eight studies did not exist. Her supervisor found the first while reading the references, the second while looking up the journal. The work was not treated as misconduct, but it was marked down two grades for poor referencing.

Declaration and source checking are two separate obligations. Neither replaces the other.

> **If your reading list came out of AI research:** check it before you submit. Paste the bibliography, wait ten seconds, see which entries can be verified. [Free check, no signup →](https://app.acurio.ch/quick-check?lang=en&utm_source=blog&utm_medium=body_cta&utm_campaign=eigenstaendigkeitserklaerung-ki)

### Grey area 3: "My institution has no rule on this"

That is the normal case, not the exception. Only about 35 per cent of German universities have an institution-wide AI policy as of 2026; in the UK and US, guidance is often set per faculty or per module rather than centrally.

A missing rule does not mean "anything goes". The declaration of authorship still applies, and it usually contains a phrase such as "without unauthorised outside assistance". If nobody has defined what is authorised, you carry the interpretive risk.

The order in which to look:

1. **Examination regulations** for your programme (binding)
2. **Module handbook** or departmental guidance
3. **Assignment brief** or the VLE page for the course
4. **Ask**, in writing, by email, so you have an answer on record

Point 4 matters most. A written answer from the person marking your work is your best evidence if it ever comes to a dispute.

### Grey area 4: "I signed the declaration before I used AI"

More common than you would expect: form filled in at the start, a tool brought in later during the work, form never updated.

The declaration refers to the state you submit, not to the moment you signed. Check it against your actual working practice on submission day.

## What belongs in the declaration

Your institution prescribes the exact wording. If you are free to phrase it yourself, three pieces of information cover what matters:

**1. Which tool.** Product name and, where known, the version. "ChatGPT (GPT-5, web version)" rather than "an AI".

**2. What for.** The step in the work, not the output. Outline, literature search, language editing, code generation, transcription.

**3. To what extent.** Whether you kept the output, reworked it, or used it only as a prompt for your own thinking.

A workable formulation looks roughly like this:

> In producing this work I used ChatGPT (GPT-5) to structure the outline and to edit the language of individual paragraphs. All substantive claims and all references were researched independently and verified against the original sources. Sources suggested by AI were checked for existence and content before being used.

The last sentence is the most valuable one, provided it is true. It pre-empts the second allegation, the one that always follows the misconduct question: the quality of your referencing.

## The AI log

Some institutions additionally require a log that breaks the use down by step, sometimes with sample prompts in an appendix. A simple format that is accepted in most places:

| Step of the work | Tool | Nature of use |
|---|---|---|
| Outline, chapters 2 to 4 | ChatGPT (GPT-5) | suggestions, reworked before use |
| Literature search, theory section | Perplexity | each hit checked against the catalogue |
| Language editing | DeepL Write | paragraphs 3.2 and 4.1 |
| Interview transcription | Whisper | full, manually corrected |

Two notes from practice. Ask **before** you submit whether a log is required, because filing it afterwards is usually impossible. And keep it **while** you work, not from memory at the end.

## What the declaration does not cover

This is where most work runs into trouble despite a correct declaration.

| Allegation | Covered by the declaration? |
|---|---|
| Unauthorised aid | ✅ yes, if properly declared |
| Fabricated source in the bibliography | ❌ no |
| Source says something other than claimed | ❌ no |
| Copied text without attribution | ❌ no |

Thomas, a business informatics student, had documented his AI use impeccably: tool, steps, prompts in an appendix. No misconduct proceedings. What he did have were three conference papers in his bibliography that had never taken place. His examiner treated that as grossly negligent referencing. The declaration protected him; the unverified sources did not.

The difference between the two classes of error is explained in [plagiarism vs. misquotation](https://acurio.ch/en/blog/plagiarism-vs-miscitation). How fabricated sources can be spotted is covered under [ChatGPT and hallucinated sources](https://acurio.ch/en/blog/chatgpt-hallucinated-sources).

> **Before you submit:** declare first, then verify. Together they take under an hour. [Check your bibliography now →](https://app.acurio.ch/quick-check?lang=en&utm_source=blog&utm_medium=body_cta&utm_campaign=eigenstaendigkeitserklaerung-ki)

## Different countries, same mechanism

The Kassel judgments apply in Germany. The underlying construction is the same almost everywhere: you sign a statement about how you worked, and that signature is the hook.

| Country | Usual name | Where the rules live |
|---|---|---|
| Germany | Eigenständigkeitserklärung, eidesstattliche Erklärung | examination regulations, module handbook |
| Austria | eidesstattliche Erklärung | Studienplan; for school-level work, the exam regulations |
| Switzerland | Selbstständigkeitserklärung, Redlichkeitserklärung | study and examination rules |
| UK | declaration of authorship, academic integrity statement | programme handbook, department policy |
| US | honour code statement, academic integrity pledge | honour code, syllabus |

One case deserves separate attention. **Universities of applied sciences** often place work with industry partners. If your project or placement report is written inside a company, the confidentiality question arrives alongside the AI question: pasting company data into a chatbot can be a problem quite independently of any examination rule. Clear that with your placement supervisor before you use a tool.

## If you notice the use only after submitting

It happens: you are tidying up and realise that back in December you had a chatbot rephrase a paragraph, and the declaration says nothing about it.

The order that leaves you the most room to act:

1. **Check whether the use was declarable at all.** Pure grammar correction is not, under many sets of rules.
2. **If it was, report it yourself**, in writing, to the examinations office or your supervisor, before somebody else finds it.
3. **Describe it factually**: which tool, which section, when. No justification, no embellishment.

Self-reporting before marking is treated differently almost everywhere from a discovered omission. Both Kassel cases were discovered omissions.

And the point that usually gets missed: while you are at it, check the sources in that section too. Where a tool was involved, the references are usually unverified. More on that under [checking your term paper bibliography](https://acurio.ch/en/blog/term-paper-source-check).

## Checklist before you submit

- [ ] Examination regulations, module handbook and assignment brief checked for AI rules
- [ ] Where unclear, asked **in writing**
- [ ] Every tool noted, including the ones used "only for rephrasing"
- [ ] Declaration reconciled with your actual practice on submission day
- [ ] Established whether an AI log is required
- [ ] All AI-suggested sources checked for existence
- [ ] Central references additionally read against the original

## What to take away

1. **The use is not the problem, the concealment is.** That is exactly what the Kassel court turned on.
2. **A single undeclared use can be enough.** There is no threshold you can rely on.
3. **Rephrasing counts as use.** If the tool produces text, declare it.
4. **No institutional rule does not mean permitted.** Two thirds of universities have none, and you carry the interpretive risk.
5. **Declaration and source checking are two obligations.** One does not protect you from the other.

The declaration takes five minutes to write. Checking your sources takes half an hour. Together they are the cheapest thirty-five minutes of your term.

**Start with the part you can automate:** paste your bibliography into the quick check and see which entries can be verified. Free, no signup.

[Check your bibliography now →](https://app.acurio.ch/quick-check?lang=en&utm_source=blog&utm_medium=body_cta&utm_campaign=eigenstaendigkeitserklaerung-ki)


## Frequently asked questions

### Do I have to declare spell checkers and translation tools?

Classical spell checking usually not, generative tools usually yes. The line does not run along the product name but along the function: does the tool correct your text, or does it produce new text? DeepL Write and ChatGPT produce text and therefore belong in the declaration, even if you only used them to rephrase.

### What happens if I used AI but did not declare it?

That depends on your institution's regulations. On 25 February 2026 the Administrative Court of Kassel confirmed in two cases that a university may treat undeclared AI use as serious academic misconduct, with the consequence of a fail plus exclusion from the resit. The judgments are not yet final but they show the direction of travel.

### Is there a standard wording for an AI declaration?

No. As of 2026 only about a third of German universities have an institution-wide AI policy, and UK and US practice varies just as much between faculties. What governs is your institution's own form, never a template from the internet.

### What is an AI log and when do I need one?

An AI log lists which tool you used for which step of the work, often with sample prompts in an appendix. Some institutions require it in addition to the declaration, others not at all. Ask before you submit, because filing it afterwards is usually not possible.

### Does declaring AI use protect me from problems with fabricated sources?

No. These are two separate allegations. The declaration protects you against the charge of using an unauthorised aid. A fabricated source is still a misquotation, and you are responsible for it regardless of where it came from.

---

Source: https://acurio.ch/en/blog/ai-declaration-of-authorship