From Australia to Germany, higher education policy points the same way: document the process behind student work.1,2
Document the process. Show the work.
A finished text rarely proves AI use in a way that holds up. That’s why Mentafy documents how a piece of work comes together and brings every signal into one Authorship Report: what students did themselves, and where AI or other help was involved.
So that good grades keep rewarding students’ own work, not access to better AI.
Plagiarism Checker and S³
also finds paraphrased passages
Reference Verifier
catches hallucinated references in the bibliography
AI Classification
an indicator, never a verdict
Writing Process Analysis
shows how the text came together
What you can count on
Complete and ready to go
Plagiarism, paraphrase, translation, fabricated sources, AI indicators, and the writing and research process: one system, one report. Up and running without a major IT project.
Future-proof
AI detectors aren’t the answer to the AI era: they’re a black box, and their false alarms hit honest students. Mentafy combines signals and documents how a text came together, where it came from, and whether its sources exist.
Privacy by design
Documentation, not surveillance: no keylogging, no webcam, no proctoring. GDPR-compliant, with data minimization, transparency, the right to erasure, and a record of processing activities. Servers in Germany. For US institutions, a FERPA compliance statement is available on request.
Doing the work should pay off. Honesty shouldn’t cost students a grade.
AI is part of everyday student life: in a survey of 28,584 students in Germany, 81 percent use it at least weekly and 46 percent daily.3 The problem isn’t AI use. It’s undisclosed AI use. Students who use AI covertly often get the better grade, and that hurts exactly the ones who work honestly.
of students rely more on AI than on traditional sources in their studies. 12% even submit AI-generated text unedited in assessed work, up from 3% in 2024.
HEPI and Kortext, UK, 20264of AI-written exam submissions went undetected in a blind test at a UK university, and on average they scored half a grade boundary higher than real students. These were unedited AI outputs, the easiest kind to spot.
University of Reading, PLOS ONE 20245Unauthorized AI use in unsupervised assessments generally can’t be proven with legal certainty. That’s the Bavarian government’s own conclusion, stated in its draft law.
Bavaria (Germany), draft BayHIG amendment, explanatory memorandum2And AI detectors? They deliver probabilities, not evidence.
No evidence
A score says how machine-like a text looks, not who did what. Every false alarm hits an honest person.
Blind to your rules
A detector can’t tell whether AI use was permitted and properly disclosed.
No finish line
Every new model forces detectors to catch up, and AI humanizers are built specifically to beat them.
A quarter of students submit text that combines an AI draft with their own revisions.4 For this hybrid, no threshold can separate honest from dishonest work. That’s why Mentafy uses AI classification only as one indicator among several.
We’ve seen this movie before
We lived through it with plagiarism in Germany: first the technology, then the tool, and action only after a scandal. AI is following the same script, with one crucial difference.
Then: plagiarism in Germany
-
The technology1998
Google launches
Other people’s texts are seconds away. Copying becomes trivial.
-
The tool2009
PlagScan launches
The technology is there, affordable and working. Adoption: low.
-
The trigger2011
The plagiarism scandal
Germany’s defense minister resigns over a plagiarized doctoral thesis, and more cases follow. Only public pressure gets things moving.
-
The standard2018
Checks become standard
Years after the scandal, plagiarism checks are part of everyday university life.
Today: AI, worldwide
-
The technology2022
ChatGPT
The finished text loses its value as evidence. One pass through a rewriting tool, and exact-match checking fails.
-
The tool2023
Mentafy Writing Process Analysis
Same founding team, same situation as in 2009: proven technology, adoption just beginning.
-
The trigger2026
Regulators move
Bavaria drafts a law: allow AI in unsupervised assessments, document its use. Australia’s regulator TEQSA names process traces as evidence of learning.
-
The standardNow
This time without a scandal
The trigger is already on the table, as regulation. Those who start now help shape the new standard.
Mentafy Check examines the text.
Trail documents how it was written.
Two products, one system, one shared report. Start with what your assessment policy allows today.
| Ready today Mentafy Check Checks the finished document. A drop-in replacement for your current AI or plagiarism checker. | Future-proof Mentafy Trail Documents how the work was written. For the evidence of students’ own work that regulation is moving toward. | |
|---|---|---|
| Key question | Were the sources used cited correctly? | How was the text written? |
| What you provide | The finished document, e.g., PDF, DOCX, or TXT, so students can write in Microsoft Word, Google Docs, or any other program. | Plus interim versions of the work:
|
| Assessment policy | No changes needed. You check submissions just as you would with traditional plagiarism software, only more thoroughly. | A clause on documenting AI use, in place before students start writing. Bavaria’s draft law calls for exactly this kind of rule.2 |
| Rollout | Create an account, upload a file. Also works as a replacement for an existing tool, with no IT project. LMS integration via LTI, e.g., with Canvas. | No IT effort if students upload their interim versions. A cloud connection adds more detailed data but isn’t required. Students write in the programs they already use. |
| Documentation frequency | Not applicable: only the finished document is checked. | You decide: automatically about once a minute, or manually. Even one version per week over a semester shows how a piece of work grew. Students can upload more versions voluntarily and strengthen their own evidence. |
| Analyses | Plagiarism matching, S³ Semantic Source Search for paraphrases and translations, Reference Verifier, and AI classification trained for English, German, French, and Spanish. | Everything in Mentafy Check, plus Research Recorder and Writing Process Analysis. Students can get real-time feedback on their citations. |
Mentafy Trail includes every analysis in Mentafy Check. Start with Check and add Trail once your assessment policy is ready, with no system switch.
Why document the process? Australia’s higher education regulator, TEQSA, lists process traces such as timestamped edits and research behavior as evidence of learning, explicitly as supporting evidence rather than proof.1 Bavaria’s draft law points the same way: it would require universities to set rules on the scope and form of AI-use documentation.2
To be upfront: Mentafy Trail is new. It was developed with several thousand students and validated on more than 10,000 documents, but its day-to-day use at universities is just beginning. We’re looking for institutions ready to take this step with us.
What a writing history looks like with dense documentation
The report doesn’t claim “copied.” It shows when text was added: valuable evidence for assessing a student’s own work.
The Authorship Report: evidence, not gut feeling
Every signal from Check and Trail in one report. It breaks the text down by origin, shows the evidence behind each finding, and says plainly what it can’t classify.
The corroboration standard: same AI signal, two findings
No finding rests on a single signal. The report brings together the signals from Check and Trail, weighs them against each other, and classifies each passage accordingly.
Case A
Plagiarism matching and S³no matching or similar text found
AI classificationmachine-like language patterns
Writing process41 saved versions over six weeks, with steady small revisions
Case B
Plagiarism matching and S³no matching or similar text found
AI classificationmachine-like language patterns
Writing processalmost the entire text appeared between two saved versions, with no earlier stages and no editing afterward
If the data isn’t sufficient, the report says “undetermined.” That’s a category, not a guess. Whether a use was permitted is decided by your assessment policy and your instructors, not by the software.
Three report levels, as deep as you need
Nobody reads sixty reports in full. So every report starts with a single rating and only opens up when you want it to.
Level 1
The traffic light
One rating. Green means no cause for concern, yellow means take a closer look, red means make sure to review it. That’s all it says.
For initial screening
Level 2
The one-pager
Shares of plagiarism, AI plagiarism, and quotations, the number of faulty references, and the work pattern: written over weeks, or in two days.
For grading
Level 3
The full text
The text with all process data, the sources in full text with links, the AI classification results, and every faulty reference.
For disputes
The traffic light doesn’t judge a person or a piece of work. It shows where a closer look is worthwhile.
Your position in a misconduct case
The institution carries the burden of proof. A case can rest on circumstantial evidence: typical indicators that together point to misconduct, which the student can rebut with a plausible alternative explanation.
| AI detector only | With the Authorship Report | |
|---|---|---|
| What you have | A probability | The origin of each passage, with potential sources and timestamps |
| When AI is permitted | Of little use, can hardly differentiate | Shows how the text developed, revisions included, so you can see whether real work went into it |
| As circumstantial evidence | A single indicator, easy to challenge | Several independent indicators that corroborate each other |
| For honest students | Little chance to clear their name | Their own work is on record |
| The decision | A score easily becomes the verdict | Stays with your instructors, backed by solid data |
How courts see it: A German administrative court in Kassel treated superficial-seeming passages as just one indicator. They only established a prima facie case of cheating in combination with further evidence, such as an interview in which a student couldn’t explain their own work, or footnotes citing court decisions that don’t exist.6 These are exactly the kinds of independent indicators the Authorship Report brings together.
What we don’t promise: legally watertight proof of AI use. No vendor can deliver that. What we deliver is a transparent basis for your decision and, above all, a record of students’ own work.
30 minutes. Real-world use. Your call.
How Mentafy compares
All three check text. The difference lies in what evidence they deliver and where they work.
| Criterion | Mentafy | Turnitin | Pangram |
|---|---|---|---|
| Exact-match plagiarism check | web sources | ||
| Trace paraphrases and translations back to a source | S³, across languages | only non-English papers against English sources, add-on | |
| AI detection | as an indicator, never decisive on its own | core focus | |
| Document the writing and research process | in the programs students already use | writing process only in its own editor (Clarity, add-on) | |
| Headquarters | Germany (EU) | USA | USA |
available limited or add-on not offered, according to public product information
As of September 2026, based on publicly available product information.7 If anything has changed, please let us know.
See Mentafy in action
30 minutes with a founder: what the Authorship Report looks like, what it can support, where it stops, and how to put AI-use documentation into practice at your institution.
- Free Integrity Baseline AuditWe check anonymized papers from your department and show you what S³ and the Reference Verifier find in them.
- Direct lineMarkus Goldbach, m.goldbach@mentafy.com, +49 221 56070319
Is a colleague responsible for this? Feel free to forward this page.
30 minutes online. Pick a time right in the calendar.
or send us a message
We usually reply within 24 hours.
Sources
- TEQSA, Tertiary Education Quality and Standards Agency, Australia (2026): Assuring quality learning in a gen AI-integrated future, June 24, 2026. The agency names process traces such as timestamped edits and research behavior as evidence of learning, and explicitly calls them supporting evidence, not proof.
- Bavarian State Government: Gesetzentwurf zur Änderung des Bayerischen Hochschulinnovationsgesetzes [Draft law amending the Bavarian Higher Education Innovation Act, BayHIG], cabinet decision of June 23, 2026, Art. 84(3), sentence 1, no. 9, with explanatory memorandum. As of September 2026: government draft, not yet passed by the Bavarian state parliament.
- CHE Centre for Higher Education, Germany: Künstliche Intelligenz im Studium, die Sicht von Studierenden im Wintersemester 2025/26 [AI in higher education: the student perspective, winter semester 2025/26], published July 22, 2026; 28,584 students at 202 institutions, surveyed in business, economics, and law programs.
- Stephenson, R. and Armstrong, C. (2026): Student Generative AI Survey 2026, HEPI Report 199, March 2026; 1,054 undergraduates in the UK, surveyed in December 2025.
- Scarfe, P. et al. (2024): A real-world test of artificial intelligence infiltration of a university examinations system. PLOS ONE 19(6), e0305354.
- Administrative Court of Kassel, Germany (VG Kassel), judgments of February 25, 2026, case nos. 7 K 2134/24.KS and 7 K 2515/25.KS, not yet final.
- Vendor product information, accessed September 2026: Turnitin (Translated Matching, Clarity), Pangram (plagiarism detection).
This page is for information only and does not constitute legal advice.