The first two parts of this series looked at two regulators approaching AI and assessment from very different directions. Bavaria's proposed higher-education reform would require universities to define how the use of AI is documented in unsupervised written examinations. TEQSA's recent work on assessment reform describes a broader concept of learning process evidence: the traces created while students plan, research, revise, seek feedback and make decisions. Neither approach prescribes a single technical format. The practical question is therefore the same: what should process documentation actually look like?

That is what we work on. This third part shows how process documentation can become part of an assessment format without turning students into data-entry clerks or examiners into log analysts.

What is process documentation in assessment?

Process documentation means adding evidence of how a piece of work was produced to the final work itself. Instead of assessing only the submitted essay, report or thesis, an institution can also consider the development of the text: writing sessions, revisions, research activity, source use and relevant reflections.

This matters because generative AI has changed the evidentiary problem (also see the UNESCO's global guidance on genAI). A polished final document can look convincing even when very little of the underlying work was done by the person submitting it. At the same time, an unusual writing pattern or an AI-classification score is not, by itself, proof of misconduct. AI-Text-Watermarks seem to have a revival, but as humanizers are already widely used to disguise AI-Text from AI-Detectors, the very same approach also destroys the AI-Text-Watermarks. Process documentation adds another layer of evidence. One that can support a more informed human judgment.

What checking after submission can and can not do

Many institutions still have to assess work after submission, with no access to a complete writing history or research timeline. In that situation, post-hoc integrity checks remain useful.

A plagiarism check compares submitted text with web and publisher sources. Mentafy's Semantic Source Search (S³) adds a semantic layer, comparing meaning, structure and chains of argument rather than relying only on identical wording. This can help surface paraphrased or translated source use that conventional word matching may miss. Reference Verifier checks citations and bibliography entries against external academic data sources and can flag problems such as inconsistent references or false DOIs.

AI classification provides another signal by assessing writing patterns associated with AI-generated or AI-revised text. These tools are useful precisely because they answer different questions: Does the text resemble existing sources? Are the cited sources real and consistent? Are parts of the text unusual or consistent with AI-assisted writing?

But there is a limit that should not be talked down: post-hoc analysis judges the finished product. If someone uses AI to produce a complete draft and then carefully revises it, the final document may contain few reliable traces of how it was created. Post-hoc checking can therefore provide important indications, but it cannot reconstruct a missing writing process.

The second layer: the path to the text

Process documentation adds that missing layer. It asks not only what was submitted? but also how did this text develop?

Writing process analysis examines the evolution of a document. Mentafy's analysis looks at signals such as writing duration, writing sessions, rhythm, revisions and the size and timing of text insertions. It can classify sections according to patterns such as human-written, copy-and-paste, copy-typed, AI-revised or paraphrased, and presents the findings in the Authorship Report.

The Research Recorder adds the research side of the picture. It records materials that students explicitly save or connect to a project, including when they engage with them and how they annotate them. Its browser plugin can also save explicitly selected pages, including AI chats, to the project's research record. It does not track general browsing.

The Writing Journal brings writing and research activity together in a chronological project record. Automatic project activity can be supplemented with the student's own written, audio or visual notes and reflections. The result is a day-by-day view of how the project developed. This further prevents a malpractice of a manually fully reconstructed diary created just before submission.

What matters just as much is what does not happen: no keylogging and no screen recording. The Writing Journal is built from project-related events such as saved document versions, uploaded files, research activity and notes, rather than general device or browsing activity. This distinction also reflects the European Commission's updated guidance on the ethical use of AI and data in education

This distinction is increasingly important. TEQSA's June 2026 work explicitly describes learning process evidence as including time-stamped edits, patterns of resource use, sequences of problem-solving actions and interactions with resources or AI. At the same timeit is also stressing student agency, privacy and transparency and warning against surveillance approaches.

Why this is a new assessment format - and still the old one

For students, a well-designed process-documentation format can leave the core writing task almost unchanged. They continue to work in familiar tools such as Word or Google Docs and in their usual cloud environment. Much of the relevant process evidence is collected automatically rather than requiring students to maintain a separate log. Students can add their own notes or reflections where these are part of the assessment design.

What changes is what counts as evidence: not only the final text, but also the path that produced it.

That is the core of what we mean by an updated assessment format. The essay does not disappear. The research paper does not disappear. What changes is that the final product is accompanied by evidence of the process behind it. That's precisely the kind of evidence that becomes more valuable when AI can generate polished text on demand.

The task can stay familiar. The marking criteria can stay familiar. What becomes stronger is the evidentiary basis for judging authorship, learning and academic integrity.

The focus shifts to personal contribution

And here lies the real difference from pure detection logic. The goal is not simply to look for where someone may have cheated. The goal is to make visible what the student actually contributed.

That distinction matters for both sides. A student who wrote the work can have evidence of that effort instead of having to defend themselves against an unexplained suspicion. A supervisor can see whether meaningful research and writing activity took place over time. And where a project appears to have stalled, process information can provide an opportunity for intervention while there is still time to help.

The feedback from our users suggests that in most cases for around 80–90% of students, the deeper in-document analysis is not necessary after the initial Authorship Report has been reviewed. That figure is a statement about Mentafy's product experience and should be understood as such, rather than as a universal benchmark for every institution or assessment format.

In other words, process documentation does not have to mean that an examiner reads every student's entire writing history. The purpose is to make additional evidence available when it is useful and to make a focused review possible when questions arise.

What process documentation does not prove

Process evidence is evidence, not a verdict. Writing patterns can indicate that a passage deserves attention, but they do not automatically establish misconduct. A student may write non-linearly, work in a second language, use assistive technology or legitimately paste and adapt material from a source. Those circumstances can produce patterns that differ from a simple "normal" writing trajectory.

There are also technical limits. Text that was typed out from another source can produce an ordinary-looking writing pattern, for example. And process data is necessarily incomplete: it provides a window into how work developed, not a perfect recording of everything a student thought or did. TEQSA makes this limitation explicit in its current guidance, describing process evidence as a partial window into learning and stressing that it is not a foolproof assurance mechanism.

That is why the right principle is simple: documentation is the basis for a decision, not the decision itself.

From a final document to an evidence-based assessment

The larger shift is therefore not from "essay" to "surveillance". It is from product-only assessment to evidence-based assessment.

That is also where the current developments in Bavaria and Australia become interesting. Bavaria's June 2026 draft law takes a legal route: for unsupervised written examinations, it proposes that AI use should generally not be excluded and that examination regulations should define the scope and manner of documenting AI use. TEQSA takes a sector-guidance route: its recent work identifies learning process evidence as a component of assessment reform, while emphasising that such evidence must respect privacy, transparency and student agency.

They are not the same policy. Bavaria's provision is still a draft and is more specific about the legal obligation; TEQSA's resources are guidance rather than binding rules. But both point toward the same practical challenge: if the final document is no longer sufficient on its own, institutions need a workable way to evidence the process behind it.

---

Ready for tomorrow's assessment requirements?

If you are considering how your examination regulations could define the scope and manner of process documentation — or how writing process analysis, research evidence and post-submission checks could work together at your institution: see the Mentafy toolset at a glance, or get in touch directly.