In June 2026, a resource commissioned by Australia's higher education regulator TEQSA set out how learning can be assured in an AI-integrated future. In the same month, the Bavarian cabinet adopted a draft law on AI use at universities. Two agencies, two continents, two fundamentally different methods for an assessment reform AI, but yet a strikingly similar conclusion: the finished document alone no longer carries the burden of proof.

That convergence is the genuinely interesting part. When two independent systems using entirely different instruments arrive at the same result, the result probably lies in the subject matter rather than in the method.

The Australian route runs through sector practice

TEQSA deliberately works without statutory force. It commissions experts from across the sector, convenes expert forums and publishes the results as resources. Both papers discussed here are explicitly non-binding Guidance Notes and, by their own statement, are not intended to be prescriptive. Its September 2025 resource sets out three pathways institutions can take to assure learning and contains a remarkably clear rejection of the detection logic: because establishing AI use with certainty is currently all but impossible, alternative approaches are needed. Rather than investing primarily in detection, the guidance argues, assessment itself should be redesigned.

The June 2026 follow-up, led by Jason M Lodge, takes a further conceptual step and introduces the category of learning process evidence. It explicitly means more than a reflective journal: the traces that accumulate anyway as students plan, revise, seek feedback, consult resources and make decisions. This includes their interactions with AI systems. Among the examples given are time-stamped edits, patterns of resource use and sequences of problem-solving actions.

So a regulator has named the instrument. What it cannot do is make it binding.

The Bavarian route runs through examination law

In Bavaria, the sequence is exactly reversed. The draft law of 23 June 2026 (we discussed it here) inverts the default rule: for unsupervised written examinations, universities should no longer exclude AI outright, but must instead set requirements on the scope and manner of documenting its use. Once the Landtag approves it, that carries full legal force.

What the draft does not do is define what such documentation consists of. That is left entirely to each institution's examination regulations. Bavaria therefore makes binding what it does not spell out, which is the exact mirror image of the Australian case.

Where the two meet - assessment reform AI

Despite opposing methods, both approaches share three assumptions. First, detection does not hold. Second, the product alone no longer suffices as evidence — process evidence sits alongside it, explicitly not in its place. Third, and consequently, assessment needs data that arises while the work is being done, not only afterwards.

The Australian university RMIT captured the June paper succinctly in a widely read commentary: the artefact on its own may no longer carry the evidentiary weight, and assuring learning means examining the journey a student takes to reach it.

The motives differ markedly. Australia argues pedagogically, from the question of how learning outcomes can still be evidenced at all. Bavaria argues from examination law, along the lines of equality of opportunity and legal certainty. That two such different lines of reasoning arrive at the same point is the strongest available argument that this is not a passing fashion.

And where they diverge

The difference fits into a single sentence: Bavaria has force without specificity, TEQSA has specificity without force.

A second difference matters just as much for institutions. TEQSA states a limitation openly that is absent from the Bavarian draft: thoughtfully designed, such data offers "partial windows" into students' regulatory activity, without claiming to capture learning in its entirety. Process data yields indications, not complete records. Anyone introducing it should say so from the outset.

The resource goes further still. Such data collection, it argues, must prioritise student agency, privacy and transparency, and must avoid surveillance approaches — because those undermine the very self-regulation they set out to foster. Nor is it a foolproof way of assuring that learning has occurred; secure forms of assessment remain necessary for now.

TEQSA also names an objection the Bavarian draft does not address: permitting AI for part of a task effectively permits it for the whole task in most cases. Which is another reason why permission alone is not enough — it needs a counterpart on the evidence side.

What follows for institutions

The question is no longer whether documentation happens. It is what counts as evidence — and that decision is made neither in Canberra nor in Munich, but in examination boards. Whoever makes it has to weigh informativeness against effort, traceability against data protection, and assessment security against students' trust.

What documentation can look like that withstands those trade-offs is the subject of part three of this series.

After we discuss and shed some detailed light on the regulations in the 2nd article of the series, we will eventually point out, how a practical solution looks like:

• 1/3 Allow AI, Document Its Use: What Bavaria’s Draft Law Means for Assessment
• 3/3 Process Documentation in Practice: A New Assessment Format for the AI Era