On 23 June 2026, the Bavarian cabinet adopted a draft amendment to the state's Higher Education Innovation Act (Bayerisches Hochschulinnovationsgesetz, BayHIG). It contains a rule that inverts the prevailing assumption in assessment: for unsupervised written examinations, universities should no longer rule out the use of AI systems outright. Instead, they must allow AI and set requirements governing the scope and manner of documenting that use (Art. 84(3) no. 9 BayHIG, draft).
One clarification up front, because it is frequently lost in the coverage: this is a government draft, not law in force. It still requires approval by the 'Landtag'. Reports stating that Bavaria has now "permitted" AI are getting ahead of the legislative process.
From prohibition to transparency - allow AI
Until now, the default at most universities has been that AI use is prohibited unless expressly allowed. For unsupervised written work, the draft reverses that logic. A ban remains possible but becomes the exception requiring justification. It's permissible where equality of opportunity between candidates cannot otherwise be assured. The explanatory memorandum calls this a "prohibition on prohibition".
Science Minister Markus Blume ties the move explicitly to assessment, arguing that new examination formats are needed to keep pace. On that reading, the documentation requirement is not conceived as a control mechanism but as the precondition for examinations to remain meaningful at all under conditions of ubiquitous AI.
Why the direction to allow AI holds up
The combination of prohibition and after-the-fact detection was never a robust strategy. A widely cited study by Weber-Wulff et al. (2023) in the International Journal for Educational Integrity tested 14 tools for detecting AI-generated text and found them neither sufficiently accurate nor reliable, and readily defeated by simple paraphrasing. Basing an examination decision on that foundation is precarious. That's a problem courts have increasingly had to confront as well.
Bavaria is not alone in this shift. Back in March 2023, Peter Salden (Ruhr University Bochum) and Thomas Hoeren (Münster) concluded in a legal opinion commissioned by North Rhine-Westphalia that a blanket ban on AI writing tools makes no sense, and that universities should instead define the conditions of legitimate use. Internationally, Australia's regulator TEQSA has pursued the same line since 2023: detection tools cannot guarantee integrity, and only redesigned assessment offers a sustainable answer (regulation update 2026).
The real work begins in the faculties
The draft creates the obligatio and leaves its design entirely to each institution's examination regulations. On enquiry, the Bavarian science ministry confirmed that neither the legislature nor the ministry will specify the scope and manner of documentation.
That is defensible, since writing cultures differ considerably between law, computer science and art history. But it also shifts the design work to where resources are already scarce. The German Association of University Professors and Lecturers (DHV) responded accordingly: the impulses are "right and important", yet still need to be made concrete. New examination formats presuppose good supervision ratios and adequate IT support.
A serious objection
The draft is not uncontested. In an analysis on Verfassungsblog published in July 2026, Stella Elmentaler and Linus Hof argue that the "prohibition on prohibition" is incompatible with the freedom of teaching enshrined in Article 5(3) of the German Basic Law. They also point to a drafting ambiguity: the provision is phrased in permissive terms while the memorandum reads as categorical. Whether the rule survives the parliamentary process, and in what form, is therefore open.
What needs resolving now
Regardless of the outcome, the practical question already arises and it is the harder one: what does documentation look like that is robust enough to support examination decisions, yet light enough not to burden students and supervisors further? A signed self-declaration alone demonstrably does not achieve this.
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:
No comment yet, add your voice below!