AI detection has improved sharply on one axis: it rarely accuses innocent writers now. On the axis that decides cases, catching text that was deliberately disguised, far less has changed. And even a perfect detector would answer the wrong question.

Does generative AI harm learning or support it? It depends on the task, the stage of learning and what the assessment is meant to demonstrate. AIAS 2.1 turns that complexity into a practical assessment-design framework.

Since August 2026, new Claude models have been weaving an invisible watermark into their text. For universities, that sounds like a useful new signal for AI detection. But a watermark answers a different question from the one examination boards actually face, and Anthropic says so itself.

Bavaria requires documentation but does not say what it consists of. TEQSA describes what process evidence is but cannot make it binding. Both aim to make documentation practically workable. Here is what that looks like and why almost nothing changes for students.