What the EU AI Act Looks Like — from the Frontlines of Teaching AI in Germany

Sunday Reflection — 2026-09-27

Thursday morning, the whiteboard was covered with hand-written model architectures and threat assessments. We were reviewing an assignment: build a sentiment classifier using Python, document the risks under the new EU AI Act template, and present the compliance reasoning. Half the cohort stared at the risk checklist as if it might bite back. The official guidance is supposed to clarify things. In reality, it reads like the user manual to a tool you haven’t bought yet. One keen student asked if their ‘Explainability’ section needed a separate audit log. I had to admit: nobody agrees on that—least of all the regulators themselves.

mermaid flowchart LR Start[Plan AI lesson] --> A[Check EU AI Act rules] A --> B{Is guidance clear?} B -- Yes --> C[Document risks, teach as planned] B -- No --> D[Ask other educators] D --> E{Consensus?} E -- Yes --> C E -- No --> F[Make best guess, hope]

A decade ago, curriculum risk meant “Don’t let students wire an Arduino to mains voltage,” maybe “Remind them not to copy code.” Now, running a toy language model in a cloud sandbox sparks debate: have we assessed bias risk? Have we documented the training data? Is the cloud provider fully EU-compliant this month? Paperwork takes more time than coding. My own Army days come back in flashes—laying plastic explosives under frozen bridges at dawn, yes, but mostly filling out paperwork to justify it. Bureaucracy kept people alive, but you always knew the forms only caught part of the truth.

There’s value in the law’s caution. We don’t want reckless AI rolled out by beginners. But these first weeks on the ground show how easily caution turns into paralysis. Students aren’t learning risk management, they’re learning civil service.

This will settle, eventually, as examples accumulate and the blurry lines sharpen up. For now, every educator I know is improvising. If you build a process where uncertainty is the norm, expect professionals who hedge—and lose sight of real learning. The EU says it wants responsible AI. Let’s see if it gets responsible builders, or just risk-averse ones.