Facing the EU AI Act: Real Changes on the Ground in IT Training and Consulting

Sunday Reflection — 2026-09-06

Last Thursday I stopped a Masterschool student halfway through his Cifar-10 demo. He hadn’t added a single line about dataset origin or preprocessing. This used to be normal—AI classes were about getting code to run and maybe tuning a few hyperparameters. Now, before models finish training, I’m checking that students write down which data licence they’re using and what biases they know about. Some complain it feels like extra paperwork for ‘throwaway’ exercises. But I see the change: two years ago, nobody would ask whether a test set had real faces or synthetic ones. As of this semester, they have to. That switch didn’t happen because of some theory; it’s the EU AI Act hitting the ground.

mermaid flowchart TD Start[Classroom Idea] Datasets[Dataset Chosen] Document[Dataset + Model Notebook] RiskCheck[Teacher Reviews for Risks] Approve[Approved] Fix[Send Back for More Detail] Start --> Datasets --> Document --> RiskCheck RiskCheck -->|Complete| Approve RiskCheck -->|Missing Info| Fix Fix --> Document

Clients are reacting too, just with different nerves. A German mid-sized energy company I work with this spring wanted a new project approval process. Their CTO worried every pilot chatbot might end up dragged into a compliance audit, so now the process starts with a checklist that looks more like a customs form than a product spec. The shift’s visible in the friction it causes—backend devs ask why yet another document is needed for a prototype, and product owners start making lists of ‘sensitive AI predictions’ before a single commit. It reminds me of Army unit inspections, lining up kit before we moved out—tedious, but sometimes the only reason a forgotten tool was found. British colleagues, by contrast, mostly shrug or stall—‘let’s see if anyone actually checks’ might as well be a national motto. Here in Saxony-Anhalt, the pendulum swings the other way: ‘better an extra file than a visit from the Data Protection Authority.’

Best advice I can give: learn to write down what your models do—and what they can’t. It’s a habit. Annoying at first, but you only need to miss one obvious problem to make it worth the trouble.