Ethics stopped being a conference panel and became paperwork with teeth. A clear-eyed tour of what is actually regulated, who has to care, and what a small business should know by name.
Practical AI ethics has three working questions, and none of them are philosophical. Who does this get wrong for, that is bias. Can you show why it said that, that is transparency. Who answers when it is wrong, that is governance. Regulators have now written all three down with penalties attached, and the writing-down is the story of this year more than anything the models themselves did.
What the EU AI Act actually does
The Act sorts AI uses by stakes. A few practices are banned outright, social scoring by authorities and manipulative systems among them. High-risk uses carry real obligations: hiring tools, credit scoring, essential services, critical infrastructure. For those you need documented risk management, human oversight, attention to data quality, logs, and conformity paperwork before the system meets the public. Limited-risk uses, chatbots and generated content among them, mostly carry transparency duties: tell people they are talking to a machine. Ordinary productivity tools sit at minimal burden. The logic is simple once you see it. The higher the stakes for a person's life or rights, the more you have to prove about how careful you were.
It matters outside Europe for the reason US companies care about California privacy law. Sell into Europe, serve European users, and the rules travel with the product.
Who must care, including the shop with eleven staff
If AI screens your job applicants, prices anything like credit, or gates access to something people need, you are inside some regime almost everywhere you look, and "the AI decided" is not a defence anyone accepts. For everyone else the duty is light and cheap: label what's synthetic, keep a human deciding anything that reaches a customer, keep the record of who approved what. Bias, in practice, is mostly an unexamined default: the hiring tool trained on ten years of resumes from the kind of person who got hired before, or the voice system that handles the accent it was tested on and fumbles the others. Nobody has to be evil for this to happen. Nobody has to be an expert either. It takes a person willing to ask who this answer is wrong for, once, before launch.
What we do here, plainly
AI Universe City teaches these tools by using them the way the paperwork describes. AI assistants handle lessons and answers, humans keep oversight and final say, every outward-facing action gets a person's yes before it happens, and the rules are published in the community rather than hidden in a settings menu. That is not a disclaimer we hope you never read. It is the syllabus.