New Software Automatically Checks If Your AI Breaks Europe's Law

Europe's AI Act is now the world's strictest set of rules for artificial intelligence. Companies using "high-risk" AI — say, a chatbot giving financial advice, or software screening job applicants — must document how it works, test it for bias, and keep humans in the loop. The problem: those rules were written with older AI in mind, the kind that sorts spam or scores loans. They leave seven big gaps when applied to modern generative AI, which writes, chats, and produces different answers every time you ask.

A pair of researchers say they've closed those gaps with code. Their system, called Governance-as-Code, turns the law's fuzzy language into 43 specific tests that a computer can run automatically. Vague phrases like "appropriate levels of accuracy" become declared numbers you can check. The tests run inside the same automated pipeline companies already use to test software before releasing it, and each result is tagged to the exact article of the law it satisfies — so the paperwork writes itself.

They tested it on two real deployments: a high-risk advisory chatbot and a simpler content generator. The software reproduced every finding from a human expert audit, including three violations serious enough to trigger penalties, while cutting the audit labor by roughly 75%. It also fixes a common misunderstanding about who owes what: companies that merely use an AI model only need to check the training-data summary their supplier provides — not document every piece of data they never had.

The catch: this is a research prototype, not an approved compliance tool, and no software can guarantee you're legal. Lawyers still sign off. But it points to where things are heading — AI rules turned into code, checked continuously like any other software, instead of reviewed once a year in a binder.