Letting an AI fix spelling mistakes in a text is one thing. Letting it screen applicants, assess an employee's performance or decide who should receive a benefit is something else entirely. The law calls that kind of thing high-risk. The requirements for those tools are strict, and they must be met before the tool is put into use — not afterwards.

An AI has learned from reality. If reality was skewed, it learns the skew along with it — and reproduces it in neat, matter-of-fact language where it's hard to spot. A quick check: swap out the name, the age or the place of residence in what you're looking at. If the recommendation suddenly looks different, it's worth mentioning. You don't need to be able to prove a bias statistically — you just need to notice it and say so.

You don't need to know the rules by heart. You need to be able to recognise the situation and say it out loud before the tool is put into use: "This assesses people — has it been checked?" That question is your whole role, and it's cheap insurance.