Because of new EU transparency rules, AI providers are adding machine-readable marks to generated content. For text, this works through a statistical pattern in word choice that can only be detected with a matching key.
Its evidential value is limited. A match merely indicates that a particular model was probably involved, perhaps only for translation. Rephrased or short passages escape detection, and the absence of a match does not prove human authorship. A public detection tool is not yet available.
We advise companies to focus on approval: define what AI text may be used for, who checks it before it is sent, and that amounts, dates and recipients come from their own systems. Such marks are not suitable for assessing job applications.