Should users know they're talking to AI, and should capabilities be documented
Across nearly every AI regulatory framework under discussion globally, a small set of policy debates keeps recurring, largely because they don't have technology-specific answers, they're genuinely contested questions about what a fair and informed interaction looks like. The first is transparency and disclosure. One strand of this is disclosure to end users: should a person interacting with a chatbot or automated customer service system be told they're talking to AI rather than a human? Several frameworks, including transparency provisions commonly associated with the EU AI Act's limited-risk tier, lean toward yes for at least some categories of interaction, on the theory that people make different judgments about trust and scrutiny when they know they're dealing with a machine.
A second strand is disclosure about the system itself, sometimes called model or capability documentation: what was a model trained on, what are its known limitations, what evaluation results exist for safety-relevant behaviors. This kind of disclosure obligation shows up in different forms depending on jurisdiction and tends to apply more heavily to larger or more capable models than to small narrow-purpose ones, though exactly where that line gets drawn is one of the genuinely unsettled details across frameworks currently being finalized.