Your chatbot needs to disclose it's an AI. Once, not every message.

I've seen two versions of a chatbot disclosure go wrong, and they fail in opposite directions. The first is a line appended to every single reply, some variation of “I'm an AI”, forever, on every message. The second is a notice that flashes once when the page first loads and is gone by the time anyone's actually reading the conversation. Neither one is doing what a disclosure is supposed to do.

The first is the more common instinct, because it feels safer: if it's everywhere, surely it counts. But a line that repeats on every message is exactly the kind of text people learn to stop seeing within the first three replies. It becomes wallpaper. Once something is wallpaper, whether it's technically present stops being the interesting question, because it's no longer doing the thing disclosure is for, which is making sure the person on the other end actually knows.

For the obligation itself, I covered it in a longer explainer on Article 50 a while back: people need to be told they're interacting with an AI system, unless it's obvious from context. That sentence tells you the goal. It doesn't tell you how often to say it, and I think that gap is exactly where both failure modes above come from: one side treats “how often” as “always”, the other treats it as “once, technically”. Both are answers to a question the obligation didn't actually ask.

The pattern that's worked better, in what I've built and in what I've seen elsewhere, splits the difference by splitting the signal from the content. The disclosure isn't a sentence baked into what the AI says. It's a separate notice your interface shows once, clearly, the first time someone starts a conversation, non-dismissable so it can't be missed or clicked past without registering. After that first message, it steps down to a small, persistent badge for the rest of that session, always visible, never interrupting, and a new session gets the full notice again. Said once, forgotten by no one who actually looked at the screen, and never competing with the conversation for attention after that.

The alternative most tools default to, because it's the easiest thing to implement, is rewriting the model's actual reply to append a disclosure line into the text itself, every time. It works, in the sense that the words are there. But it means every single response is partly not the model's output anymore, it's the model's output plus a fixed string glued onto the end, and the fixed string is precisely the part people's eyes learn to skip. Making the disclosure a genuine part of the interface instead of an edit to the content keeps it visible without holding the conversation hostage to it.

This is, again, how Disclarion's SDK handles it by default: the disclosure is metadata your own frontend renders, not text your model is forced to say, and a ready-made reference widget implements exactly the once-then-badge pattern above so nobody has to design that interaction from scratch. The older approach, rewriting the reply text on every message, still exists in the SDK for backward compatibility, and I'd be lying if I said I thought it was the better default. It just came first.

None of this changes what the underlying obligation is, people still need to know they're talking to an AI, and I'm not claiming the law specifies a frequency, because it doesn't. What changes is whether the disclosure still means anything by the tenth message. A notice nobody registers because it's everywhere isn't meaningfully different from one nobody sees because it flashed once and vanished. Worth checking which failure mode your own product is closer to — and if the honest answer is “we rely on the bot's name,” that's a third failure mode, covered separately in why a name alone isn't disclosure.