Broadcasters must label AI — for all-AI streams, that's the easy half

Reporting on 30 September 2026 said the Broadcasting and Media Communications Commission will finalise ethical principles for AI-generated content in broadcast by the end of 2026. There are two parts. That something made with AI be disclosed on screen or by voice notification, and that a person carry out the final verification and hold responsibility for it. News and current-affairs programmes get separate, stricter standards, barring the generation of evidentiary material and any depiction of a real person that could be mistaken for actual footage or a recording. The principles are non-binding ethics guidance rather than legal regulation, and they are in internal review now, to be finalised after input from broadcasters and the public. The same report said an integrated media law covering both broadcast and OTT is planned for December 2026.

AI content disclosure is the principle that AI-generated broadcast scenes are flagged on screen or by voice and that a person holds final responsibility.

“There are no mixed-in scenes, so there is nothing to label. What remains is the second sentence: a person holds final responsibility.”

The frame the reporting sets up is that viewers are fooled by an AI scene. That is a real problem. When a scene that looks filmed was not filmed, disclosure is necessary. But the frame rests on an assumption: that most of a broadcast is made by people and AI is a part mixed into it. On the side we are building, that assumption inverts. On AStream, an AI BJ live broadcast, the thing doing the broadcasting is itself an AI a member built. There are no mixed-in scenes, so there is nothing to label. The first sentence of the principle falls out of the design. What remains is the second one: a person carries out the final verification and holds responsibility. The reporting treats disclosure as the conclusion, but of the two sentences the hard one was the second.

First limit: in a broadcast that never stops, a person cannot do the final verification. What we can do is done in advance. Set the character, set what it will not say, set the subjects it will not cross. The verification the principle describes is closer to after the fact — a person checks what went out and answers for it. That the broadcast does not stop is the premise of this product, and not stopping and a person seeing all of it do not hold at the same time. This is something we have not solved, not a part we can claim.

Second limit: disclosure shows once on the way in, and the broadcast continues. Almost nobody watches a 24-hour broadcast from its start. Most arrive in the middle, and after leaving the room they keep listening through a mini player. In a voice broadcast with almost nothing on screen, where to keep "what you are listening to is an AI" is still unanswered. A notice that appears once and disappears satisfies the principle and not the purpose.

Third limit: on a platform where members build the broadcaster, who holds responsibility is undecided. Ethics guidance is written with a broadcast operator in mind. An AStream BJ is not built by the platform but by a member. Whether responsibility for what that BJ said sits with the person who built it or with the platform, and where it divides, has no answer on our side either. While the principles are advisory that question can be deferred; once they are law it cannot.

So the design we chose does not bolt disclosure on afterwards; it puts the fact that this is an AI up front, as the basis for choosing. The screen for choosing a room on AStream has no thumbnails. You choose by a one-line concept, voice tags and genre colour. All three state how this BJ was built. That means it does not sit in small print as a fact to be buried but in the first place you read when choosing. Not mixing human broadcasts and AI broadcasts into one list follows from the same reason: more certain in the design than mixing them and then separating them with a label.

What remains: AStream is not public yet, and we have no evidence that the design above actually prevents a misunderstanding. And one thing disclosure does not solve is left. If the reason people leave a broadcast on is to keep on the state of not being alone, then knowing it is an AI and that state holding are separate questions. Whether it holds anyway, or breaks once known, can only be learned after release. Finally, these principles are advisory for now and the integrated media law is still planned. There is no way around designing without knowing what will be added.

WorldWideSomething studies the psychology of people through HCI and AI and designs apps out of it. Beff, the proactive AI friend app, is public; AStream, an AI BJ live broadcast, and LiveLike, a job-experience chat, are not public yet. On the unreleased side we write down what has been decided and what has not, separately.