Disclosure is now a feature

On 2 August 2026, the transparency obligations in Article 50 of the EU AI Act began to apply. If a person is interacting with an AI system, they have to be told — clearly, at the latest at the first interaction — unless it is obvious from the circumstances. Synthetic image, audio and video content has to be marked in a machine-readable way. Deepfakes have to be disclosed. None of this requires a legal department; it requires an honest interface.

The same month, plenty of commentary described the AI Act as delayed. That is only half right. The Digital Omnibus deferred the high-risk regime, but the Article 50 transparency duties were not part of that deferral. A support assistant, a booking agent, a voice helper — the kind of thing a small business actually deploys — sits in the transparency conversation now, not in 2027.

What a small site actually has to do

Three things, in plain terms. Say that the assistant is an AI, in the interface rather than a footnote. Say when a demo is a simulation rather than a live system. And if you publish generated imagery or audio, label it and keep the provenance where a tool can read it. The first two are design decisions. The third is a file-format decision.

The US picture is narrower but not empty. California's SB 243, in force since 1 January 2026, targets companion chatbots — systems built for sustained, human-like personal interaction — and requires a clear AI disclosure plus safety protocols. It expressly leaves transactional customer-service bots outside its scope. That exemption is not a licence to be vague: the FTC's rule on fake reviews and testimonials has been in force since October 2024, and its Operation AI Comply enforcement sweep went after businesses that made deceptive claims about what AI could do. A chatbot that pretends to be a person, or a case study that pretends to be a client, is a claims problem before it is an AI problem.

Disclosure as an interface decision

The cheapest way to comply is also the best way to build trust: put the disclosure where the decision happens. A small 'AI' badge in the chat header. A first message that says what the assistant can and cannot do. A 'Simulated example' label that stays visible for the whole demo, not just the first frame. A transcript so a reader can check what was said without replaying an animation.

What does not work is the paragraph at the bottom of the page. Disclosure that arrives after the interaction is not disclosure; it is indemnity. Readers notice the difference, and so do regulators.

The commercial upside

We have watched visitors relax when a demo says 'this is simulated'. The label removes the question 'am I being sold something fake?' and lets them look at the workflow instead. The same is true of an assistant that opens with 'I'm an AI; here's what I can do'. Expectation-setting is not a cost. It is the reason someone trusts the handover to a human when the assistant reaches its limit.

The compliance deadline is a useful forcing function, but it is the weaker argument. Build the label because it makes the product easier to trust, and the filing gets easier as a side effect.