What actually changed on 2 August?

The transparency obligations in Article 50 of the EU AI Act became applicable on 2 August 2026. In plain English, three things now have to be said out loud that previously didn't: an AI system people can converse with must tell them it's an AI; images, audio and video that are AI-generated or meaningfully altered must be marked; and AI-written text published to inform the public on matters of public interest must be disclosed, unless a human took editorial responsibility for it.

That's the whole user-facing change. It is smaller than the coverage suggests and lands on more people than the coverage suggests, which is an unusual combination and worth ten minutes of your attention.

The wording matters on the first one. Article 50 requires the information be given "in a clear and distinguishable manner, at the latest at the time of the first interaction or exposure." Not in your terms. Not when asked. Before or at the start — which for a chatbot means the greeting, and for a DM automation means the first message.

Who does this actually land on?

Here is the part most summaries get wrong, and it's the reason this is a small-business story rather than a big-tech one.

The Act splits duties between providers — the people who build and place an AI system on the market — and deployers — the people who use one in the course of their business. Article 50 puts obligations on both. An organisation with no high-risk AI anywhere near it can still have real duties here simply because it runs a customer-facing chatbot, publishes marketing creative made with generative AI, or uses a system that produces synthetic imagery.

The practical consequence is blunt: your vendor cannot comply on your behalf. If you bolt a chatbot onto your website using someone else's platform, the platform's obligations are the platform's. The disclosure that appears — or doesn't — in your customer's chat window is yours. The same logic that makes an agent a supplier you should give a one-page contract to makes it a supplier whose paperwork does not cover you.

On territory: scope follows where your users are, not where you are. A UK or Kenyan business whose chatbot answers anyone who lands on the page is worth checking properly rather than assuming out of. That check is a job for someone qualified, and this post is not that.

Not legal advice — this is a plain-English summary of a long instrument, and your position depends on facts this post doesn't know.

What is it not?

Worth being equally precise about the things that are not true, because a fair amount of August's commentary implied them.

It is not a ban on using AI in your business. It is not a licence or a registration — there is no form to file and no body to notify for these obligations. It is not retroactive: content generated before 2 August 2026 does not have to be labelled after the fact.

And the transitional period is narrower than most write-ups made it sound. The extension to 2 December 2026 covers the machine-readable marking and detection duty for providers of generative systems already on the market. The user-facing disclosures — the chatbot telling someone it's an AI, the deepfake label — applied on the 2nd. If you have been telling yourself you have until December, check which obligation you were thinking of.

Two more facts to hold. The European Commission's AI Office has published a voluntary Code of Practice on Transparency of AI-Generated Content, which exists to give providers a recognised way to demonstrate compliance rather than to add a burden. And the penalty ceiling for non-compliance is up to €15 million or 3% of worldwide annual turnover, whichever is higher — a number that tells you the category of obligation, not what a sole trader should expect.

So what do you actually do this week?

For most one-person businesses: about an hour, spread over four things.

• Your chatbot's greeting. One sentence at the top: "This chat is answered by an AI assistant." If it hands off to you, say when.
• Your generated media. Anything AI made or meaningfully altered — hero images, ad creative, social graphics — gets a label. A caption line is enough for most of it.
• Your published writing. If you publish anything news-like or in the public interest and a human didn't review it, disclose. If a human did take editorial responsibility, that's the exemption — but "a human took responsibility" has to be true, not aspirational.
• Your policy page. One place a cautious client can look before they ask.

None of that is a project. It is configuration, and it is the fastest half of the week's work — which is why tomorrow's post is the 20-minute audit that finds every surface before you start writing lines.

The strategic read is simpler than the legal one. This is the first time a regulator has put a date on the middle step of vet → disclose → verify. The businesses that already had a written trust boundary have almost nothing to do. Everyone else has an hour.

For the checklists we use to run this — free, 5-min read: 7 AI Prompts That Run a One-Person Business.

Sources (verified 9 Aug 2026): Article 50, EU AI Act (artificialintelligenceact.eu/article/50/); European Commission, Transparency obligations under Article 50 FAQ and Code of Practice on Transparency of AI-generated Content (digital-strategy.ec.europa.eu); Cooley, EU AI Act: Transparency Obligations Take Effect 2 August 2026.

Related Decodes

• Say It Out Loud: The Solo Founder's Guide to AI Disclosure — the week's hub: the audit, the three sentences and the one-page policy in full.
• The 20-Minute Disclosure Audit — find every surface before you write a single line.
• Agents Got File Access in July — the launches that put AI in front of your customers in the first place.

AI the Brief (@aithebrief) — AI advice you can act on and trust, for time-starved founders.

TL;DR: Article 50 of the EU AI Act became applicable on 2 August 2026, requiring that conversational AI tells people it's an AI at the first interaction, that AI-generated or altered images, audio and video are marked, and that AI-written text published to inform the public on matters of public interest is disclosed unless a human took editorial responsibility. The obligations fall on deployers — businesses using AI — as well as on providers, so a vendor's compliance does not cover yours, and scope follows where your users are rather than where you are. It is not a ban, a licence or a registration, and it is not retroactive; the transition to 2 December 2026 covers only machine-readable marking by providers of systems already on the market. For most one-person businesses the work is roughly an hour: a chatbot greeting line, labels on generated media, disclosure on unreviewed published writing, and a policy page.