On 2 August, a rule came into force in Europe that most business owners have not read and never will. Article 50 of the EU AI Act. Stripped of the legal language, it says three things. If your AI talks to a person, that person has to be told they are talking to an AI. If your AI generates audio, images, video or text, the output has to carry a machine readable marker saying a machine made it. If you publish a deepfake, you have to label it clearly, on first exposure, not in a footnote.

That is the whole thing.

I have watched the reaction for two weeks and nearly all of it has been about compliance. Law firm briefings. Readiness checklists. Webinars on scope and penalties and whether a business in Bangkok selling to a customer in Berlin is caught by it. For the record, it is, if the output is aimed at people in the EU.

Almost nobody is asking the more interesting question.

Why did we need a law for this at all?

What the rule actually asks for

Chatbots, voice agents and avatars that speak directly to people now have to make it obvious at the start of the interaction that they are not human. Not buried on page four of a privacy policy. At the start, in a way a normal person would notice.

Synthetic content has to be marked in a machine readable format so it can be detected as artificially generated.

Deepfakes have to be labelled visibly or audibly the moment someone encounters them.

There are carve outs, and the enforcement picture is still forming, and the guidance will be argued over for a year. But the spirit of it is not complicated. People get to know when they are dealing with a machine.

The part that should make this industry uncomfortable

A rule only gets written when enough people are doing the opposite.

Nobody legislates the obvious. There is no law requiring a shoe shop to admit it sells shoes. Article 50 exists because a meaningful number of businesses have been deploying AI in ways that quietly depend on the customer not knowing. The endlessly patient agent called Sarah who replies in eleven seconds at three in the morning. The product review that reads like a real customer wrote it. The voice on the phone that breathes in the right places and says "sorry, could you repeat that" to seem more human, not because it failed to hear you.

That has been the design brief on a lot of projects. Make it pass. And the market rewarded it, because "indistinguishable from a human" sounded like the highest compliment you could pay a system.

I have never accepted that as a goal, and now there is a legal reason not to. The legal reason is the least interesting one.

Disclosure makes AI work better, not worse

This is the part that gets lost in the compliance panic, and it is the only part I would argue about at length.

Telling people they are talking to a machine improves the interaction. It does not damage it.

When someone knows they are speaking to a system, their behaviour changes immediately and helpfully. They stop performing politeness. They ask direct questions instead of building up to them. They give the information the system needs in the order the system needs it. They stop apologising for asking twice. And crucially, they escalate honestly when they want a person, instead of getting quietly annoyed and abandoning the conversation.

Undisclosed AI has a very specific failure mode, and it is not embarrassment. It is betrayal. A customer who suspects halfway through a conversation that they have been talking to a bot does not simply update their expectations. They go back through the whole exchange and reinterpret it. The reassurance they were given becomes worthless. The empathy becomes offensive. They do not think "ah, clever system". They think "you let me talk to a machine about something that mattered to me and you did not tell me".

You cannot recover from that with a better model. It is not a quality problem.

The businesses I have watched get real value out of AI over the past year did not win by hiding it. They won by putting one sentence at the top of the conversation. Something like: you are chatting with our booking assistant, it can check availability, book, reschedule and answer questions on pricing, and it will pass you to a person for anything else.

That sentence does more work than another three weeks of tuning. It sets scope. It sets expectations. It gives the customer permission to ask for a human without feeling rude, which means the handover happens early, while the customer is still calm, rather than late, when they are furious.

Disclosure is not a tax on the system. It is part of the system.

The bit that changed while everyone was arguing about chatbots

Here is the development that got much less attention than it deserved, and it landed on the same day.

On 2 August, Anthropic began embedding an invisible watermark in every piece of text Claude generates. Worldwide. No opt out. They confirmed the details publicly on 11 August. The approach is based on Google DeepMind's SynthID-Text, and it works by biasing the model's word choices during generation so that a statistical pattern is detectable across enough text. Files carry C2PA provenance metadata on top of that. Anthropic was one of roughly 190 signatories to the EU Code of Practice on Transparency of AI-Generated Content in July.

The important detail for anyone using AI in their business is this. The watermark is in the word choices, not in the file. It survives copy and paste. Pasting into a document, an email, a CMS or a LinkedIn box does not remove it.

The reaction online was not warm. A lot of people who use these tools every day discovered that the output they had been treating as theirs now carries a marker they did not ask for and cannot turn off.

I understand the frustration. I do not really share it.

If I am going to argue that businesses should disclose when a machine is involved, I do not get to be annoyed when the machine discloses itself. Those are the same principle. You cannot want transparency to apply to your suppliers and not to you.

There are honest limitations, and it is worth being clear about them rather than pretending this is settled. The signal needs a reasonable volume of text to be detectable, so a short social post carries far less than a long article. Running the text through a different model to rewrite it substantially degrades the original signal, although you may simply be swapping one provider's marker for another. And detection requires the provider's own detector, which is not something a normal platform is running today.

So the practical exposure right now is small. The direction of travel is not.

While I am here: these are my ideas and thoughts, then drafted with AI assistance and then again rewritten by me. It is probably watermarked. I would rather tell you that than have you find out.

What this changes if you run a small business

In terms of work, almost nothing.

If you already tell people when they are talking to your assistant, you are finished. Go and do something useful with your afternoon.

If you do not, the fix is a sentence and about ten minutes.

The genuinely uncomfortable question is a different one, and it is worth sitting with. If disclosure would meaningfully hurt your results, what does that tell you about what you built?

Because there is a version of AI deployment where the disclosure is a non event. The system is useful, it does a defined job well, the customer gets an answer faster than they would have from a person, and nobody minds at all. I have seen that many times. The disclosure line gets read and ignored, the way people ignore the word "recorded" at the start of a support call.

And there is a version where the disclosure is fatal. Where the entire value came from the customer believing they had been given human attention. In that version the AI was never doing the work. The illusion was doing the work.

The first version survives Article 50 without noticing it. The second version was always going to break, and Europe has simply set the date.

Where I think this goes next

My guess, and it is a guess, is that within two years disclosure stops being a legal question and becomes a marketing one. The label moves from the small print to the sales page. Businesses will advertise which parts of their operation are handled by a machine and which are handled by a person, because customers will start asking, and the ones who answer clearly will look confident while the ones who dodge will look shifty.

We are already most of the way there in other industries. Nobody thinks less of a bank for having an automated balance check. They think less of a bank that pretends the automated balance check is a relationship manager.

The AI industry spent three years chasing a benchmark nobody actually wanted, which was fooling people. The businesses that quietly ignored that benchmark and built systems that were obviously machines, but good ones, are in a considerably better position this month than the ones that did not.

The line I keep coming back to

If your AI only works when the customer does not know it is AI, you have not built an AI system. You have built a disguise. And disguises have a shelf life, whether or not a regulator sets one.

Europe has now set one anyway.

So here is what I actually want to know, because I think reasonable people land in different places on this.

If a customer found out tomorrow that a machine handled their last enquiry with your business, would they be annoyed or relieved?

And if the honest answer is annoyed, is that a problem with the disclosure, or a problem with the system?