On 2 August the transparency obligations of Article 50 of the European AI Act started to apply. The simplest of them says that if a system is built to talk to you, you have to know you are talking to a machine.
Saelyx is exactly that. So this article covers what the rule asks for, what we were doing on 2 August —not much— and what we have changed since.
01First, a date almost everyone gets wrong
Regulation (EU) 2024/1689 has been in force since 1 August 2024. What happened on 2 August 2026 is not that it came into force: it is that its general obligations started to apply, Article 50 among them.
It sounds like a lawyer's quibble and it isn't. A text can be in force for two years without anyone having to do anything yet, and those two years are precisely the window to get ready. Whoever confused the two had twenty-four months in which not to notice.
02What Article 50 asks for, without the jargon
There are four obligations and they do not all fall on the same person. Two are on whoever builds the system and two on whoever deploys it professionally:
- If the system talks directly to people, it has to be designed so the person knows they are interacting with an AI. With one carve-out: not needed if it is obvious to someone reasonably well-informed and observant.
- If it generates synthetic audio, image, video or text, the output must be marked in a machine-readable format and detectable as artificially generated.
- Whoever deploys emotion recognition or biometric categorisation must inform the people exposed to it.
- Whoever generates deep fakes must disclose that the content is artificial.
And one rule that ties them together: the information is given at first contact, clearly and distinguishably, and accessibly. Not on page twelve of some terms.
03The delay that wasn't
In July the headlines said Europe was delaying the AI Act. Regulation (EU) 2026/1744, in force since 27 July 2026, did delay things: high-risk systems under Annex III move to 2 December 2027, and those under Annex I to 2 August 2028.
Article 50 did not move. All that text does to it is remove a power the Commission had to adopt implementing acts, and give a four-month transitional period —until 2 December 2026— to the marking obligation, and only for those who already had the system on the market before 2 August.
So: the part that says "tell them you are an AI" started on time. Anyone who read the headline and relaxed, relaxed about the wrong obligation.
04What we were doing on 2 August
Not much, and it is worth saying before anything else. Nothing in Saelyx declared that it is an AI. Not the consent sheet before the first voice session, not the onboarding, not the agent's greeting.
And there was something worse than the silence. The agent's internal instructions carried a mandatory directive that pushed in the opposite direction on exactly the question the rule contemplates: faced with "are you an AI?", "am I talking to a person?", the agent was steered into redirecting the conversation towards the product's technology.
It was not put there in bad faith —it came from not wanting to give away technical detail— but the effect was the effect: the most direct question a user can ask was the one answered worst.
05What we changed, and when you get it
The directive is fixed. Faced with "are you an AI?", "are you human?" or "am I talking to a person?", the agent says so clearly and first time, without redirecting.
And it pays to be exact about the "when", because that is where the easy lie creeps in: that instruction travels inside the application, not on a server. It is fixed in the code and is already published in the Mac app, from 0.10.72. If you have an older copy, update and you will have it.
Not on the iPhone yet. The instruction travels inside the application, and the iPhone version was built before the fix landed, so it will arrive in a later one. We say so because “Saelyx already says it” in the singular would be false: today it is true on the Mac and not on the phone.
What does not change, deliberately: we don't tell you which models sit underneath. Article 50 requires saying that you are an AI; it does not require publishing what you are made of. Those are different things and the rule does not confuse them.
In the text chat, which does run on a server, there was never an instruction in either direction: neither to declare it nor to dodge it. There the agent declares it out of the model's own behaviour, which is defensible but not the same as having asked it to. We say so because the difference matters.
06What we still owe
Two things, both open:
Telling you before you ask. The rule wants the information at first contact, not at the first question. The consent sheet you see before your first voice session still doesn't declare it. That sheet also travels inside the application, so it will arrive the same way: in a version, not in a server change.
Marking the voice. The obligation to mark synthetic output in a machine-readable format applies to audio, and the Saelyx voice is synthetic audio. When we published this article we said we didn't know whether the provider applied any marking of its own, and that we would publish the answer whatever it was. We have it now.
We ran two different audio files through the provider's own detector: one generated through the text-to-speech path, and one captured from the production agent over its own channel, without going through a speaker or a microphone. Both came back the same: no watermark detected.
The same detector says, separately, that the audio was "very likely" generated with that provider, at 98 % probability. That is not a watermark: it is a statistical classifier inferring it from the signal. Article 50(2) asks for machine-readable marking embedded in the output, and a classifier getting it right is no substitute. We spell this out because that 98 % is easy to present as if it counted, and it doesn't.
So, as of today: the Saelyx voice is not marked, and we do not configure any marking either. The transitional deadline is 2 December 2026.
07Why we tell it this way
It would have been more comfortable to publish an article explaining the rule and let it be understood that we comply. Nobody would have checked: there is no public register to look it up in.
But an article about transparency that isn't transparent about itself is worth nothing, and it ages badly. We would rather it be on record that we were four days late to one part and that the other is still open, with its date in plain sight.
08Frequently asked questions
Since when does Article 50 apply?
Since 2 August 2026. Regulation (EU) 2024/1689 has been in force since 1 August 2024; what began on 2 August 2026 is the application of its general obligations.
Wasn't it delayed?
High-risk systems were delayed, to December 2027 and August 2028. Article 50 was not. Only the marking obligation has a four-month transitional period, until 2 December 2026, and only for systems already on the market before 2 August.
Does Saelyx tell me it is an AI?
On the Mac yes, from 0.10.72: if you ask, it says so clearly and first time. On the iPhone not yet, because that version was built before the fix and the instruction travels inside the application. Declaring it in the consent sheet as well arrives the same way, on both.
Does the Saelyx voice carry a watermark?
No. We checked with the provider's own detector, on both paths —text-to-speech and the production agent— and neither shows a watermark. The same detector identifies the audio as AI-generated at 98 % probability, but that is a classifier, not machine-readable marking, and it does not cover the obligation.
Does this only apply in Europe?
The Regulation is European. Saelyx applies it without looking at where you are from, because building two different products depending on the user's passport strikes us as the exact opposite of what the rule is after.
09Sources
Both texts, in their official versions, checked on 6 August 2026:
- Regulation (EU) 2024/1689 — the AI Act. Article 50 (transparency) and Article 113 (application).
- Regulation (EU) 2026/1744 — the one that delays high-risk and sets the four-month transitional period for marking.
This is not legal advice: it is what we read and what we did.