Must you tell a candidate they are talking to an AI?

Yes if they converse with it, no if a human reviews and sends. The line the AI Act draws does not run where everyone places it: it does not separate what was written by an AI from what was written by a person, it separates direct interaction from text produced under human responsibility.

The deadline, meanwhile, has passed. The transparency obligations of Article 50 have applied since 2 August 2026, and they should not be confused with the postponement of the Annex III high-risk systems, moved to 2 December 2027. These are two distinct calendars, and postponing one suspends nothing of the other.

What does Article 50 actually say?

It sets out three obligations with neither the same addressee nor the same trigger, and confusing them is behind nearly every misreading I hear on the subject.

Paragraph 1 covers direct interaction. A system designed to interact directly with natural persons must be built so that those persons are informed they are dealing with an AI, unless that is obvious to a reasonably well-informed, observant and circumspect person. The information must be given clearly and distinguishably, at the latest at the time of the first interaction.

Paragraph 2 covers marking synthetic content. Providers of systems generating audio, image, video or text must mark the outputs in a machine-readable format. The provisional “omnibus” agreement of May 2026 gives systems already on the market before 2 August until 2 December 2026 for that specific obligation.

Paragraph 4 covers publication. A deployer distributing AI-generated text must disclose it, but only where the text is published for the purpose of informing the public on matters of public interest, and the obligation falls away where the content has undergone human review with editorial responsibility assumed.

The case everyone thinks is covered, and is not

The follow-up email drafted by an agent, reviewed by a recruiter and sent under their name. It is the case I get asked about every time, and it falls under none of the three obligations.

Paragraph 1 does not apply, for lack of direct interaction: the candidate reads a message, they are not addressing the system. Paragraph 4 does not apply either, because a recruitment message is not text published to inform the public on a matter of public interest, and because human review with assumed responsibility is precisely what the text retains as exempting.

That conclusion surprises people, and it deserves two warnings rather than a victory lap. First, the review must be substantive and not a rubber stamp, which the guidelines state without ambiguity: clicking send without reading is not editorial responsibility, it is a blank signature. Second, the AI Act is not the only text that applies, and the GDPR carries on living its own life on automated decisions as on informing data subjects.

The case that genuinely is covered

The agent replying to the candidate live, by instant messaging, by chat on your site or by phone. Here the obligation applies in full, without argument, and it is settled by one sentence at the start of the exchange.

The point not to miss is the timing. The information is due at the latest at the first interaction, which rules out the footer mention, the line in the terms and conditions and the reminder sent the next day. It must also be distinguishable, meaning it cannot blend into the message carrying it.

The same reasoning applies to an interview conducted by an AI, which incidentally stacks this obligation on top of far heavier questions about how the answers are then used. We have written about why a score attached to a candidate is the wrong tool, and transparency about the nature of the interlocutor repairs nothing if what comes out of the exchange is a number nobody can contest.

Why the obviousness exemption will not protect you

It is written for cases where nobody could reasonably be mistaken, and it gets invoked for cases where one would rather say nothing.

The standard is a reasonably well-informed, observant and circumspect person, and the guidelines call for a two-step assessment: identify the audience concerned, then judge its actual level of awareness. They add that it should not be over-relied upon, which is a careful way of saying the exemption will be read narrowly.

Apply it honestly to a candidate. They apply in the evening, and receive a well-built, relevant reply within twelve seconds, at ten at night. Somebody in the industry will conclude it is a machine; a thirty-year-old engineer applying for the first time in six years will conclude nothing at all, and they are the person concerned. What is obvious to you is not obvious to them, and that is the whole subject.

What we do, and why it is not for compliance

We announce the agent in every case where it addresses someone outside the company, including those the text does not require, and that decision predates the August deadline.

The reason is not legal, it is economic. A candidate who finds out afterwards that they were dealing with a machine does not conclude the practice was lawful, they conclude their time was wasted, and they tell others. The cost of a disclosure sentence is nil; the cost of a reputation for having robots talk to people without saying so is paid on every subsequent hire, in a market where silence is already the first grievance candidates raise.

There is a second effect, less expected, that we observed by doing it. Announcing the agent changes candidate behaviour for the better: they ask more direct questions, they test the form less, and they accept a fast answer on a factual point more readily. Transparency removes an awkwardness both sides were maintaining.

How to word the sentence

It is harder to write than it looks, because the two possible mistakes do not cancel out and neither is visible on a reread.

The first is burying the information. “This service uses artificial intelligence technologies to improve your experience” does not tell the person they are addressing a machine, it tells them something modern is happening somewhere. The text calls for clear and distinguishable information, which rules out the compliance formula slipped between two marketing sentences.

The second is overdoing it, and it is more common among those who take the subject seriously. A solemn three-line warning with a link to a policy lands on a candidate like a legal notice and suggests something worrying is going on. You have informed and you have alarmed, which was not the point.

What works, in our experience, is one sentence that names the thing, says what it can do and says how to reach a human. Something like: “You are talking to such-and-such firm’s assistant, which can answer questions about the role and book a slot. Ask for a person and I will pass you on.” It is short, it sits at the start, and the last clause does most of the work: knowing an exit exists makes people markedly more tolerant of the machine.

What it changes in a process

Three things to check, and none of them needs a project.

Identify every place a candidate can receive an answer without a person having released it, which is also the right exercise for working out who approves what. Each of those points is a site of direct interaction, and therefore a place where the disclosure sentence has to exist.

Then check that your review is one. If the recruiter approves forty messages in four minutes you do not have editorial responsibility, you have a button, and the exemption you believe you hold does not exist.

Finally, write down somewhere which systems are concerned and on what basis. That document does not need to be long, it needs to exist before somebody asks for it, and it is the same reflex that makes a usage charter written in advance useful rather than one written after an incident.