Where Your AI Disclosure Actually Has to Appear

With the transparency obligations in Article 50 applying from Sunday, most of the writing about them has explained what the Article says. Less of it has answered the question a delivery team actually asks on the Monday after: where does the disclosure go? A rule about informing people is easy to agree with in principle and surprisingly awkward to place in practice, because professional work does not arrive in one format through one channel.

The Article distinguishes between two situations that get conflated constantly. One concerns systems that interact directly with a person, where the person must be told they are dealing with AI rather than a human, unless that is already obvious from the context. The other concerns content: AI-assisted or manipulated material, with particular attention to synthetic media and to text published to inform the public on matters of public interest. These are different duties with different triggers, and a firm can easily fall under one and not the other.

The obviousness exception is doing real work

The phrase “unless it is obvious from the context” carries more weight than its length suggests, and it is the reason blanket disclosure is usually the wrong instinct. Nobody needs a notice explaining that a chatbot on a support page is software. Where the exception stops helping is the middle ground: a client receiving a written analysis has no contextual signal at all about how it was produced, and that ambiguity is exactly what the transparency direction is aimed at.

For professional services, the practical reading is that the risky cases are not the obvious robots. They are the deliverables that look entirely human because a human did in fact write, shape, and sign them, while AI did meaningful work somewhere in the middle. Whether any specific deliverable falls within Article 50 is a question for counsel and depends on what the work is and who it is for, but the direction of travel is not ambiguous.

Four places disclosure tends to belong

The first is the engagement letter or statement of work, which is where AI use should be established rather than confessed. A standing clause describing how the firm uses AI assistance, and what review it applies, converts disclosure from an awkward per-document act into an agreed term of the relationship. Clients who know the answer in advance stop treating the question as an accusation.

The second is the deliverable itself, usually as a short methodology note rather than a disclaimer. There is a meaningful difference between a defensive line saying AI was involved and a confident one describing that analysis was AI-assisted, checked against named sources, and reviewed and approved by a named person before release. The second reads as process maturity, and it is also considerably more useful to the reader.

The third is anything published rather than delivered privately. Public-facing material carries the tightest expectations, and also the highest cost when a claim turns out to be unsupported, because the correction is public too. The fourth is the internal record, which is not a disclosure to anyone outside the firm but determines whether you can substantiate the other three when somebody asks six months later.

Disclosure you cannot substantiate is a liability

This is the part that gets skipped. A methodology note claiming that AI-assisted analysis was reviewed before delivery is a statement about your process, made in writing, to a client. If a dispute arrives and the only support for that statement is somebody’s recollection, the disclosure has quietly made the position worse rather than better, because now there is a documented claim with nothing behind it.

Which means the sequence matters, and most teams run it backward. Being able to show what was checked has to come first, and the disclosure describes something real. Otherwise the note is a promise the firm cannot keep, written on the record, which is a strictly weaker position than saying nothing would have been.

Start with the highest-exposure workflow

The realistic move is not to solve disclosure across every channel at once. Pick the workflow where AI does the most work and the audience matters most, usually client-facing analysis or anything published, and get that one to a state where the disclosure you write is backed by a record you could produce on request. One workflow that genuinely holds up beats a policy covering everything and evidencing nothing.

Qonera is the AI governance platform for professional teams, and the review and approval workflow is built so that the record writes itself: which sources the answer drew on, where independent models agreed and disagreed, which claims carry citations, and who approved the result before it left the building. How that maps to the Act’s transparency obligations is set out on the EU AI Act page. Disclosure is a sentence in a document, but what makes it worth writing is everything standing behind it.

This article is for general information only and does not provide legal advice. Organisations should consult qualified legal counsel about how Article 50 and the EU AI Act apply to their specific systems, workflows, and obligations.