What is AI Transparency?

Ask an organisation what its AI system does and you will get the brochure. Ask what data the system runs on, and what it cannot do, and the room goes quiet. AI transparency is the discipline of being able to answer all three questions, in plain terms, to the people who deserve the answers.

What transparency actually covers

AI transparency means telling people what the system does, on what data, and with what limits. Three questions, none of them technical. What it does: the system’s actual function inside your process, described as behaviour rather than aspiration. Does it rank, filter, score, recommend, draft or decide, and at which step? On what data: what it was trained on, in outline, and what it consumes in operation. Not schemas; categories. The application a candidate submitted, or things gathered from elsewhere? The contract as signed, or a summary someone typed? With what limits: where the system is known to be weak, what it was never designed to judge, and where its output should not be trusted without a human.

Notice what is absent. Transparency does not require publishing source code or model weights. A system can be entirely open source and still opaque to everyone it affects, and a proprietary system can be honestly transparent. The obligation is legibility, not disclosure of internals.

Transparency to users, and transparency to those affected

The word hides two different audiences. The first is the user: the recruiter, underwriter or planner who works with the system’s outputs. They need operational transparency: how much to trust an output, in which situations the system is unreliable, and when they are expected to overrule it. This is what makes a human in the loop real rather than ceremonial.

The second audience is the person the system affects: the candidate who was screened, the customer who was scored, the employee whose shift was scheduled. They may never see the interface. They need a different transparency: that a system was involved at all, what it considered, and how to contest the result. Most transparency programmes serve the first audience thoroughly and the second one barely, which is exactly backwards as regulation sees it: the EU AI Act’s transparency obligations point mainly at the people on the receiving end. (This entry is educational context, not legal advice.)

A transparency notice is a decision record, not marketing

Here is the part that gets missed. To write an honest transparency notice, someone must first have decided the things the notice describes: what the system is for, what data it may and may not use, which limits the organisation accepts, who oversees it, and how a result is contested. The notice is downstream of those decisions. If the decisions were never explicitly made, the notice does not describe the system; it describes what its author hopes is true. An honest transparency notice is the visible edge of decisions someone actually made and recorded. Anything else is marketing wearing the costume of disclosure.

This is why the strongest test of a transparency programme is not the prose but the record behind it. When a notice says a human reviews every outcome, can you point to the review record? When it says the system considers only the application, can you show what the system consumed? A notice you can defend line by line is a decision record. A notice you cannot is a liability with a publication date.

One concrete example

Clearly illustrative, with no customer implied. A firm of a few hundred people uses a matching system to shortlist applicants. Its first draft notice reads: “We use AI to improve your application experience.” True, and empty. The rewritten notice says what the system does (ranks applications against the stated requirements of the role), what it uses (the application and the role description, nothing gathered from elsewhere), what its limits are (it assesses nothing outside the written application, and it does not make the decision), who does (a named reviewing role, with the authority to reorder the list), and how to ask for reconsideration. The second notice took longer to write, not because the words were harder but because each sentence forced a decision the firm had been deferring. That is the point. The notice was the record of those decisions.

Transparency and decision intelligence

Transparency is a property of records, not of prose, which is why it belongs to decision intelligence rather than to communications. In ONX, every fact carries a state on the evidence hierarchy (measured, modelled, inferred, stated or unmeasured), so “on what data” has a precise answer with a quality attached. Consequential calls are made by humans, and overrides are recorded with who, when and against what evidence, so “who oversees it” points at a record rather than a role description. That record is a decision audit trail, and ONX applies the same stance to its own AI: candidate matching in Hiring is treated as high-risk under the strict reading of the EU AI Act, with human review gates, and the AI is kept explainable and challengeable. A transparency notice written on top of records like these is not an exercise in wording. It is a report.