The Moment of Disclosure

On July 20, 2026 the European Commission adopted the final Guidelines on the transparency obligations in Article 50 of the EU AI Act. The obligations apply from August 2, 2026. Every one of them attaches to a moment in the run rather than to a document held on file.

What the Commission adopted.

The Guidelines are practical guidance for national authorities, providers, and deployers, issued less than two weeks before the article applies. Article 50 carries four duties.

A provider of a system that interacts directly with a person has to make it apparent that the counterpart is an AI system. A provider of a system that generates synthetic audio, image, video, or text has to mark the output in machine-readable form so it is detectable as artificially generated. A deployer of an emotion recognition or biometric categorisation system has to inform the people exposed to it. A deployer that generates or manipulates a deep fake, or publishes AI-generated text on a matter of public interest, has to disclose it.

The Commission also fixed the manner. The disclosure has to be clear and distinguishable and meet accessibility requirements. Buried in terms of service does not discharge it.

The duty attaches to an instant.

Read the timing language. The interaction notice is owed at the latest at the time of the first interaction or exposure. The marking duty attaches at output generation. The emotion-recognition and deep-fake duties attach before or at the moment of exposure or presentation.

None of those is a state a program can be in. Each is a point in the run, and each binds a particular output rather than the system that produced it. Compliance here is decided one output at a time.

Most builds will treat this as a design-time property. A banner in the interface. A marking library configured once at deploy. That discharges the duty for a system that generates and presents in one step, to one person, in a shape the designer anticipated.

An agent run separates generation from presentation.

An agent run is not that shape. The step that generates content and the step that puts it in front of a person are different steps, often several apart. The agent calls a model, writes the result into state it holds mid-run, passes it through a tool that transforms it, composes a response out of the pieces, and emits. Provenance crosses those hops only if something carries it.

The duty attaches to the step that reaches the person, and the step that reaches the person is rarely the step that generated the content.

A banner configured at deploy time makes a claim about the system. Article 50 asks a question about the output. Between those two sits the part almost nobody has built. Something that knows, at the instant of emission, what this particular output is made of and whether the marking survived the trip.

Where the duty gets discharged.

A duty that binds an individual action is discharged by a control standing in front of that action. A log written afterward records what happened. A control in the path decides what the action is permitted to do while it can still be changed.

That control is a governor. It sits ahead of the emit, reads what the action is about to put in front of a person, and sets the authority of that action against a declared envelope. Where provenance is intact and the marking holds, it grants. Where provenance is thin, the grant narrows. Emit with the disclosure attached. Route it for review. Hold it. The reduced-authority pass is the normal case, and it is the one a binary allow-or-block gate cannot express.

Disclosure becomes a condition of the grant, not a property the system asserts about itself after the fact.

The output that reads cleanest.

There is a case here that runs backward the first time you see it. A fluent, well-formed, confidently generated output is the one a person is least likely to recognize as machine-made. Its polish is what makes the undisclosed version misleading.

So the governor can grant less authority to the step whose content reads better. Trust in the output and authority to emit it are two quantities, and a governor that keeps them separate is free to move them in opposite directions. It reads where the emission is heading, not only how clean it looks in hand, and when the forward look is uncertain the grant gets smaller rather than larger.

That is the failsafe direction, and it is a construction rather than a policy. Anticipation is only allowed to add restriction. A governor that could hand authority back on the strength of the acting system's own report would share the failure it exists to bound.

What the Guidelines leave to the builder.

The Commission specified the standard and left the mechanism open. Clear and distinguishable. Accessible. Owed at the latest at first interaction. Detectable in machine-readable form. Those are properties of a system while it is running, and the guidance does not say how a multi-step agent is supposed to hold them across a chain of tool calls.

That gap belongs to the deployer, and August 2 is nine days from this writing. Nothing in the guidance requires a new model or a redesigned agent. It requires a place in the path where the question gets asked before the output leaves.

What we are building.

Wayfinder Systems Group builds a runtime governor. It sits above control and below intelligence. On every action it sets how much authority the system is granted against a declared safety envelope, then enforces that grant through a path the system cannot step around. It does not retrain your models or redesign your autonomy stack. Where a rule attaches a duty to one particular output, the governor is the thing standing in front of that output at the instant the duty attaches, and it signs what it granted. Patents held in The Wayfinder Trust. We call her Velma.