Explainability versus auditability: what you must be able to demonstrate about AI content from 2 August 2026

You must now demonstrate per workflow how AI content was generated, marked, checked and logged. From 2 August 2026, Article 50 of the EU AI Act requires auditability: the ability to identify, trace and attribute AI use through notifications, machine-readable markings, provenance detection and logs.

The prompt is an analysis of 8 September 2026 of transparency obligations under Article 50 of the EU AI Act, which argues that the practical requirement has shifted from explaining how a model works to demonstrating how AI content moved through your workflow. The European Commission published its implementation guidelines on 20 July 2026, effective from 2 August 2026. In our assessment, this distinction between explainability and auditability reshapes what you must be able to show: not insight into model behaviour, but a verifiable chain of marking, detection, notification and human control that survives later scrutiny.

What is the difference between explaining a model and auditing its output?

Explainability gives a user or supervisor insight into why a model reached a particular output. It is useful for decision-making, but it is not evidence after the fact. Auditability fills that gap. It is the ability to reconstruct, after publication, how a specific piece of AI content came about, who checked it, what choices were made, and what records exist to prove it. A model may be explainable—its logic comprehensible—yet produce content you cannot later account for. Conversely, you may not fully understand a model's internals and still maintain an auditable record of what it generated and how you handled it.

Which concrete controls do the guidelines now require?

Article 50 sets out obligations that apply across your workflow. You must be able to demonstrate the following:

1. Record the model and its purpose — document which model each workflow uses and the lawful basis for the data it touches.
2. Apply machine-readable marking — embed a mark in AI-generated content that identifies it as such, separate from and in addition to any user-facing notification.
3. Provide clear notification to users — ensure the notification is separate, clear and directly visible; a watermark alone is not sufficient.
4. Maintain logs of generation and handling — record when content was created, by which model, who reviewed it and what changes were made.
5. Document human editorial control — where you claim an exception for content subject to genuine human responsibility, show who assessed which version and what choices they made.

When does the exception for human editorial control apply?

The European Commission's guidelines recognise an exception where AI-generated public information was subject to genuine human editorial responsibility and control. This is not a back door. The Commission treats it as an exception in defined cases, not as a default exemption. In practical terms, an informal "someone looked at it" is insufficient. You must be able to show who assessed which version, when, and what editorial choices they made. Without a record, there is no evidence. The exception brings you back to the same principle as the rest of the obligation: auditability through documentation.

What should you secure in contracts with suppliers?

Those making arrangements with suppliers about AI content should secure audit rights and evidence obligations in the contract itself. Suppliers must be able to provide you with the records you need to demonstrate compliance. This includes logs of model use, marking applied, notifications issued and human review undertaken. Without contractual clarity on what evidence the supplier must retain and provide, you cannot later discharge your own obligation to demonstrate auditability.

How do tools support this, and what remains your responsibility?

Verification tools can help make verification steps, corrections and sources visible for inspection. They support review and recording. They do not replace your professional final judgement. The decision on publication and accountability remains with you. Auditability is a governance duty, not a technical one. Tools carry the record-keeping; you carry the choice.