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

Explainability shows how an AI system reaches its output; auditability proves, through notifications, machine-readable markings, provenance detection and logs, how AI content moved through your workflow. From 2 August 2026 Article 50 of the EU AI Act requires the latter: you must be able to identify, trace and attribute AI use and AI content demonstrably.

On 20 July 2026 the European Commission published its guidelines on the transparency obligations of Article 50 of the AI Act. According to those guidelines the obligations apply from 2 August 2026. In our assessment this shifts the practical question for teams from ‘can we explain what the model does?’ to ‘can we demonstrate per workflow how AI content was generated, marked, checked and logged?’. That distinction, explainability versus auditability, is central to this article: explaining how the model works as opposed to being able to prove how the content was handled. You will find more context in our topic hub on the EU AI Act and compliance.

What does Article 50 of the AI Act specifically require from 2 August 2026?

In its guidelines the European Commission sets out a number of obligations side by side. The implementation guidelines on Article 50 describe them in detail. According to those documents the following applies, among other things:

• Providers must inform users when they are communicating directly with an AI system; that information must be clear and recognisable at first contact.
• AI-generated or manipulated audio, image, video and text must be marked in a machine-readable way.
• Deployers must disclose deepfakes.
• Text published to inform the public on matters of general interest must be disclosed as AI-generated, unless it was subject to genuine human editorial control.

According to the guidelines, the Commission stresses that a machine-readable watermark alone is not sufficient for the notification to the user: that notification must be separate, clear and directly visible.

Why is explainability not enough for the new transparency obligations?

Explainability is about comprehensibility: it gives a user or supervisor insight into why a model reached a particular output. That is useful, but it is not evidence after the fact. A systematic review in Artificial Intelligence in Medicine states that explainability is still fragmented, insufficiently validated and only weakly translated into compliance across the entire lifecycle. In other words: being able to explain how a model works does not cover the question of whether you can later reconstruct how a specific piece of AI content came about and was handled.

Auditability fills that gap. A preprint on arXiv, A Framework for Responsible AI Systems, describes responsible AI in high-risk scenarios as a coherent lifecycle process with logging, documentation, human oversight, incident analysis and re-examination when something goes wrong. That aligns with what Article 50 requires in practice: not just explanation, but a verifiable chain of marking, detection and notification.

Which pieces of evidence make your AI content auditable per workflow?

In our assessment it is sensible to record per workflow what you need in order to demonstrate that you meet the transparency obligations. A practical checklist:

1. The notification to the user that there is AI interaction, with a record of how and when it was shown.
2. The machine-readable marking on generated or manipulated content, plus a means of detecting that marking.
3. A provenance record: which model or which chain produced the output, and at what moment.
4. For public information on matters of general interest: the disclosure, or the evidence that the exception for human editorial control applies.
5. Logs and version control with which you can reconstruct the steps after the fact.

Those making arrangements with suppliers about this can best secure them in contractual audit rights and evidence obligations in AI contracts. For the marking obligation itself, our explanation of marking AI content and the AP's supervision from 2 August 2026 is a more concrete starting point.

When does the exception for human editorial oversight apply?

According to the European Commission's guidelines, the disclosure obligation for AI-generated public information is not absolute. The exception applies where the content was subject to genuine human editorial responsibility and control. That is not a back door: the Commission treats it as an exception in defined cases, not as a default exemption.

In practical terms this means you must be able to demonstrate the human control. In our assessment an informal ‘someone looked at it’ is insufficient; you want to be able to show who assessed which version and what choices were made in doing so. That brings the exception back to the same principle: without a record there is no evidence.

How do you set up explainability and auditability together?

Explainability and auditability are not opposites; they complement each other. Explainability helps your people make better decisions, auditability helps you account for those decisions after the fact. The broader shift the European Commission illustrates here fits a movement from AI governance principles to concrete control duties.

That this governance is now also practically urgent for European providers is clear from the attention on the continent: Reuters reported on 8 September 2026 that the French company Mistral closed a large funding round, co-led by investors with EU involvement. That is not central to these transparency rules, but it is a sign that the verification and governance question around European models is gaining weight.

Where teams process sensitive documents, a verification layer such as Vera can help to make verification steps, corrections and sources visible for inspection. This supports review and recording, but does not replace your professional final judgement. The decision on publication and accountability remains with you.

Sources: The article draws on the European Commission's guidelines on Article 50 of the AI Act, a review in Artificial Intelligence in Medicine, an arXiv preprint on responsible AI systems and reporting by Reuters on Mistral.

← All articles in this topic ← All articles