Regulatory field note · Current as of 15 July 2026
Article 50(4): the human-review route — what it actually requires
Most marketing teams meet the EU AI Act through one anxious question: “Do we have to stamp ‘AI-generated’ on everything now?” The answer, for published text, is more interesting than yes or no.
Legal mechanics for orientation, not legal advice. Whether your process satisfies Article 50(4) for a given publication is a judgment you make with counsel.
What Article 50(4) says, shortly
Deployers of AI systems that generate or manipulate text published with the purpose of informing the public must disclose that the text is AI-generated — unless the content “has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content.”
Read that twice, because both halves carry weight:
- A process of human review or editorial control. A process. Not a vibe, not a policy PDF nobody follows — an operational step content actually passes through.
- A person holds editorial responsibility. Someone identifiable stands behind the publication, the way an editor stands behind a masthead.
If both are true, the public “AI-generated” label obligation for that text does not apply. Your blog post reads like your blog post — reviewed, owned, accountable.
The trap: an invisible process is indistinguishable from no process
Here’s the practical problem nobody warns you about. Suppose your team genuinely reviews every AI draft. A regulator inquiry, a procurement questionnaire, or opposing counsel asks: prove it — for this specific article, from last March.
What do you produce? A Slack message that says “looks good”? A doc revision history that shows edits but not accountability? For most teams the honest answer is: nothing durable. A review process that leaves no record gives you the cost of review with none of the evidentiary value. This is the gap between doing Article 50(4) and being able to show Article 50(4).
What “showable” human review looks like
Work backwards from the questions an auditor would ask, and the requirements write themselves:
- Who reviewed it? → the reviewer’s name on the record, not “the team.”
- What exactly did they review? → the exact bytes that published, not “a draft” — content changes after approval must void the approval.
- When, and against what? → a timestamp, the model, the source material.
- Can the record be trusted? → the record must be tamper-evident, so a convenient retroactive edit is visible.
- Was it real review?→ per-asset decisions. Bulk “approve all” buttons undermine the substantive-review basis the provision rests on. (This is why we refused to build one.)
How CrawlQ implements it — the shipped workflow, not a roadmap
In CrawlQ Studio, a workspace owner can turn on compliance mode — deliberately, with a typed confirmation. From that moment:
- No AI-generated asset can publish until a named human approves it. The review queue shows the reviewer exactly what will publish — full, untruncated — plus the model and prompt behind it.
- Approval is a considered act. The button reads “Approve & commit to audit trail.” The confirmation asks the real question — “Take editorial responsibility?” — and records that the reviewer saw those exact bytes. If the content changed since the reviewer loaded it, the approval is refused and they re-review.
- Every approval is committed to a tamper-evident, Merkle-chained audit trail with a per-asset inclusion proof — exportable whenever someone asks you to show your process.
That is Article 50(4)’s workflow step, made durable. Available today, opt-in per workspace.
One honesty note, because it matters: this is a record of your review process, not a certificate of exemption. Whether your process satisfies Article 50(4) for a given publication is a judgment you make with counsel. What we guarantee is narrower and more valuable: when you’re asked to show the process, you’ll have something real to show.
Beyond the EU: the same record answers everyone
The named-human-review record isn’t only an Article 50 artifact. It’s the same evidence US FTC enforcement orders demand for AI documentation, the human-oversight signal UK CMA principles emphasise, the documented review point Quebec Law 25 asks for on automated decisions, and monitoring evidence toward ISO 42001 clause 9.1. One workflow step, multi-jurisdiction evidence. (The full ceiling argument: Does Article 50 apply to my company?)
Would your review survive a “prove it” moment?
Check yourself in 90 seconds.
The free Article 50 Readiness Diagnostic scores your stack across five obligation areas — including whether your current review process leaves evidence. Not legal advice.
Frequently asked questions
Does Article 50(4) mean AI text never needs a label?
No. It provides that the disclosure duty for AI-generated text published to inform the public does not apply where the content underwent human review or editorial control and a person holds editorial responsibility. Other Article 50 duties — deepfakes, interaction disclosure — have their own rules.
Is a general “we review everything” policy enough?
The provision describes a process of review with editorial responsibility. A policy with no per-content record leaves you unable to demonstrate the process happened for any specific publication — which is the practical failure mode.
Who should the named reviewer be?
Someone genuinely exercising editorial control — an editor, content lead, or manager who reads the actual content. The title matters less than the substance: real review of the real bytes, with their name on the record.
Does CrawlQ certify my content as exempt?
No — and be wary of anyone who says they do. CrawlQ records that a named human reviewed and approved the exact published content, in a tamper-evident audit trail. The legal conclusion is yours and your counsel's; the evidence is ours to provide.