Human review and editorial responsibility for AI

AI ACT · EDITORIAL CONTROL

A person at the end of a workflow does not automatically create human control. Review becomes real when the reviewer has competence, time, sources, authority to change the work and responsibility for what is published.

Review is not a formality

The Commission defines human review as deliberate examination of a content’s substance by people with relevant knowledge and professional judgement. Editorial control requires a responsible entity able to approve, change or reject the text on substantive grounds, including fact-checking and source-trustworthiness assessment. ^{[EU-A50] [EU-A50-G]}

A purely grammatical or procedural check is insufficient for the labelling exception concerning certain public-interest texts. Review should not be reduced to fear of sanctions: it is how an organisation preserves accuracy, first-hand experience and editorial voice while using automation. ^{[EU-A50] [G-AI]}

Step 1: fix the purpose and boundaries

Review starts before generation. The brief should define audience, question, commercial or informational purpose, subjects requiring expertise and claims the system must not invent. Without those boundaries, the editor receives fluent prose but has no standard against which to check it. ^{[G-HELP] [G-AI]}

For subjects affecting health, money, safety, rights or compliance, identify the competent validator from the outset. An SEO editor may verify structure and intent but should not independently confirm a legal, medical or financial conclusion they cannot professionally assess. ^{[EU-A50] [G-HELP]}

Step 2: build a source ledger

Every verifiable claim should link to a directly reviewed source with URL, issuer, date and relevant passage. For rules and specifications use the legal text, authority or vendor documentation rather than circular summaries. Distinguish observed information, inference and recommendation. ^{[EU-A50] [G-AI]}

The editor opens the source, checks currency and confirms that it supports the exact sentence. A thematically relevant link is not necessarily evidence. If a source has changed or remains draft, the text should say so and avoid presenting an interpretation as a final obligation. ^{[EU-A50-G] [G-HELP]}

Steps 3 and 4: verify substance and add experience

Break the draft into claims: names, dates, percentages, conditions, exceptions, causes and conclusions. Confirm each item, seek contradictions and remove false precision. Then check whether it answers the real question rather than repeating definitions. A reasonable conclusion should be identified as interpretation, not presented as fact. ^{[G-AI] [G-HELP]}

Original value appears when the editor adds real examples, operational limits, trade-offs and decisions the model cannot infer from public documents. Google recommends clarity about who created content, how it was produced and why it exists. Experience is not a decorative anecdote but information that changes the reader’s understanding. ^{[G-HELP] [G-AI]}

Step 5: review the page, not only the prose

Titles, meta descriptions, headings, links, alt text, structured data and social images can introduce claims absent from the body. Google requires accuracy in these elements too. Check that schema faithfully describes the page, links support claims and translations preserve legal or technical meaning. ^[G-AI]

On a multilingual site, review does not end with canonical-language approval. English and German titles, descriptions and key wording must be read as editorial texts, not treated as substituted strings. A fluent translation can still alter an exception, degree of certainty or operator identity. ^{[EU-A50] [G-AI]}

Step 6: name the approver and retain evidence

The final record should show who reviewed substance, who holds editorial responsibility, which sources were consulted, which material changes were made and when the piece will be revisited. Retain both the version entering review and the published version; their difference evidences the ability to intervene. ^{[EU-A50] [EU-A50-G]}

A recommended model for AYSA.RO and AYSA.AI workflows would retain research, sources, draft, changes, approval status and approver identity in one trail. This is an operating recommendation, not a claim that every feature is already implemented or that its mere existence proves compliance. ^{[EU-A50] [G-HELP]}

Official sources and verification date

1. European Commission — transparency obligations under Article 50
2. European Commission — guidelines on transparency obligations
3. Google Search Central — using generative AI content
4. Google Search Central — optimising for AI features in Search
5. Google Search Central — helpful, reliable, people-first content