AI public-interest text needs an editorial review record

A person clicking “approve” beneath an AI-written draft is not necessarily an editor. From today, that distinction matters for businesses publishing text about public health, consumer safety, finance, politics or other matters that shape public debate.

Article 50 of the AI Act applies from 2 August 2026. The Commission’s final guidance says that in-scope AI-generated or manipulated text must be clearly labelled unless it has undergone substantive human review or editorial control and a natural or legal person holds editorial responsibility for the publication. A spelling pass is not enough.

For an SME, the practical response is not a new policy with a broad promise that “all AI content is reviewed”. It is a short record tied to the exact version that went public: who examined the substance, what they checked, what they changed and who accepted responsibility for publication.

Start with the publication test

The labelling rule does not attach to every sentence that has touched an AI system. The Commission’s questions and answers set out three cumulative characteristics. The text must be published, it must inform the public, and it must concern a matter of public interest.

That final category is broad. The Commission lists politics and democratic processes, public administration and services, justice and law enforcement, fundamental rights, public security, public health, environmental protection, consumer safety, and economic, financial, scientific or cultural developments that may be relevant to public debate.

Purpose and context therefore matter more than document labels. A private draft circulated to two managers is not published to the public. A product description for an ordinary spare part may not inform public debate. A public blog post claiming that the same part reduces pollution, improves safety or qualifies for a subsidy can cross into public-interest territory.

Record the classification before choosing the control:

Question	Evidence to keep
Was the text published?	Final URL, channel, publication time and intended audience
Was its purpose to inform the public?	Brief, page purpose and distribution decision
Did it concern public interest?	Named topic and one-sentence rationale linked to the Commission’s categories
Did generative AI generate or manipulate it?	System, task and the part of the text affected
Was there substantive review or editorial control?	Reviewer, checks, changes and approval evidence
Who holds editorial responsibility?	Named natural or legal person and publication authority

Source: European Commission Article 50 questions and answers and transparency guidelines. Last verified 2026-08-02.

Do not turn the first three rows into an automatic “no” for all commercial material. Economic, financial, scientific and consumer-safety developments can be matters of public interest. The person making the classification should be able to explain the publication’s real purpose, not merely select a content type from a dropdown.

Review substance, not punctuation

The Commission defines human review as deliberate examination of the substance by one or more natural persons with relevant knowledge and professional judgement. Its examples include academic peer review and professional validation chains.

Editorial control is a related but distinct route. It is control exercised in practice by a responsible editorial entity with authority to approve, alter or reject the substance on substantive grounds. The Commission expressly includes fact-checking and checking the trustworthiness of sources.

That makes several familiar approval steps too weak on their own:

• running spelling, grammar or tone checks;
• checking that a template is complete;
• confirming that a manager opened the draft;
• approving only layout, brand colours or legal boilerplate;
• accepting a contractor’s assurance that the output “looked right”;
• comparing the final text only with the AI prompt rather than with evidence.

The Commission calls superficial, solely formal or procedural checks insufficient. A reviewer does not need to rewrite every line, but must be able to challenge the claim, source, context and omission that could change what the reader understands.

Relevant knowledge is tied to the subject. A communications manager may review tone and publication purpose. A safety claim may also need an engineer. A health statement may need someone qualified to assess the evidence. Splitting those checks across people is compatible with a validation chain, provided the record shows who examined what.

Keep one version-linked record

Article 50 does not prescribe a particular review form. The Commission’s guidance instead explains what qualifies and how compliance may be demonstrated. A compact operational record is therefore a defensible way to show that the claimed process happened in practice.

Use one row per published item, not one annual declaration for the whole company:

Field	Minimum entry
Publication	Title, URL, channel and planned publication time
Final version	Document version, CMS revision or cryptographic hash
AI involvement	System used, task performed and affected sections
Public-interest test	Published / informative / topic rationale
Reviewer	Name, role and relevant subject knowledge
Substantive checks	Claims, dates, calculations, quotations, sources, context and material omissions
Review result	Changes required, rejected claims and unresolved limits
Editorial decision	Approve, alter, reject or publish with an AI disclosure
Responsibility	Natural or legal person holding ultimate responsibility
Evidence	Source copies, marked draft, comments and approval timestamp

Source: Flint Brief operational record derived from the Commission’s definitions of human review, editorial control and editorial responsibility. Last verified 2026-08-02.

The version link is the control most likely to disappear. A review of draft 7 does not prove that draft 9 was reviewed after an automated rewrite. Store the CMS revision, exported final copy or hash alongside the decision. If the text changes materially, reopen the record.

Keep prompts only where they help explain the role AI played and where retention is compatible with confidentiality and data-protection duties. The objective is not to archive every keystroke. It is to connect the published claim with the evidence and accountable judgement that permitted it.

Contractors do not move the duty away

The Commission explains that a legal person remains the deployer when employees use an AI system under its authority. The same can remain true when contractors or freelancers operate the system on the business’s behalf and under its responsibility and control.

An outsourced content workflow should therefore assign evidence, not assume the duty was outsourced. The brief should state whether AI may be used, which sources must be checked, who performs substantive review, who may alter or reject the text and who holds editorial responsibility when it is published on the SME’s channel.

Ask for the marked draft and source trail as part of acceptance. A supplier certificate saying “human in the loop” does not identify the human, the loop or the judgement exercised. If the SME itself gives final substantive approval, its reviewer must have access to the underlying evidence rather than only the polished output.

This distinction also prevents accidental overclaiming. A freelancer may be responsible for their professional work, but the business publishing under its own name still needs to know whether it can rely on the review exception or must disclose AI involvement.

Label when the evidence is thin

If the three publication criteria are met and the business cannot establish qualifying review plus editorial responsibility, the safer operational branch is disclosure.

The disclosure must be clear, distinguishable, accessible and provided no later than first exposure. It should tell the reader that the text was artificially generated or manipulated. Hiding the fact in terms and conditions, metadata or a machine-readable mark does not give the reader the perceivable disclosure expected from a deployer.

The EU offers optional icons, but the Commission says an icon does not establish compliance by itself. Contextual wording can make the meaning clearer, particularly where “modified” could refer to ordinary human editing. Keep the label attached when the text is syndicated, exported or republished in another channel.

The limited transition to 2 December 2026 concerns the provider-side marking and detection obligation for certain AI systems placed on the market before 2 August. The Commission’s current questions and answers say the deployer rule for public-interest text applies from today. Do not use the marking grace period as a general delay for publication controls.

Run a five-item day-one audit

Select the last five public texts in which generative AI played a meaningful role. For each one, ask:

1. Can we explain whether it was published to inform the public on a matter of public interest?
2. Can we identify the exact AI-generated or manipulated portion?
3. Can a named reviewer show substantive checks against reliable evidence?
4. Can we identify the person or entity holding editorial responsibility?
5. Can we connect that decision to the version the public actually saw?

Where all five answers are supported, preserve the record. Where the text is out of scope, preserve the short classification rationale. Where review evidence is missing, fix the workflow and use a clear disclosure for the next in-scope publication.

The useful control is not a universal ban on AI drafting or a universal label on everything. It is a fork that can be audited: classify the publication, prove substantive editorial judgement, or disclose AI involvement at first exposure. Test that fork on five real items before the next automated draft reaches the publish button.