The useful question is not “Was AI involved?” It is: what was generated or manipulated, who is presenting it, and could a reasonable person mistake it for authentic human or real-world content?

Article 50 of the EU AI Act began applying on 2 August 2026. It contains different transparency duties for providers and deployers of certain AI systems. This checklist turns the European Commission’s July 2026 guidance into an editorial intake process. It is not legal advice, and it cannot decide a borderline case for you.

Start with the role, not the file type

Write down who did what before debating wording.

Question Record
Who supplied the AI system? Provider, platform and product version
Who used the output publicly? Legal entity and responsible editor
What did the system generate or manipulate? Text, image, audio, video or interactive output
What does the audience see? The final asset, surrounding context and distribution channel
Could it be mistaken for authentic content? Yes, no, or uncertain—with reasons

This avoids a common failure: treating the software’s “AI-generated” toggle as the entire compliance analysis. A platform control may be one part of the evidence, but it does not identify your role or explain the final presentation.

Classify the content before choosing a disclosure

Use four working buckets. They are an editorial triage tool, not legal categories.

1. Direct human communication through an AI system

If a person interacts directly with an AI system, check whether the system must make clear that the user is interacting with AI. Do not hide an automated support agent behind a human name or portrait.

2. Synthetic or manipulated media

Record whether an image, voice or video depicts real people, places or events and whether the alteration could create a false impression. A cloned executive voice reading words the executive never approved carries a different risk from a clearly stylised animation.

3. Public-interest text

Escalate AI-generated or manipulated text published to inform the public on matters of public interest. The Commission guidance discusses an exception where there has been human review or editorial control and a person or entity holds editorial responsibility. Do not invoke that exception with a checkbox alone; retain evidence of the review and responsibility.

4. Ordinary production assistance

Spell-checking, noise cleanup, caption timing and layout assistance may not create the same transparency issue as synthetic content. Record why the use did not materially generate or alter the meaning instead of assuming that every automated function is exempt.

Use a two-layer disclosure

Where disclosure is required or editorially appropriate, separate the immediate notice from the fuller explanation.

Layer one: visible at the point of encounter. Use plain language a reader can understand before relying on the content: “AI-generated voice,” “Synthetic reconstruction,” or “Image materially altered with AI.” Avoid a vague sparkle icon without text.

Layer two: method note. Explain the system used, the material human contribution, what was checked and who approved publication. This is useful when a short label cannot communicate the production history.

A model sentence:

This explainer uses an AI-generated narrator. A human editor wrote and fact-checked the script, selected the sources and approved the final audio. The voice does not imitate a named person.

The sentence is specific enough to be useful without turning into a defence of the technology.

Keep a six-field decision record

For each asset, retain:

  1. Asset ID and public URL — so the decision can be matched to the published item.
  2. AI system and version — including the feature used, not just the vendor name.
  3. Input rights — who supplied the text, image, voice or likeness and on what authority.
  4. Material changes — what the system generated, removed or transformed.
  5. Disclosure decision — label text, placement, exception relied on, or reason no notice was used.
  6. Human responsibility — reviewer, approval date and retained source files.

Keep the original and final asset. A disclosure log without the underlying version history is difficult to audit after a complaint.

Do not confuse provenance with disclosure

Content Credentials or another machine-readable marker can preserve useful provenance. A visible notice tells the audience what matters at the moment of use. One does not automatically replace the other.

Use both when the context justifies it: tamper-evident provenance for inspection, plus a visible explanation that does not require specialist software.

A publication gate for small teams

Before release, ask one editor who did not create the asset to answer:

  • Is the identity of an AI system or synthetic person misleading?
  • Could the output reasonably be read as a record of a real event?
  • Are a real person’s voice, face or statements involved?
  • Does the material inform the public on a matter of public interest?
  • Is the notice clear before the audience relies on the content?
  • Can we prove the human review or editorial responsibility we claim?
  • Do platform, advertising, copyright, privacy or employment rules add stricter duties?

If the reviewer cannot answer the fourth or sixth question, pause publication rather than polishing the label.

Limitations

Article 50 is only one part of the legal picture. A disclosure does not create rights to a person’s likeness, cure a misleading advertisement, satisfy data-protection duties or make an unsupported claim accurate. National enforcement practice will develop after the rules begin to apply.

Organisations publishing political, medical, financial, employment or other high-impact content should obtain qualified advice for their facts and jurisdiction.

Bottom line

The strongest workflow is not “label everything AI.” It is a documented classification: role, content, risk, notice and responsible human. That record improves both compliance work and editorial credibility because the team can explain what it actually did.

Sources