Label AI-Generated Content: Does This Text Need an AI Disclosure?

Not every item created with AI needs a visible disclosure. Article 50 primarily requires disclosure for deepfakes and certain public-interest text, while providers must support machine-readable marking of synthetic output. Businesses therefore need an approval process that evaluates content type, degree of AI involvement, publication purpose, and editorial responsibility before release.

Legal status reviewed on July 14, 2026. This article provides operational guidance and is not a substitute for legal advice. At that date, the European Commission (https://commission.europa.eu/) had published the final voluntary Code of Practice but was still preparing the final Article 50 guidelines.

Why does the labeling question often produce the wrong answer?

A mid-sized company prepares a new thought-leadership article. A language model creates the first draft. A subject-matter expert then checks the sources, rewrites major sections, adds industry examples, and approves the final publication. Does the website need to display “This article was written by AI”?

The marketing team creates an illustration of a modern service facility that does not represent any real location. Elsewhere, the same team modifies a photograph of an actual project. AI adds safety barriers, removes visible defects, changes the weather, and makes unfinished work appear complete. Should both images receive the same label?

Other common cases include synthetic narration, generated product descriptions, social media graphics, virtual presenters, artificial customer testimonials, translated videos, AI-assisted press releases, and photographs extended with generative fill to fit a new aspect ratio.

A rule stating that every use of AI must be disclosed goes beyond Article 50. The opposite rule—that human approval always removes the need for disclosure—is also unreliable. The EU AI Act separates technical marking duties placed on system providers from audience-facing disclosure duties placed on professional deployers. It also treats text differently from image, audio, and video.

The right starting point is therefore not merely whether AI was used. A business must ask what was generated or modified, whether the content resembles an existing person or event, whether an audience may take it as authentic evidence, whether public-interest information is involved, whether meaningful editorial review occurred, and who assumes responsibility for the final publication.

What does Article 50 require for AI-generated content?

Article 50 creates two different layers of responsibility.

The first applies to providers of generative AI systems. Providers of systems that generate synthetic audio, images, video, or text must generally design the system so its outputs are marked in a machine-readable format and can be detected as artificially generated or manipulated.

Possible techniques include provenance metadata, digital signatures, watermarks, fingerprints, or combined technical methods. The chosen approach must account for the content type, technical feasibility, implementation cost, and the state of the art. Article 50 provides an exception where a system only performs an assistive standard-editing function or does not substantially change the input or its meaning.

The second layer applies to deployers, including businesses, agencies, publishers, and professional creators using generative systems. A human-facing disclosure is required when:

• generated or manipulated image, audio, or video constitutes a deepfake; or
• generated or manipulated text is published to inform the public about matters of public interest without sufficient human review or editorial control and without a person or organization assuming editorial responsibility.

These obligations apply from August 2, 2026. Under the 2026 transitional arrangement, providers of relevant generative AI systems placed on the market before that date have until December 2, 2026 to implement the machine-readable marking requirement.

What is the difference between machine-readable marking and an audience-facing disclosure?

Machine-readable marking is designed for software, platforms, detection services, and provenance tools. It enables a file or output to carry information indicating that AI generated or modified it.

The information may be embedded inside the asset, connected through a cryptographically secured provenance record, or supported by complementary detection methods.

The Coalition for Content Provenance and Authenticity (https://c2pa.org/) publishes the open C2PA standard for Content Credentials. A credential can record how an asset was created, which tools modified it, and whether an AI system participated in the workflow. C2PA also emphasizes that provenance data does not determine whether the underlying claim is true and cannot solve misinformation by itself.

An audience-facing disclosure is intended for the reader, viewer, or listener. It may appear as text, an icon, an overlay, a spoken statement, or a combination of elements.

Examples include:

AI-generated illustration

This image was partially modified using AI

The narration uses a synthetic voice

This video contains AI-generated scenes

This text was generated using AI and was not editorially reviewed

Metadata alone does not satisfy an audience-facing disclosure obligation when Article 50(4) applies. A visible label does not replace the provider’s technical marking obligation either.

An SME does not need to invent its own watermarking technology for every model it uses. It should, however, ask whether the provider supports machine-readable provenance, whether the information survives export, and whether the particular publication also requires a disclosure to the audience.

What content qualifies as a deepfake?

The EU AI Act definition extends beyond manipulated recordings of politicians or celebrities. A deepfake is AI-generated or manipulated image, audio, or video content that resembles an existing person, object, place, entity, or event and could falsely appear authentic or truthful to a person.

This can cover ordinary-looking corporate media.

A construction business might publish an AI-generated image that appears to document a completed customer project. A traffic-management provider could add barriers, signs, equipment, or workers to a photograph of a real work zone. A marina could present a synthetic aerial image as though it showed actual operations. A field-service company could generate a fictional “before and after” image of a real facility.

Audio and video create similar issues. Examples include a cloned executive voice, an artificial customer interview, a digital version of a real employee, or synthetic footage showing a real machine performing an action that never occurred.

An abstract diagram, an obviously graphic illustration, or a fictional person with no meaningful resemblance to a real individual does not automatically constitute a deepfake. A fully generated product image may nevertheless mislead an audience when it is presented as an actual photograph, an existing reference installation, or evidence of completed work.

A corporate review should therefore consider more than the reproduction of human faces. Existing products, facilities, projects, offices, job sites, vessels, machines, and events can also be represented deceptively.

Must every AI-generated image receive a visible label?

No. Article 50(4) requires audience-facing disclosure for image, audio, and video when the content constitutes a deepfake. It does not create a universal visible-label requirement for every synthetic image regardless of its context.

A broader internal disclosure policy may still be commercially sensible. A business can label material that falls outside the strict deepfake definition when an audience could reasonably interpret it as factual documentation.

Examples include:

• product visualizations produced before manufacturing,
• artificially furnished properties or offices,
• simulated industrial or construction scenes,
• generated employee and customer scenarios,
• artificial before-and-after demonstrations,
• fictional case studies,
• AI-created emergency or safety scenarios.

Terms such as “concept visualization,” “AI-generated example,” or “illustrative rendering” may communicate the purpose more accurately than a generic AI label.

Decorative textures, abstract backgrounds, icons, and visibly stylized illustrations will often require neither a statutory disclosure nor an extensive explanation. Existing machine-readable provenance should nevertheless not be intentionally removed without a documented reason.

When must AI-generated text be labeled?

The Article 50 rule for text is narrower than the deepfake rule. Disclosure applies when an AI system generates or manipulates text that is published for the purpose of informing the public about matters of public interest.

Not every piece of corporate writing meets that condition. Product descriptions, internal reports, individual quotations, private email, routine sales copy, and ordinary customer correspondence are not public-interest publications merely because AI assisted with the wording.

Potentially relevant content includes publications concerning public safety, health, environmental incidents, political or regulatory developments, major social events, public infrastructure, or other subjects affecting a broader group of people.

A corporate article explaining the EU AI Act may concern a matter of public interest. That does not automatically require a label. Article 50 excludes text that underwent human review or editorial control when a natural or legal person assumes editorial responsibility for the publication.

A professionally reviewed article published by KrambergAI GmbH (https://krambergai.com/) can therefore be treated differently from an automated news page that republishes AI summaries without substantive review.

What should human review mean in a corporate publishing process?

The EU AI Act does not prescribe a fixed number of review rounds or a particular approval form. The text exception requires both a human-review or editorial-control process and editorial responsibility assumed by a person or organization.

A nominal approval click provides limited evidence of a genuine editorial process. In practice, the responsible reviewer should:

• verify important factual statements and statistics,
• inspect sources and quotations,
• remove misleading generalizations,
• align statements with the company’s actual services and expertise,
• escalate legally or technically sensitive claims,
• determine the final tone, structure, and conclusions,
• approve the publication in an attributable capacity.

A spelling review alone may not demonstrate meaningful editorial control when substantive claims remain untouched and unverified. On the other hand, an editor does not need to rewrite every sentence from the beginning. The essential point is that the organization evaluates and adopts the final text as its own publication.

For ordinary corporate articles, a proportionate record may consist of the reviewer’s name or role, review date, source list, approval status, and final version.

How should AI-assisted and primarily AI-generated text be distinguished?

Text production rarely follows a single-step process. An employee may create the outline, use a model for selected paragraphs, add professional examples, verify sources, change the argument, and then ask the assistant to produce an executive summary.

A fixed percentage dividing human from artificial contribution would be difficult to apply. Neither the EU AI Act nor the 2026 Code of Practice requires companies to calculate an exact AI share for every publication. During development of the code, the European Commission (https://commission.europa.eu/) also moved away from a rigid taxonomy separating AI-generated and AI-assisted content.

A functional internal classification works better:

Minor AI assistance includes spelling, formatting, or limited language refinement that does not substantially alter the underlying meaning.

Substantive AI assistance occurs when the model proposes wording, arguments, summaries, or structure, while a person subsequently evaluates and revises the material.

Predominantly AI-generated text with editorial review begins largely as model output but becomes a company publication through substantive review, correction, rewriting, and approval.

Unreviewed AI publication occurs when a system creates material that is published automatically or almost unchanged without a responsible person validating its claims and impact.

For Article 50, the final category is particularly relevant when public-interest information is involved. The business should still classify all publication workflows because the categories determine the appropriate approval level.

What applies to websites, blogs, and downloadable documents?

A corporate website may contain several different cases.

An editorially reviewed blog article does not require a visible label merely because a writing assistant participated. This can also apply to a public-interest subject when meaningful editorial control occurred and the company assumes responsibility.

An automated news section that summarizes external reports and publishes them without review is different. When it informs the public about regulatory, financial, health, environmental, or social developments, Article 50 disclosure may be required.

White papers, management briefings, and downloadable PDFs should follow the same review approach. If they contain generated images of real people, facilities, or events, the image analysis must remain separate from the text analysis.

A generic statement in the website footer that AI “may be used” does not satisfy the obligation for a specific deepfake. Article 50 requires the information no later than the audience’s first exposure to the relevant content.

For images, a caption or embedded label may be appropriate. Video may require an opening disclosure and an additional label during the affected scene. Audio can use a short spoken notice supported by text in the player interface.

What applies to LinkedIn, Pinterest, YouTube, and other social platforms?

Social platforms create a practical difficulty: upload processes can modify or remove metadata, posts are reshared, and media may appear outside the original profile.

The European Commission’s recommendations for the new EU icons state that disclosure should be perceivable no later than the audience’s first exposure. Where possible, the label should remain connected to the media and survive sharing or downloading.

A disclosure placed only at the end of a long caption may therefore be ineffective when the user encounters the image or video first. Deepfake disclosure should appear on the asset, at the beginning of the post, or through a dependable platform overlay.

Automatic platform labels can help. The publishing company should still test whether the label appears after upload, remains visible on mobile devices, and continues to describe the content accurately.

LinkedIn diagrams, Pinterest graphics, and abstract B2B illustrations are not automatically subject to visible disclosure. The analysis changes when a publication artificially portrays a real executive, customer, product, facility, reference project, or event.

Possible labels include:

AI-generated illustration — not an actual customer project

Visualization of a possible implementation

Image partially modified using AI

Synthetic scene — depicted individuals are fictional

What applies to advertising and product representations?

Article 50 is not the only relevant framework for advertising. Even where a statutory AI label is not required, a representation should not create a false impression about product characteristics, completed work, customer experience, testing, or availability.

A manufacturer may use a generated concept image for a future machine. The presentation should not suggest that the equipment has already been produced, certified, installed, or used by a customer when that is not the case.

A contractor may show an AI-rendered renovation concept. Calling it a “design visualization” or “illustrative concept” prevents confusion with an actual completed reference.

Testimonials require particular caution. A generated individual who appears to be a real customer describing a supposed experience may implicate Article 50 and may also misrepresent the existence of an authentic customer opinion.

An additional internal advertising question is therefore useful: Would an ordinary viewer interpret the image, voice, text, or video as a factual claim about a real product, project, organization, or customer experience?

When the answer is yes, the company should use an explanatory label, change the presentation, or avoid the asset.

What applies to synthetic voices and AI-generated audio?

A generic synthetic voice is not automatically a deepfake. The classification becomes more relevant when it resembles an existing person and could be mistaken for that person’s authentic statement.

Common deepfake examples include:

• a cloned executive voice,
• an artificial statement attributed to a known employee,
• a simulated customer interview,
• a voice modeled on a professional expert,
• audio presented as a recording of a real conversation.

A neutral narrator that is not associated with a real person may fall outside the statutory deepfake rule. Voluntary disclosure may still be useful when the voice performs a personal or authoritative role.

An AI phone assistant presents a different issue. Its primary obligation concerns informing callers that they are interacting with an AI system under Article 50(1), rather than labeling the audio file itself as a deepfake.

Businesses should therefore separate three questions: whether audio is synthetic, whether it imitates a real person, and whether it is part of an interactive AI service.

What applies to video avatars and virtual presenters?

Corporate video can combine several AI elements: synthetic imagery, generated speech, translated lip movement, artificial backgrounds, and an automatically drafted script.

An avatar modeled on a real person can qualify as a deepfake when the output appears to be an authentic recording of that person. This may remain the case even where the individual consented to the production. Intent to deceive is not necessarily required by the definition; audience perception is central. Consent remains important for separate legal and contractual reasons.

An obviously fictional avatar, animated character, or heavily stylized presenter is less likely to qualify. A voluntary statement that the presenter is virtual can still help the audience interpret the format.

For evidently artistic, creative, satirical, fictional, or analogous works, Article 50 permits an adapted form of disclosure that does not interfere with the display or enjoyment of the work.

Corporate video labels may appear at the beginning, in the end credits, or alongside a particular scene. A note available only in the description may not reach viewers who encounter an embedded player, excerpt, or downloaded clip.

How do common publication scenarios compare?

Classification depends on the actual presentation. An abstract illustration can become misleading when a caption describes it as documentary evidence. A deepfake does not avoid disclosure merely because it appears in advertising rather than a news publication.

How should disclosures be designed and placed?

The disclosure must be available no later than first exposure, noticeable, distinguishable from surrounding information, and accessible. The optional EU icons published by the European Commission (https://commission.europa.eu/) are available for this purpose. Their use is voluntary, and an icon alone does not establish compliance.

For images, a statement can appear in the image or immediately below it. A social graphic may include a small but legible disclosure near the lower edge.

A video label should remain on screen long enough to be perceived. Where only selected scenes are manipulated, an additional scene-specific notice may be appropriate.

Audio can begin with a spoken statement and include equivalent text beside the player or podcast description for accessibility.

Text labels can appear above the article, directly below the headline, or in the byline. A statement available only in general terms or on a separate AI policy page will not identify the particular content.

The wording should match the actual degree of modification. “Created with AI” may be too broad when only the background changed. “Background partially modified using AI” is more informative.

Which sample statements can businesses use?

For fully generated imagery:

AI-generated illustration. The depicted scene is not an actual customer project.

For partially modified photography:

Original photograph partially modified using AI. Added elements were not present in the actual installation.

For product concepts:

AI-assisted concept visualization. The final product may differ in design and equipment.

For a generic synthetic voice:

The narration uses a synthetic voice that does not belong to a real person.

For a cloned voice:

This voice was artificially reproduced with the depicted person’s authorization.

For virtual presenters:

This video uses an AI-generated presenter and synthetic narration.

For unreviewed public-interest text:

This text was generated using an AI system and was not editorially reviewed.

For voluntary disclosure on reviewed text:

This article was created with AI assistance, professionally revised, and editorially approved.

The final statement is not necessarily required by Article 50 when the review exception applies. A company may still adopt it as part of a voluntary transparency policy.

How can an approval process for corporate communications work?

A useful process does not need to send every social graphic to outside counsel. It should ensure that potentially misleading or regulated publications are identified before release.

The first stage records production.
The creator identifies the tool used and whether text, image, audio, or video was fully generated or materially modified.

The second stage evaluates the relationship to reality.
For image, audio, and video, the reviewer asks whether an existing person, object, place, entity, or event is portrayed and whether the asset could be mistaken for authentic documentation.

The third stage evaluates publication purpose.
For text, the business determines whether the publication informs the public about a matter of public interest. Individual communication and routine product copy are treated separately.

The fourth stage assigns editorial responsibility.
A named person checks facts, sources, corporate relevance, and potential misinterpretations. That person decides whether the asset may be released and what disclosure it requires.

The fifth stage preserves technical provenance.
Available metadata, Content Credentials, or other machine-readable marks are retained through export and publication where feasible.

The sixth stage tests the channel.
The company reviews the result on the website, mobile device, social platform, embedded player, and downloaded file. A label visible inside the content system may be cropped or covered after publication.

The final stage records approval.
The company stores the final asset, publication date, responsible person, disclosure decision, and wording used.

Which labeling and documentation checklist should SMEs use?

Content and production

• Was AI used to generate or materially modify the content?
• Is the asset text, image, audio, video, or a combination?
• Was the AI function limited to standard editing?
• Which system and version were used?

Deepfake review

• Does the asset resemble a real person, object, location, company, or event?
• Could the audience treat it as an authentic recording or statement?
• Were real voices, faces, products, projects, or facilities reproduced?
• Is an audience-facing disclosure required?

Text review

• Will the text be published?
• Does it inform the public about a matter of public interest?
• Were claims, sources, and statistics professionally reviewed?
• Is a responsible person or organization identified?
• Is editorial approval recorded?

Technical provenance

• Does the file include machine-readable provenance?
• Does the information survive export and upload?
• Does the provider support Content Credentials, watermarking, or another method?
• Were existing marks preserved rather than intentionally removed?

Publication

• Is a required disclosure noticeable at first exposure?
• Does the wording match the actual modification?
• Does it display correctly on mobile devices?
• Does it remain associated with shared or downloaded content?
• Are accessibility and alternative text addressed?

Records

• Is the final version stored?
• Is the reviewer or editor recorded?
• Is the decision to label or not label documented?
• Is there a review date for the internal policy?

What is the role of the voluntary EU Code of Practice?

The European Commission (https://commission.europa.eu/) published the final Code of Practice on Transparency of AI-Generated Content on June 10, 2026. It addresses both providers of generative systems and professional deployers publishing deepfakes or certain public-interest text.

Participation is voluntary. The statutory Article 50 obligations apply regardless of whether an organization signs the code.

For providers, the code addresses technical marking and detection. For deployers, it provides an implementation framework for label design, placement, and presentation. It also includes optional EU icons for fully generated and partially AI-modified content.

An SME does not need to sign the code before using its methods. The icons are freely available, and the underlying practices can be incorporated into an internal communications policy, approval matrix, and evidence process.

Using an icon alone is not enough. The label’s content, timing, placement, and relationship to the asset must work together.

Why does labeling matter commercially as well as legally?

The disclosure requirement enters an online environment in which confidence in digital media is already under pressure.

Adobe (https://www.adobe.com/) found that 74 percent of surveyed consumers considered it important for retailers to disclose the use of AI-generated content, imagery, or recommendations. Only 26 percent believed brands were meeting that expectation.

Ofcom (https://www.ofcom.org.uk/) reported that 43 percent of surveyed UK adults had encountered at least one deepfake online during the previous six months.

The Reuters Institute Digital News Report 2025, published by the Reuters Institute for the Study of Journalism (https://reutersinstitute.politics.ox.ac.uk/), found that 58 percent of respondents worldwide were concerned about distinguishing real from false online news.

These figures are not a representative survey of German SMEs. They nevertheless show why artificial customer imagery, fictional reference projects, and synthetic endorsements can damage corporate credibility when published without context.

Disclosure does not have to diminish the value of an asset. A term such as “concept visualization” can set appropriate expectations and prevent later disputes. The wording should correspond to the medium and the production method.

How can individual decisions become a dependable approval system?

Businesses should not postpone the labeling decision until moments before publication. It begins with the creative brief, choice of tool, and assignment given to an agency or employee.

A practical foundation can combine an internal policy for generative content, classifications for recurring media formats, approved sample statements, a review matrix, and an evidence repository.

A smaller company may designate one person to review realistic AI imagery. A larger mid-sized organization may involve marketing, the responsible business unit, privacy, information security, and legal specialists depending on the publication.

KrambergAI GmbH (https://krambergai.com/) supports mid-sized companies in establishing these controls within an AI policy and governance foundation. The deliverables can include an AI inventory, labeling rules, an approval matrix, sample disclosures, employee training, and a documented publication workflow.

Which sources support the figures used in this article?

Adobe: 2025 AI and Digital Trends for Retail
https://business.adobe.com/assets/pdfs/resources/reports/retail-digital-trends/2025-ai-and-digital-trends-retail.pdf

Ofcom: A Deep Dive into Deepfakes That Demean, Defraud and Disinform
https://www.ofcom.org.uk/online-safety/illegal-and-harmful-content/deepfakes-demean-defraud-disinform

Reuters Institute for the Study of Journalism: Digital News Report 2025 – Overview and Key Findings
https://reutersinstitute.politics.ox.ac.uk/digital-news-report/2025/dnr-executive-summary

Which resources provide useful further reading?

European Commission: Code of Practice on Transparency of AI-Generated Content
https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content

European Commission: EU Icons for Labeling AI-Generated Content
https://digital-strategy.ec.europa.eu/en/policies/eu-icons-labelling-ai-generated-content

Coalition for Content Provenance and Authenticity: C2PA Technical Specification
https://spec.c2pa.org/specifications/specifications/2.4/index.html

Must every AI-generated text be labeled?

No. Article 50 primarily covers generated or manipulated text published to inform the public about matters of public interest. Even in that category, disclosure is not required when meaningful human review or editorial control occurred and a natural or legal person assumed editorial responsibility for the publication.

Is briefly reading an AI-generated text enough?

A quick read provides limited evidence of editorial control. The responsible person should review important claims, sources, statistics, and possible misinterpretations, make necessary changes, and approve the final version under their own authority. A proportionate record should identify the reviewer, review date, final version, and approval status.

Does an editorially reviewed blog article need an AI label?

Not necessarily. When the article concerns a matter of public interest, the exception for human review and assumed editorial responsibility may apply. Voluntary disclosure remains possible. If the content is produced and published substantially automatically without a person checking its factual claims, Article 50 disclosure may be required.

Must every AI-generated social media image be labeled?

No. Article 50 primarily requires audience-facing disclosure for deepfakes. Abstract artwork and obviously fictional imagery do not automatically qualify. When an image appears to document a real customer project, employee, product, facility, or event, the business should assess whether deepfake disclosure or another explanatory label is needed.

What constitutes a corporate deepfake?

A corporate deepfake can be generated or modified image, audio, or video that resembles an existing person, object, place, entity, or event and could falsely appear authentic. Examples include manipulated project photographs, cloned executive voices, synthetic customer interviews, and artificial footage presented as a real facility or installation.

Is technical metadata sufficient?

No, when an audience-facing Article 50(4) disclosure is required. Machine-readable marks are intended for platforms, software, and detection services. Deepfakes and specified unreviewed public-interest text also require information that people can perceive. The two layers serve different purposes and should be evaluated separately in the publication workflow.

Must the disclosure appear inside the image or video?

Not in every case, but it must be noticeable no later than first exposure and associated with the relevant content. On social platforms, embedding the label in the asset is often more durable than placing it later in a caption. EU recommendations favor methods that remain present when media is reshared or downloaded.

Can a company rely on an automatic platform label?

A platform label can support compliance. The publisher should still verify that it appears after upload, remains visible on mobile devices, and accurately describes the asset. Platforms may strip metadata, convert media, or omit labels from embedded and downloaded files. The company’s operational responsibility does not disappear.

Must a synthetic narrator be labeled?

A generic synthetic voice is not automatically a deepfake. Disclosure becomes more relevant when it resembles an existing person or appears to be an authentic statement from an executive, employee, customer, or expert. Interactive AI phone services are subject to a separate requirement to inform callers that they are communicating with an AI system.

Which records should the company retain?

Useful records include the final asset, publication date, AI system used, responsible person, editorial approval, and labeling decision. For deepfakes or realistic synthetic content, the company should also retain consent records, a description of modifications, the final disclosure wording, and information about technical provenance marks. The depth can reflect reach and risk.