AI-generated text, images, audio, and video now appear across corporate websites, online stores, news platforms, educational resources, and marketing campaigns. In 2026, labeling this material is no longer merely an editorial preference. Depending on the content, audience, jurisdiction, and role we play in producing or publishing it, AI disclosure may be a legal obligation, a consumer-protection requirement, or an essential trust measure.
An effective label must tell visitors what they are seeing without forcing them to interpret vague terminology. We should place the disclosure where users encounter the content, describe the degree of AI involvement accurately, preserve the label when media is shared, and support visible disclosure with reliable provenance records.
This guide explains how we can label AI-generated content correctly, consistently, and accessibly in 2026.
AI-generated content is material produced or materially altered by a machine-learning system. It can include:
We should distinguish AI-generated content from AI-assisted content.
AI-generated content usually means that an AI system created a substantial part of the final output. AI-assisted content means that a human created the underlying work while AI performed a narrower supporting task, such as correcting grammar, transcribing an interview, removing background noise, or suggesting alternative headlines.
The distinction matters because a label stating “AI-generated” may inaccurately minimize a human author’s contribution, while a vague “AI-assisted” label may conceal that an entire article or realistic image was produced by a model.
There is no single worldwide disclosure rule covering every use of generative AI. We must evaluate the countries in which our website operates, the people we target, the type of content we publish, and whether we are a provider or professional deployer of an AI system.
Article 50 of the EU AI Act applies from 2 August 2026. It establishes transparency obligations for providers and deployers of certain AI systems.
Providers of systems that generate synthetic text, images, audio, or video generally must make the outputs machine-readable and detectable as artificially generated or manipulated. Deployers have disclosure duties for deepfakes and certain AI-generated or manipulated text published to inform the public about matters of public interest. Providers of systems that interact directly with people must also ensure that individuals know they are interacting with AI unless that fact is already obvious. These requirements are summarized in the European Commission’s Article 50 transparency guidance.
A limited transition applies to the machine-readable marking obligation for qualifying generative AI systems placed on the market before 2 August 2026: providers of those systems have until 2 December 2026. Content created before 2 August does not generally require retroactive labeling, although voluntary labeling is encouraged.
For website publishers, two categories demand particular attention:
The public-interest text obligation is subject to an important exception when the content has undergone human review or editorial control and a person or organization holds editorial responsibility. A spelling check or purely procedural approval is not enough. The Commission describes human review as a substantive examination by someone with relevant knowledge and professional judgment.
Creative, fictional, satirical, and artistic works may qualify for a more proportionate form of disclosure that does not interfere with the work’s display or enjoyment. That exception should not be treated as permission to hide a realistic synthetic portrayal.
The United States does not have one universal federal rule requiring every AI-generated webpage to carry an AI label. However, existing consumer-protection, advertising, endorsement, intellectual-property, privacy, and sector-specific rules still apply.
The Federal Trade Commission requires advertising claims to be truthful, evidence-based, and non-deceptive. A disclosure cannot cure an otherwise false claim. The FTC’s Consumer Reviews and Testimonials Rule also prohibits specified fake or false reviews and testimonials, including reviews that misrepresent the experience or existence of the purported reviewer. The rule has been effective since 21 October 2024, as explained in the FTC’s official questions and answers.
We therefore must not create a synthetic customer, generate a testimonial, and assume that adding “AI-generated” makes it lawful. An AI label does not transform fabricated social proof into a genuine customer experience.
Additional state laws may regulate political deepfakes, impersonation, digital replicas, election communications, and synthetic sexual content. Regulated industries—including healthcare, finance, employment, insurance, and education—may carry further disclosure, recordkeeping, or human-oversight duties.
We can create a consistent disclosure policy by classifying content according to the degree and significance of AI involvement.
| Content category | Recommended label |
|---|---|
| Human-written text with spelling or grammar correction | No page-level label normally required |
| Human-created work substantially reorganized or rewritten with AI | AI-assisted and human-reviewed |
| Initial text drafted primarily by AI and substantively verified by an editor | AI-generated draft, reviewed and edited by [name/team] |
| Fully automated text without substantive human review | AI-generated content |
| Illustration created from a text prompt | AI-generated image |
| Photograph materially changed by generative AI | Image materially altered using AI |
| Synthetic narration | AI-generated voice |
| Realistic depiction of a real person saying or doing something that did not occur | AI-generated or manipulated depiction—not an authentic recording |
| Customer-facing chatbot | You are interacting with an AI assistant |
This framework avoids treating every automated task as equivalent. It also tells visitors whether a qualified person has reviewed the material.
A disclosure hidden in a general terms page is insufficient when visitors can encounter the content without seeing that page. We should place labels at or before the first meaningful exposure.
For an AI-generated or substantially AI-assisted article, we should place the disclosure:
A useful label is:
AI disclosure: This article was initially drafted using generative AI. Our editorial team reviewed the claims, verified the cited sources, revised the text, and approved the final publication.
If no substantive human review occurred, we should say so plainly:
AI-generated content: This page was created automatically and has not been independently reviewed by a human editor.
We should not claim that content was “reviewed” when an employee merely clicked an approval button or checked its formatting.
An AI-image disclosure should appear in the caption, immediately beside the image, or as a persistent overlay when confusion is likely.
Appropriate wording includes:
Alt text should still describe the visual content. We should not replace a useful image description with only “AI-generated image.” Where the disclosure itself conveys important information, we can add that information to the accessible name or accompanying caption.
For synthetic video or audio, a description-page notice alone may be missed when the media is embedded, downloaded, clipped, or shared.
We should therefore:
A suitable high-risk label is:
Synthetic media: This video was generated or materially manipulated using AI. It is not an authentic recording of the depicted person or event.
The EU’s optional AI icons may support a disclosure program, but an icon alone does not establish compliance. The Commission recommends that labels be clearly perceivable at first exposure, unobstructed, understandable, and accessible to assistive technologies. Its AI-generated content icon guidance also emphasizes preserving visibility when content is downloaded or reshared.
A customer-facing chatbot should identify itself before or at the beginning of the interaction:
You are chatting with an AI assistant. Its responses may be inaccurate. For account-specific or consequential decisions, contact our support team.
We should repeat the disclosure when a conversation is transferred, resumed, embedded in another interface, or presented through a realistic voice or avatar. Visitors should also have a clear route to human assistance when the interaction involves complaints, payments, health, legal rights, safety, or other consequential matters.
A strong disclosure answers four questions:
Good labels use direct language such as AI-generated, generated using artificial intelligence, or materially altered using AI.
We should avoid ambiguous expressions such as:
“Digitally enhanced” may describe ordinary color correction and does not clearly communicate that a model invented visual elements. “AI-powered” identifies a technology but does not explain whether the final content is synthetic.
The primary disclosure should remain short. We can link it to an expandable explanation or editorial policy containing the model category, review procedure, creation date, responsible team, correction process, and known limitations.
A text disclosure can be implemented as a semantic note:
<aside class="ai-disclosure"
role="note"
aria-label="Artificial intelligence disclosure">
<strong>AI disclosure:</strong>
This article was initially drafted using generative AI.
It was fact-checked, edited, and approved by our editorial team.
<a href="/editorial-policy/ai">Read our AI editorial policy</a>.
</aside>For an image, we can associate the disclosure directly with the asset:
<figure>
<img
src="/images/future-city.webp"
alt="Illustration of a low-carbon city with parks and electric transit">
<figcaption>
<strong>AI-generated illustration.</strong>
This is a conceptual image, not a photograph of an existing city.
</figcaption>
</figure>The disclosure should remain readable with CSS disabled, keyboard-accessible, visible at high zoom, and distinguishable through more than color alone.
A visible label and machine-readable provenance serve different audiences. The visible notice informs people. Provenance data helps compatible systems inspect how an asset was created or modified.
We should not invent unsupported metadata such as:
<meta name="ai-generated" content="true">There is no reason to assume that search engines or compliance authorities recognize a custom tag simply because we add it.
For article structured data, we should identify the real publisher and responsible author or reviewer accurately. Schema.org supports properties such as publishingPrinciples, which can link to an editorial policy, and backstory, which can describe how an article was produced. We must not attribute AI-generated copy to a fictional human author.
For media, Content Credentials based on the C2PA standard provide a more robust provenance mechanism. A credential can record an asset’s origin, edits, cryptographic hash, signature, and AI-related actions. Compatible software can validate whether the credential or associated asset was altered. C2PA’s official explainer also makes clear that provenance does not determine whether content is accurate or trustworthy; it provides evidence about its history.
Because metadata may be removed during compression, screenshots, or platform processing, we should combine:
Google does not prohibit content merely because generative AI contributed to it. Its published guidance focuses on usefulness, originality, reliability, and compliance with spam policies. Generating large numbers of pages without adding value may violate the policy on scaled content abuse, according to Google Search Central’s generative AI guidance.
Google also recommends considering an AI or automation disclosure when visitors could reasonably ask how the content was created. Its people-first content guidance emphasizes explaining the use of automation where users would expect that information.
A disclosure is therefore not a substitute for editorial quality. We should still provide:
A site-wide policy should define when labels are required and who is accountable for applying them. At minimum, we should document:
We should also maintain a content register recording the page URL, asset identifier, creation date, AI system category, responsible editor, review status, disclosure text, and any embedded credentials. This record allows us to audit the website efficiently instead of relying on individual authors’ memories.
Before publishing, we should confirm that:
The most reliable approach is a layered one: a clear visible label, truthful editorial attribution, accessible presentation, machine-readable provenance, and documented human accountability. When we apply those elements consistently, visitors can understand how content was produced and who stands behind it.
Helping real businesses prove they are authentic, transparent, and human-responsible in the age of AI.
Copyright© 2026 Verified Human Created, All rights reserved.