Clarivo – AI Content Disclosure

FAQ

Does Clarivo make my website automatically legally compliant?

No – and be wary of tools that promise this. Whether a piece of content
falls under the labelling obligations of Art. 50 EU AI Act is a
case-by-case decision for the operator (when in doubt, with legal advice).
Clarivo is a technical tool: it finds indications of AI origin, makes
marking and visible labelling easy, and documents the status in the Media
Library. It cannot and does not want to take over the decision itself or
the legal assessment.

What do I have to label as a website operator?

The EU AI Act splits the transparency obligations across two roles. As a
website operator you must visibly disclose AI or deepfake content for
image, audio and video to your visitors (Art. 50(4), from 2 August 2026).
What is required is the visible disclosure – not that you yourself write a
machine-readable note into the file. That is exactly what Clarivo provides
the visible labelling for, across all media types: badge/overlay on the
image, notice line or overlay on the video and audio player, optionally
with the official EU icons. Note: Clarivo is a technical tool and is no
substitute for legal advice – whether a specific piece of content falls
under the obligation is for you to decide, when in doubt with legal advice.

Why doesn’t Clarivo write a note into video/audio files?

Because machine-readable marking of the file itself is the job of the AI
provider (Art. 50(2)) – that is, the AI system that creates the content
(ChatGPT, Gemini, Veo, ElevenLabs & co.), not the website operator. Your
operator obligation is the visible disclosure (Art. 50(4)), and Clarivo
covers that at the player. Clarivo deliberately does not take on the
provider’s job and therefore does not alter video/audio files. For images,
Clarivo additionally offers the machine-readable IPTC/XMP note as a
voluntary bonus (it also survives downloading the image file) – that is
not an operator obligation either.

Why doesn’t the scan detect my AI image? (The clipboard trap)

Automatic detection reads the file’s metadata. Anyone who takes an image
via “copy image”, through the clipboard or as a screenshot loses ALL
metadata in the process – automatic detection is then no longer possible.
For that reason, always download AI images as the original file
(“Download” or “Save as”) and upload that file to the Media Library. Some
image editors and “optimisation” plugins also strip metadata on save.
Undetected images can be marked manually in the Media Library’s attachment
dialog.

Which image generators are detected automatically?

Tested with freshly created images, each downloaded as the original file
(as of June 2026):

• Google Gemini: yes (C2PA manifest, source “Google (Gemini)”).
• ChatGPT / GPT-Image (OpenAI): yes (C2PA manifest).
• Midjourney: yes (XMP note in the PNG file; source attribution
  generic, as the file contains no manufacturer name).
• Seedream (ByteDance): yes (C2PA manifest, BytePlus ModelArk).
• Recraft: no – the export file contains no metadata; please mark
  images manually.
• DALL-E (classic), Adobe Firefly, Stable Diffusion, FLUX, Ideogram,
  Leonardo AI, Canva AI, Bing/Copilot Designer, Grok: not yet tested.
  Many of these providers write C2PA or XMP metadata that the scan detects
  generically – but a verified test result is not yet available.

Important: a C2PA manifest alone does not mean “AI” – for example, Pixel 10
smartphones embed C2PA in every normal camera photo. Clarivo therefore
evaluates the digitalSourceType and correctly classifies genuine camera
photos as “no finding” (verified with a test file). Conversely: providers
can change their export behaviour at any time – a missing finding is no
proof that an image is not AI-generated.

Which file formats are supported? (Format matrix)

Principle: the machine-readable in-file note (IPTC or XMP in the original
file) exists exclusively for images. Video and audio files are labelled
visibly only – the media files themselves remain unchanged. Manual marking
and visible labelling work for all media types.

• JPEG: scan yes · in-file note yes (IPTC)
• PNG: scan yes · in-file note yes (XMP)
• WebP: scan yes · in-file note yes (XMP)
• GIF: scan yes · in-file note no
• AVIF: scan yes, with a limitation – the scan reads the first 512 KB
  of the file; if the metadata of a large AVIF file lies beyond that
  (behind the image data), it is not found. Rare in practice; when in
  doubt, mark manually. In-file note no.
• HEIC/HEIF: WordPress 6.7 and later automatically converts HEIC
  uploads to JPEG – the JPEG file is scanned and, where applicable, given
  a note.
• MP4 / MOV (video): scan yes (C2PA manifest in the container, e.g.
  Google Veo 3) · no in-file note
• M4A (audio): scan yes (same container as MP4) · no note
• MP3: scan yes (AIGC note in the ID3 tag, e.g. MiniMax; C2PA/XMP in
  ID3 provenance frames) · no note
• WebM, WAV, OGG, FLAC, MKV, AVI …: no evaluable provenance metadata
  standard (as of 06/2026) – the scan honestly returns “no finding”,
  manual marking possible at any time.

Can Clarivo display the official EU labelling icons?

Yes, optionally. In June 2026 the EU Commission published official icons
for labelling AI-generated content together with the voluntary Code of
Practice (three categories: fully AI-generated, AI-edited, basic “AI”).
On the “Images” tab, under “Display” you can choose the “Official EU icon +
text” variant – the free-text badge remains the default. The matching icon
is chosen automatically based on the detected content (overridable per file)
and appears next to the label text; the choice applies to image, video and
audio. The icons are provided by the Commission for free use without
attribution.

Important and honest: the visible icon is part of your web page, not of the
media file – it does not travel along when the file is downloaded or shared.
What survives a downloaded image file is solely the optional
machine-readable IPTC/XMP note (images only). Video and audio files receive
no in-file note. And: Clarivo uses the official EU icons – on its own that
does not establish legal compliance and is no substitute for legal advice.

What can video/audio detection do – and what can’t it?

• Read limits: the metadata areas of the container are evaluated
  (C2PA/XMP boxes in MP4/MOV/M4A, ID3 tag in MP3). Very large manifests are
  only read up to 512 KB; for heavily fragmented streaming MP4s
  (fMP4/CMAF) the search ends after 1024 container sections. Not read are
  the very old ID3v2.2 format and XMP in the moov/udta atom (older Adobe
  pattern).
• No detection without metadata: converted or trimmed files lose their
  provenance metadata; many AI providers do not supply any in the first
  place.
• Embedded content (YouTube, Vimeo, Spotify & co.): embeds show content
  from external platforms that does not reside in your Media Library –
  Clarivo can neither read metadata there nor attach a label to the player.
  Alternatively, label the post via the “Contains AI-generated content”
  checkbox or place the “AI notice (Clarivo)” block above the embed.
• Fullscreen mode: in native fullscreen the browser shows only the
  video element itself – the overlay badge cannot be rendered there
  technically; in the normal page view it remains visible.
• Offload/CDN plugins (S3, Cloudflare R2 & co.): if a plugin serves
  Media Library files from an external domain, Clarivo no longer recognises
  the player URL of the classic [video]/[audio] shortcode as a Media
  Library file – no label appears there. The editor’s video/audio blocks
  are not affected.
• No in-file note: video/audio files are never written to – the
  machine-readable note remains reserved for images.

Which AI plugins in WordPress are detected?

If an AI plugin creates images directly in WordPress, Clarivo marks them at
the point of creation – even when the file itself carries no metadata:

• AI Engine: tested for real. Detected are the image generator and
  chatbot images that are saved in the Media Library. NOT detectable is the
  editor command “generate image” – it leaves no provenance data.
• AI Power: covered on the basis of the plugin source code (all Media
  Library saves).
• Elementor AI: covered on the basis of the plugin source code – all
  images from the Elementor AI dialog, alongside generated ones also
  AI-edited images (“remove background”, “expand image”).
• WordPress AI plugin (“AI”, wordpress.org): covered on the basis of
  the plugin source code, checked with versions 0.9.0 and 1.0.1 – its own
  image interfaces record the provenance (ai_generated meta) on save. NOT
  detectable are images that third-party code delivers directly via the
  plugin’s import interface without passing the note along.
• WordPress 7.0 core AI (AI Client/Connectors): an honest gap – the
  core substructure itself leaves no provenance note on save. Safety nets:
  the upload scan catches it if the AI provider embeds metadata in the file,
  and third-party plugins can use the open interface.

Existing images from AI Engine, AI Power and the WordPress AI plugin are
also detected after the fact by the Media Library scan based on the plugin
notes in the database; Elementor AI existing stock cannot be re-detected
technically. Other AI plugins can declare images via the open interface
(action clarivo_mark_ai_generated with the attachment ID).

How do I also place the AI notice directly in the chat widget?

The Clarivo banner sits as its own layer above the page and does not
interfere with third-party widgets. Anyone who also wants to show the
notice INSIDE the widget configures it directly at the provider (labels as
of June 2026, provider interfaces may change):

• Tidio: in the Tidio panel, edit the welcome message and add a
  sentence such as “You are chatting with an AI assistant”.
• Crisp: in the Crisp console, adjust the bot scenario or the welcome
  text; additionally you can rename the chatbox title (e.g. “AI assistant”).
• Chatbase: extend the agent’s “Initial Messages” with the notice and
  choose the display name accordingly.
• Other widgets: almost all chat providers have a configurable welcome
  message – the AI notice belongs at its start.

The provider’s built-in disclosure is thereby supplemented, not replaced.

I changed the alt text in the Media Library – why does the post show the old one?

That is WordPress behaviour, not a Clarivo feature: the image block stores
the alternative text fixedly in the post when inserting – later changes to
the Media Library alt text have no effect there. Maintain the text for
existing blocks in the block sidebar; featured images, by contrast, always
use the current Media Library alt text. Clarivo’s optional alt-text
addition (“AI-generated image”) is appended at runtime and does not change
the stored alt texts.

Does the plugin send data to external servers?

No. All scans run locally on your server, there are no external HTTP
requests, no tracking and no forced account.

What happens on uninstall?

When the plugin is deleted, all plugin settings and caches are removed. The
markings on your media and posts (AI status, source, detection method) are
deliberately retained – they are your documentation and are immediately
usable again after a reinstall. Anyone who also wants to remove the
markings completely sets the following in wp-config.php before deleting:
define( ‘CLARIVO_UNINSTALL_PURGE_META’, true );
IPTC/XMP notes already written into image files are not touched on uninstall
(the plugin does not change any files without an explicit instruction); they
can be reverted beforehand per image by removing the marking.