AI transparency and detection

What this page is for

Kontiqo is software that people use to create posts for social networks. Parts of those posts are generated by artificial intelligence: text and images.

Article 50(2) of the EU AI Act requires two things from us: to mark AI outputs in a machine-readable format, and to make sure that marking can actually be detected. This page sets out what we embed and how you can verify it yourself, without having to ask us.

It is addressed to anyone who comes across content created with Kontiqo: readers, platforms, researchers, media and authorities.

Images: what is embedded

Every image generated by Kontiqo receives a machine-readable marking when it is stored, following the IPTC standard for photo metadata. It is invisible and does not alter the image.

We embed an XMP packet with these two entries:

• Iptc4xmpExt:DigitalSourceType = http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia
• xmp:CreatorTool = kontiqo (AI-generated)

The first value is the official IPTC NewsCode for media created entirely by a trained algorithm. It is a public industry standard, not a Kontiqo format.

How to verify it

• With any tool that reads XMP metadata - for example the free command-line program ExifTool, or the file information panel of common image editors.
• Without any extra software: the XMP packet sits in the file as readable text. Open the file in a text editor and search for DigitalSourceType.

You need neither an account with us nor a request to us. That is deliberate.

What this marking does not do

Social networks re-encode uploaded images and handle metadata differently from one another. When we measured this on 1 August 2026, the picture was: on LinkedIn the provenance entry had been removed entirely after upload; on Facebook and Instagram it was still present. **Please do not rely on this.** It is a snapshot, not a promise - networks change their image processing at any time and without notice. We re-check regularly and update this page; if you need certainty, inspect the specific file yourself using the tools described above. All of this happens after we deliver the file and is outside our control. An image without a marking is therefore not proof that it was not AI-generated. And wherever the marking does not reach the viewer, the visible labelling (see the next section) is the only remaining signal.

The marking is not yet cryptographically signed. A signature following the C2PA standard is built but not yet active. We will update this page once it goes live.

Text: currently no marking, and why

We do not embed a machine-readable marking in AI-generated text. That is a deliberate, examined decision, not an oversight.

There are two recognised techniques, and neither is available to us:

• Invisible text watermarks require intervening in how the language model produces text. Our model provider's interface offers neither that option nor a tool with which such a watermark could be detected. We measured this directly against the interface on 1 August 2026 and documented it.
• Metadata can be attached to text as long as it sits in a file. A post, however, is handed to the social network's interface as plain text, and those interfaces have no field for provenance information. A marking would never reach the reader.

Article 50(2) expressly requires technical solutions only as far as they are technically feasible. We reassess that feasibility continuously, at least every six months and whenever something changes about the models or platforms we use.

What exists instead: for every generated text we keep an internal record of which model produced it and when, together with a checksum of the text. We do not store the text itself for this purpose. This lets us answer questions about a specific text - see below.

Visible labelling in the post

Separate from the machine-readable marking is the visible labelling of a post. Under Article 50(4), responsibility for that lies with the person or company publishing the post, not with us as the software provider.

Kontiqo provides its customers with tools for this: a visible AI label that can be burned into images, and a notice line in the post text. Whether they are used is decided by whoever publishes.

Provenance request for a specific piece of content

If you want to know whether a particular text or image came from Kontiqo, write to support@kontiqo.io. Please provide the content as unchanged as possible - text as a copy, images as the original file.

Two limits, so you know what to expect: we delete our provenance records after twelve months, so we cannot match older content. And a text that was edited after generation no longer matches our checksum - in that case we cannot give a definitive answer.

Responsible entity

DIGITALPASSION Ltd., Tassou Papadopoulou 16, House 2, 8270 Tremithousa, Cyprus

Email: support@kontiqo.io

Information current as of 1 August 2026