What is produced with AI on this page
We think it is right to say this clearly from the start: all of the body text on this site is written with an AI language model. That covers the buying guides, the knowledge articles, the data checks on individual devices, the brand and watch brand pages, and also this page that you are reading now. Anyone reading our texts should know how they come about before basing a purchase decision on them.
What does not come from the AI is the substance. The choice of topics, the selection of products, the rating logic, the research into calibre data and safety standards, and the decision on which recommendation is ultimately made all come from people. The AI is the tool for the wording, not the authority that decides what is correct.
The figures do not come from the AI either. Rotation values, power reserves, resistance grades of safes and technical details of devices come from manufacturer specifications, technical data sheets, standards and the manufacturers' settings databases. Where we cannot substantiate a figure, we say so instead of inventing a number.
How the editorial review works
Between the AI draft and publication there is a fixed review process, and it is not a fig leaf. Technical details are checked against manufacturer data and data sheets, product availability and purchase links are checked via the retailer's programme interface, and ratings are checked for contradictions between text and score. Only then does a page go live.
In addition there are fixed editorial rules that every text must follow. We do not claim measurements we cannot substantiate, we do not state current daily prices but only price ranges, and we expressly advise against a device if it is unnecessary for the situation described. These rules are set out in detail in our section on assessment methodology.
Editorial responsibility lies with Seydhan Cakmak as a natural person, not with a tool and not with an anonymous editorial abbreviation. That is the decisive point legally, and it is also the honest one: if something here is wrong, a human being is responsible for it and corrects it. We accept reports on this via the contact page.
Where the images come from, and why none of them is an AI image
We do not create images with generative models. All product images on this site are manufacturer and retailer photos from the Amazon catalogue and are embedded directly from there, without our altering them. So we show the image that also appears on the retailer's product page, not a reconstructed depiction.
One limitation belongs here, and we state it instead of leaving it out. For some time now, Amazon and the retailers selling there have been using generative tools in their listings themselves, for example for backgrounds or image variants. From the outside, we cannot check for an individual catalogue image whether and to what extent this has happened, and we therefore claim neither one nor the other.
The only image we produce ourselves is the preview image for shared links. It is created by a browser outputting an HTML template with our logotype as an image file. This is pure rendering by program code, involving no generative image model and no training data. Under the IPTC provenance code classification, this is purely algorithmically generated graphics and expressly not an AI image.
How we label: the two labels
We work with exactly two labels so that the labelling remains understandable at a glance. The label AI text appears on every content page directly below the heading, that is, where you see it at first glance at the page and not only after scrolling. It means: this text was created with AI and then checked and approved by a human being.
The label AI image works differently, because the risk of deception is greater with images. It is not shown next to the image or in a footnote but visibly in the corner of the image itself, so that it is present immediately when viewing. As of today, not a single image on this site carries this label, because we do not use AI images.
We deliberately decided against labelling all images as AI images as a precaution. A label that makes a blanket claim that is not true for the individual image is not transparency but a false statement, and it devalues the label precisely where it will one day really be needed. The mechanism is fully built in and takes effect without further changes as soon as an AI image is added.
What the EU AI Act requires from 2 August 2026
Article 50 of the AI Act regulates the transparency obligations for AI content and has applied since 2 August 2026. For texts, the disclosure obligation applies, according to its wording, when they are published in order to inform the public on matters of public interest. For images, the focus is on the deepfake, that is, the deceptively realistic depiction of persons or events.
At the same time, the regulation contains an express exception for texts. The obligation does not apply if the content has undergone human review or editorial control and a natural or legal person bears editorial responsibility. Exactly this situation applies here, because every text is checked and approved before publication.
Where the labelling takes place is also specified. It must be clearly recognisable, be present at the latest at the first perception of the content, and be accessible without barriers. For texts it therefore belongs at the beginning and not at the end, and for images visibly on the image or directly next to it. We have aligned our placement with exactly this.
Why we label even though the obligation is disputed
A buying guide for watch accessories is not reporting on matters of public interest in the narrower sense, and the exception for editorially reviewed content also applies here. By both routes there is much to suggest that we would not be obliged to label at all. That is, however, a question of interpretation on which there is as yet no established case law.
We label anyway, for two reasons. The first is simply decency towards our readers: anyone who bases a purchase decision on our recommendation has a right to know how that recommendation came about. The second is practical, because a label that is not needed costs nothing, whereas a missing label can become expensive in a dispute.
There is also a third point that matters to us. Some of our knowledge articles touch on topics involving value and safety, such as the resistance grades of watch safes, the anchoring of a safe, or the basics of insuring a collection. It is precisely there that the standard for disclosure is higher, and we do not want to make it depend on how a standard is ultimately interpreted.
What we expressly do not do with AI
We do not invent experience reports. There is no sentence on this site that claims practical trials that did not take place, and we do not generate user voices, star ratings or averages from feedback that does not exist. Where we give a rating, it is labelled as an editorial rating and recognisable as a reasoned assessment.
We also do not generate product photos or depictions of devices that do not exist in that form. A generated product image could show details that look different on the real device, such as a holder or a control unit, and that is exactly what would mislead a purchase decision. We therefore show only the catalogue images of the suppliers.
Finally, we do not let the AI decide on the product selection. Which devices are included depends on availability, verifiable technical data and fit with the topic in question, and this decision is made by the editorial team. Commission levels play no role in this, as described in the section on funding on the About us page.
Machine-readable labelling
In addition to the visible labelling, our pages also carry the information in structured form so that search engines and AI systems can evaluate it. For this we use the digitalSourceType property from the Schema.org vocabulary, which since early 2024 has been based on the IPTC provenance codes. It appears in the article and review nodes of the respective page.
We state one limitation openly here too. This markup is standards-compliant, but it does not lead to any visible indication in search engines, because the evaluation of image provenance data there is tied to metadata embedded in the image file. As our product images sit on third-party servers, we cannot set such metadata.
For the same reason, we do without some constructions that are occasionally recommended online. Entering an AI tool as author or creator would be impermissible under the vocabulary, because it provides only for persons and organisations, and it would blur the human responsibility that we want to document. The author node therefore stays with the responsible person.
Found an error or doubt a statement?
Transparency is only worth as much as the ability to verify it. If a statement seems wrong to you, a source is missing, or an image does not match the device described, write to us via the contact page. We check every report and correct substantiated errors promptly.
This expressly applies to this page itself as well. Should our assessment of the legal position turn out to be outdated, for example through a ruling or official interpretive guidance, we will adjust the labelling and note that here. The date at the top of this page shows you when we last revised it.
If you want to know who is behind this site and the rules we use to rate products, you will find it on the About us page. The legally required details are in the Legal notice, and the information on data processing is in the Privacy policy.