Asked briefly, answered honestly.

I do not recognize AI images by how they look, I read evidence: provenance data that many AI tools write into the original file when saving it, namely C2PA / Content Credentials, IPTC DigitalSourceType, and XMP. If I find such a marker, I propose the matching EU label. Not every tool writes a marker, so I do not find every AI image; you label the rest by hand.

A machine-readable provenance marker in the original file. The tool that created or edited an image writes it in. It is a claim made by the creator: I read it, but I do not verify the signature of a C2PA manifest. On the back of the image card I show you, for every finding, the marker and the tool that wrote it.

Evidence is a marker in the original that the creating tool wrote. A hint is the output of a detector that estimates a probability from image patterns. I deliver only evidence, no hints. How all methods are classified is explained in the detection taxonomy.

Detectors estimate a probability from pixels; the result is a hint. A marker sits in the file itself and comes from the tool that created the image. That way every proposal I make rests on data you can look up on the back of the image card. I do not use a probability detector, and I do not read invisible watermarks from providers.

Because the copies your store serves usually no longer carry the marker. lakör measured: on September 1, 2026, 5 of 34 markers survived storefront delivery, while all 34 were readable from the stored original. So I read the unchanged original from your file library (Content > Files), never the delivered image.

Some tools write no marker at all; with others it gets lost on the way, for example through a screenshot or a format conversion during download. Images you know to be AI-generated you label by hand under “AI detection.” That is not a fallback but the second main path.

There is no original of this image that I can read, usually because it was uploaded directly on a collection and is not a file under Content > Files (measured September 10, 2026). I read no marker there and do not claim “No AI evidence” either; the image stays open, and you decide. If you upload it under Content > Files and pick it from there in the collection, I can check it.

I do not find every AI image, I do not judge whether an image requires a label, and I only reach your Online Store. I do not use a probability detector, and I change nothing in images, products, or theme. I name these limits before you install so that you know what stays with you.

No, and no app can. I read provenance markers in the original (C2PA, IPTC, XMP); an AI image without a marker looks like a photo to me. I show you what I can identify with certainty, and you add the rest.

I found no marker in the original. That is all it means: a real photo, an AI image without a marker, and an AI image that lost its marker all look the same to me. A missing marker says nothing about the image, and the responsibility for labeling stays with you, even with automation.

I do not know that, and I say so. I only find data that the generating tool wrote into the file: of 47 image models lakör measured on September 1, 2026, 13 wrote nothing. An empty report means “I found nothing,” not “you have no AI images.”

Either the tool wrote none, or the export path destroyed it. A screenshot never contains provenance data, and a format conversion during download destroys the manifest (lakör measurement, September 1, 2026). Right-click saving from a preview, compressors, and messengers were not measured but are expected to behave the same. AI images stay readable this way: official download button, native format, nothing in between.

Adobe Firefly writes no manifest into the file, only a pointer to a manifest in Adobe’s cloud (measured September 1, 2026). I do not follow that pointer, so Firefly images appear as “No AI evidence” (as of September 16, 2026). Label them by hand.

Images your theme hard-codes, and image sources owned by third-party app blocks. Images from metafields or from product and page text I do check if they are in your file library; I cannot tell where they are used, so they appear as “Not used.” Only the published theme counts as usage (as of September 7, 2026).

Outside your Online Store: on marketplaces, in social media, in newsletters and e-mails, and in headless storefronts. On top of that, images that the theme loads as a CSS background. There, you add the label yourself.

No. I do not touch images, alt texts, products, or theme code; I write only two app-owned settings fields to your store. The label is an overlay that an app embed shows in your storefront and that you turn on and off in the theme editor.

Wherever you turn it on under “Display and placement”: product pages, collections, home page, search, and cart, and depending on the theme also blog, header, and footer. You choose the corner and size.

No. In the app, the button “Open theme editor” takes you straight to the “Easy AI Act Image Labels” block under App embeds. Toggle it on, save. Then I check myself whether it worked and show “App embed active in your theme.”

With automation, I check new images once a day and, after product changes, within about 15 minutes. Images with a readable marker get the label, images without a marker I close as “No AI evidence” without a label, and “Not verifiable” stays open for you. Automation is included in Pro and Unlimited, it is off by default, and every decision it makes can still be changed. In Basic you start the scan yourself.

Corner, size, and language per location, under “Display and placement.” Not the icon, not the color, not the wording: I use the official EU icons with the abbreviation AI, as provided for in the EU Code of Practice on labeling; the Code names the top right corner of the image as the standard position, which is why it is my default. The label carries no lakör logo.

No. Every theme has its own app embed settings. After a theme switch I show you a notice with the link to activate it again.

Read-only: products, files, theme settings, and the languages of your store. I do not read customers, orders, or payments, and I store no images. What I store where and for how long is on Trust and security.

The label disappears from your store because it was only an overlay. Images, products, and theme stay as they were. I delete the access token immediately, and all other data of your store with the deletion request that Shopify sends 48 hours after uninstall, no later than 30 days after uninstall.

Fixed: one language for all labels, English by default, with the official files of the European Commission (“AI GENERATED,” “AI MODIFIED”). Dynamic, from Pro: the label follows the language your customer is currently viewing the store in, in 24 EU languages. The abbreviation “AI” stays in every language. The non-English versions are translations by lakör, not official files, and not reviewed by any authority.

The regulation prescribes no specific language but a disclosure your audience clearly understands (Article 50(5) AI Act, Guidelines para. 142, as of August 31, 2026). Whether English is enough for your customers or a national language law applies is your decision for your target markets.

Since August 2, 2026, Article 50(4) of the EU AI Act requires deployers to visibly disclose AI-generated or AI-manipulated images that show authentic-looking persons, objects, places, entities, or events. The deployer is whoever operates the AI tool, which is usually your store. The guide explains the criteria.

I read the originals of your images for AI markers, propose the matching EU label for each match, show confirmed labels visibly in your store, and, with automation from Pro, check new images daily. Images without a marker you label by hand. Whether an image requires a label and whether your implementation is sufficient is for you or your lawyer to assess: I am a tool, not a legal opinion.

No. Whether an image must be disclosed under Article 50(4) AI Act depends on whether it shows real persons, objects, places, or events in a deceptively authentic way, not on whether AI was involved (Commission Guidelines C(2026) 5054, para. 113). You or your lawyer decide; I deliver the marker finding and the check log.

According to a European Commission FAQ, content from before August 2, 2026, does not have to be labeled retroactively, but the Commission expressly recommends that deployers do so (the Commission’s view, not a legal rule, as of August 31, 2026). The text of the regulation itself knows no grandfathering. From December 2, 2026, providers of older AI tools must also write provenance data; a new scan after that date can bring new matches.

In Germany, fines under the AI Act are imposed by the Federal Network Agency (Bundesnetzagentur) (Section 2(1) and Section 17(1) KI-MIG); for small and medium-sized enterprises the lower amount of the range applies (Article 99(6) AI Act). The more likely risk is warning letters (Abmahnungen) under the German Act against Unfair Competition (UWG); the Wettbewerbszentrale has run its own AI complaints office for this since July 28, 2026. As of September 14, 2026, no court decision and no documented warning letter for missing AI image labels is known.

Three plans by the number of images in your store, billed through Shopify (as of September 14, 2026): Basic at EUR 4.90 per month for up to 100 images, Pro at EUR 9.90 per month for up to 1,000 images, Unlimited at EUR 19.90 per month with no image limit. From Pro, the label follows your customer’s language, and automation checks new images daily. On the yearly plan you pay 20% less. All details are on the pricing page.

7 days free, once per store, for the plan you start with; if you switch during the free trial, the remaining days carry over. Development stores use the app for free at the Unlimited scope.

Your quota covers the oldest images in your store by upload date. I do not scan newer images outside it and show no label for them, even if you have already labeled them; they carry the note “Outside quota,” and your decisions stay saved. For these images, the disclosure stays with you until you switch to the next plan.

Yes, in the app or by uninstalling, effective at the end of the current billing period. On the yearly plan the paid year runs to its end, and there is no prorated refund. After your plan ends, your labels stay visible for another 7 days.