Do AI images from before August 2, 2026 need a label?

Since August 2, 2026, Article 50 of the EU AI Act applies to stores that show AI images. Many merchants uploaded AI images long before that, and the regulation does not say directly whether those older images now need a label. There is a statement from the European Commission, a reading by a German law firm, and one task that sits with you: in a dispute, you need to be able to show when an image was created.

What does the law say about images from before August 2, 2026?

Nothing that exempts older images. Article 50 of Regulation (EU) 2024/1689 has applied to deployers since August 2, 2026, without a transition period. The deployer (Betreiber) is usually the store itself as soon as it uses a ready-made tool such as Midjourney, Firefly, Canva, or Shopify’s image editor, or instructs an agency to use AI (Guidelines, paras. 12 and 14). Article 50(4) falls on the deployer: the visible disclosure of AI images that show people, objects, places, or events in a deceptively realistic way.

The amending Regulation (EU) 2026/1744 (Digital Omnibus on AI, in force since July 27, 2026) left Article 50(1) to (6) untouched but added a new Article 111(4): providers (Anbieter) of image generators placed on the market before August 2, 2026 have until December 2, 2026 to retrofit the machine-readable marking under Article 50(2). That deadline concerns providers only, and only the marking inside the file, not the visible label the store is responsible for; Article 111 governs systems, not individual outputs. The regulation contains no grandfathering (Bestandsschutz) for images that already exist (both regulations retrieved from EUR-Lex on August 31, 2026).

What does the European Commission say about older images?

The regulation itself does not say that older images need no retroactive label. That statement comes from an FAQ answer by the Commission on the AI Act (wording as of August 31, 2026): “Content generated prior to 2 August 2026 does not need to be labelled retroactively. Nonetheless, the Commission encourages relevant deployers to do so, where possible, since it contributes to the goals pursued by Article 50 of the AI Act.”

• The criterion is the creation date, not the date an image went online.
• It is an authority’s view, not a legal norm. It matches the natural reading that an obligation does not reach back in time without an express provision. Only the text of the regulation itself is binding (as of August 31, 2026).
• The Commission still recommends labeling. If you are going through your catalog anyway, label as you go.

The Commission Guidelines C(2026) 5054 final of July 20, 2026 are non-binding (para. 5) and will be updated (para. 155). Anyone who has not signed the Code of Practice must be able to show how they implement the obligations (para. 148); that takes a record.

How does IT-Recht Kanzlei read the situation?

IT-Recht Kanzlei, a German law firm focused on e-commerce law, updated its article on labeling AI images in online stores on July 31, 2026 (third-party opinion, retrieved September 14, 2026). Its reading: deepfakes created or edited before August 2, 2026 do not have to be labeled retroactively; as an exception, the firm names re-publication (erneute Veröffentlichung).

A second case follows from the Guidelines: if an older image is edited further with AI after August 2, 2026, a new, manipulated output is created under para. 59; the older-image question then no longer arises.

This is a law firm’s opinion, not an authority’s position and not case law. As of August 31, 2026, no court decision on labeling AI images was known in Germany, Austria, or Switzerland. Both sources point the same way; the Commission looks at creation, the firm additionally at publication.

One group never raises the question in the first place: AI color correction, background retouching for design reasons, and upscaling of a real photo do not, as a rule, trigger a labeling obligation under para. 116 of the Guidelines, subject to case-by-case assessment in each instance. For such older images the cutoff question does not arise.

What counts as re-publication in a store?

This is where it gets uncertain. Neither the Commission FAQ nor the Guidelines define “re-publication” for an online store, and the law firm does not spell the term out for store practice, as far as its article shows (retrieved September 14, 2026). Only the yardstick can be derived: Article 50(5) requires disclosure at the latest at the first exposure, and para. 143 of the Guidelines applies the obligation to every output and to every person exposed to it. That yields a working hypothesis, not a ruling:

lakör’s house rule for doubtful cases: with photorealistic images, label when in doubt. The reason: the Wettbewerbszentrale (Germany’s self-regulatory body against unfair competition; guide, version 2 as of July 29, 2026) and the Austrian Schutzverband gegen unlauteren Wettbewerb (position of February 4, 2026) read the deepfake concept broadly, both bodies issue warning letters (Abmahnungen) themselves, and a label takes little effort to add.

Why does the upload date belong in the review record?

Because it is a fact that Shopify keeps, and because it answers the question the Commission FAQ turns on. An image cannot be uploaded before it exists, so an upload date before August 2, 2026 establishes creation before the cutoff. The reverse does not hold: an upload after August 2, 2026 says nothing about the creation date. Then you need other evidence: for marked images the timestamp in the C2PA manifest, otherwise the export file or the tool account.

Where Shopify keeps the date:

• Files (Content > Files): creation date per file, readable through the Admin API as createdAt. The same source keeps the unchanged original with all markers (originalSource; measurement of September 1, 2026: 34 of 34 preserved).
• Collection images uploaded directly on the collection: no entry in Files, so no original and no marker check (measurement of September 10, 2026). The upload date can be read approximately from the version stamp in the image URL. They become checkable once you upload them under Content > Files and pick them from there for the collection.

The date is a fact, not a verdict. A review record should carry “uploaded on May 12, 2026 (before August 2, 2026)” as a column, but not derive a status of “needs no label” from it. Whether the Commission FAQ, the law firm’s exception, or the recommendation to label voluntarily applies to an image is for you or your lawyer to decide. A review record, for example the one kept by the app Easy AI Act Image Labels, therefore captures marker findings, the visibility of labels, and your decisions; it does not establish that images without a marker are not AI images.

A record per image contains:

1. File name and location in the store (product, collection, blog, theme section)
2. Upload date from Shopify, noted as “before” or “after August 2, 2026”
3. Tool, if known (from the marker or from memory)
4. Marker finding: found (type, signer, digitalSourceType) or not found
5. Image type according to the example lists in the Guidelines (paras. 113 to 116): photo in front of an AI background, AI person, AI product image, lifestyle scene, stylized
6. Your decision with date: label, no label, image replaced

What about older images without a provenance marker?

First the reverse case: an existing marker does not replace a label. The machine-readable marking is the provider’s obligation under Article 50(2), and under para. 117 of the Guidelines the store cannot rely on it.

A missing marker says nothing: the file may be a real photo, an AI image from a tool that does not mark, or an AI image whose marker was lost along the way. With older images the last two cases are common:

• Many tools did not mark until recently. In the measurement series of September 1, 2026 (47 image models in the Magnific/Freepik catalog), 34 wrote a marker and 13 did not; Shopify’s image editor (“Generate background”) wrote none. Providers of older tools are required to mark only from December 2, 2026, and only for images created from then on.
• The export path destroys markers. A screenshot never contains metadata, right-click saving from a preview often yields only the web copy without a marker, and a format conversion on download takes the manifest with it (measurement of September 1, 2026).
• The storefront is the wrong place to check. The copies Shopify serves in the store mostly no longer carry the marker (measurement of September 1, 2026: 5 of 34; re-measurement of September 10, 2026: 0 of 3). The marker can only be checked on the original in Files.

If you know that an older image is AI, your knowledge counts, not the marker: enter it in the record and decide as with any other image.

What applies regardless of the cutoff date?

1. The image must show the product as delivered. Article 50(6) of the AI Act leaves other transparency obligations untouched, and § 5 UWG (the German Act against Unfair Competition) covers misleading statements about essential characteristics of the goods, a product image included. An older AI image that shows a product differently from what is delivered was open to challenge before August 2, 2026 and remains so afterward, with or without a label. The Wettbewerbszentrale has run its own AI complaints office with an online form since July 28, 2026.
2. Channels have their own rules. Meta (IPTC and C2PA signals) and TikTok (C2PA) label automatically according to their own statements (as of August 31, 2026). Anything you publish there after August 2, 2026, you check like a new image.

How do you proceed in practice?

1. Pull an inventory: all images with their upload date, with the August 2, 2026 cutoff clearly marked.
2. Work on images from after the cutoff first; the older-image question does not arise there.
3. Split older images into two groups: unchanged in their old place (the Commission’s recommendation applies) and newly published or re-uploaded (label when in doubt).
4. Replace every AI image that flatters the product or shows it differently from what is delivered. As long as it is online, it needs the label (Guidelines, positive list for product images), and the label does not cure the misleading impression under § 5 UWG.
5. Keep a record, with a date per decision. From December 2, 2026, providers of older tools must mark as well (Article 111(4)); that concerns newly created images, and your existing files do not change as a result. For them, your knowledge and the record continue to count.

The legal situation is a few weeks old, the Guidelines will be updated, and case law on older images is not known (as of August 31, 2026). If you keep records today, you can follow a later clarification in either direction.