Part images rarely reach a listing unedited these days. The background is removed, serial numbers are made illegible, a new part is placed in a workshop scene, or a view is generated outright because no photo exists. Since 2 August 2026 this raises a legal question in a new form: what has to be labelled? The answer sits on two levels — the transparency obligations of the AI Act and the ban on misleading practices in unfair competition law and marketplace rules. How good part images are taken is covered in Part photography: how images sell used parts; this article deals only with labelling.
Two obligations, two addressees
Article 50 of the AI Act, Regulation (EU) 2024/1689 distinguishes two roles. Under Article 3(3), a provider develops an AI system or has it developed and places it on the market or puts it into service under its own name or trademark. Under Article 3(4), a deployer uses an AI system under its authority, except in the course of a personal non-professional activity. A dealer, workshop or shop that uses a third-party image tool in its business therefore meets the definition of a deployer. Each role carries its own obligations:
| Art. 50(2): marking | Art. 50(4): disclosure | |
|---|---|---|
| Who it applies to | Providers of AI systems generating synthetic audio, image, video or text content | Deployers of an AI system that generates or manipulates image, audio or video content constituting a deep fake |
| What is required | Outputs marked in a machine-readable format and detectable as artificially generated or manipulated | Disclosure that the content has been artificially generated or manipulated |
| Which content | All synthetic outputs, unless an exception applies | Only deep fakes within the meaning of Article 3(60) |
| Exceptions | Assistive function for standard editing; no substantial alteration of the input data provided by the deployer or its semantics; law enforcement | Law enforcement; limited disclosure for evidently artistic, creative, satirical, fictional or analogous works or programmes |
| Form and timing | Technical solutions that are effective, interoperable, robust and reliable as far as technically feasible | Clear and distinguishable, at the latest at the time of first exposure, and accessible (Art. 50(5)) |
| Application | Since 2 August 2026; for systems placed on the market earlier, by 2 December 2026 | Since 2 August 2026, with no transition period |
When a part image is a deep fake
The term suggests forged videos of politicians, but the definition is broader. Under Article 3(60), a ‘deep fake’ means “AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful”. Objects are expressly included. The Commission's guidelines stress that the assessment is objective: no intention to deceive is required; what counts is resemblance, message, deployment context and audience expectations. They expressly list as a deep fake an AI-generated image of a product in advertising or on packaging that can mislead as to the product's actual appearance, characteristics or use. Not a deep fake, by contrast, is a real product shown against an AI-generated background, as long as the advertisement is not likely to mislead about the product's actual representation, characteristics and use.
- Used-part listing. For a single item, buyers expect a photo of exactly that item. A generated image without a notice appears as a photograph in this context — precisely the constellation the definition describes.
- New part in a catalogue. A generated image of a series part resembles an existing object. Whether it falsely appears authentic depends on whether it looks like a photograph and whether shape, colour or markings deviate from the real part.
- Only the background is new. According to the guidelines, adjusting or replacing backgrounds for clearly aesthetic purposes is likely to have only a minor impact on whether a product is perceived as authentic — provided the part itself stays unchanged.
- The part is embellished. If it looks cleaner, more complete or of higher quality than in reality, that matches the guidelines' own example: a product made to appear not identical to the real one, more appealing or of improved quality.
The core question: individual item or depiction?
| Edit | Classification under Article 50 | What follows for the listing |
|---|---|---|
| Background removed, pixels of the part unchanged | The guidelines list deleting and obscuring backgrounds that are visible in the original file as an example of standard editing, to which paragraph 2 does not apply | Remains a photo of the individual item. Defects must not disappear in the process |
| Serial numbers and codes made illegible | Intervention in individual image areas; the object itself remains the photograph. The guidelines list pixelation or blurring of faces as a minor alteration but do not mention identifiers expressly. Classification depends on the case | Notice that identifiers were anonymised — see Anonymising identifiers on part images without losing the OE number |
| New part placed into a scene | For paragraph 2, the guidelines cite composite images that modify the representation of objects as requiring marking. For the deep-fake question, background changes for aesthetic purposes and arrangements of existing products in advertising are likely to have only a minor effect | Label it if the image gives the impression of a real photograph or shows the part altered |
| View generated from reference images | Synthetic content under paragraph 2; a deep fake under paragraph 4 if it appears to be a photograph | Label visibly as a depiction; for a used part never in place of the photo |
| View generated without a reference, from model knowledge | As above; here the closeness to the real part is least supported by evidence | Only sensible for new parts, always labelled as a depiction |
Misleading practices: the obligation that applies regardless of AI
Article 50 governs whether content is recognisable as artificial. Whether a listing misleads is still judged under unfair competition law; the Commission's guidelines themselves stress that the deep-fake criterion is to be understood independently of the concept of deception in the Unfair Commercial Practices Directive. For listings in Germany the relevant statute is the Act against Unfair Competition (UWG). Under Section 5(1) UWG, anyone who engages in a misleading commercial practice that is likely to cause consumers or other market participants to take a transactional decision they would not otherwise have taken acts unfairly. Under subsection (2) no. 1, misleading practices include untrue or otherwise deceptive statements about essential characteristics of the goods such as their type, design, accessories and nature. Section 5a UWG covers misleading by withholding essential information. An image can therefore satisfy the AI Act and still mislead.
For a used part, the condition of the individual item is the essential characteristic. A generated image shows no condition; it shows a part that is not lying on the bench in that form — without the scratch, the missing tab, the corrosion. A notice reading “AI-generated” does not turn that depiction into a statement about condition. Anyone offering a used part with a depiction alone leaves open the very question the image is meant to answer.
How to label
- At the image, not in the small print. Where a disclosure obligation exists, Article 50(5) requires the information to be provided in a clear and distinguishable manner at the latest at the time of first exposure. The Commission's FAQ cite visible labels as an example.
- In wording that names the difference. “AI-generated depiction – not a photo of the part offered” tells the buyer more than a bare “AI”.
- In the caption or listing text where text in the image is prohibited. Where the marketplace does not allow text inside images, the caption or the description remains.
- With a flag in your own item master. A field separating photographed from generated images prevents a later export from presenting a depiction as a photo.
- Without removing the provider's marking. The machine-readable marking under paragraph 2 is applied by the provider of the system. If it sits in the metadata, it is lost when the export discards the metadata.
What the tapinomahub API returns for this
The image services of the tapinomahub API distinguish these cases in the result itself. The following fields are part of the published contract; the complete schemas are described in the developer documentation.
| Call | Fields in the response | Meaning |
|---|---|---|
POST /vision/part/remove/bg | sourcePixelsPreserved, limitations | sourcePixelsPreserved: true confirms that the pixels of the object come from the submitted photo. The response contains no separate labelling field |
POST /vision/identifiers/redact | per image in assets[].provenance: assetType, synthetic, requiredLabel, disclosureEmbedded | Classification of the result, the intended notice text and the statement that a notice is embedded in the file |
POST /vision/part/composite | asset.provenance with the same fields, plus partIdentityPreserved | Classification, notice text, embedded notice and the statement that the inserted part is preserved |
GET /vision/part/generation-jobs/{jobId} | in result: synthetic, disclosure, basis, and per view in views synthetic and partIdentityVerified | Result of a job started via POST /vision/part/generate; synthetic is always true there |
For generation, the source is part of the order. The default for sourceMode is references_required; generation from model knowledge alone is ordered explicitly with knowledge_only. The result states in result.basis whether it rests on reference_images or on knowledge. Views that carried unsupported markings appear in result.coverage.skippedAngles with the reason generated_view_carried_unsupported_markings instead of in the image set, and result.finish.colorAppearanceIsAuthoritative is always false: the colour impression of a depiction is not a colour specification. The whole image path up to the listing is described in From part photo to listing: the image route through the tapinomahub API.
This article reflects the legal position as of 13 September 2026 and is not legal advice. The classifications rest on the text of the regulation and the Commission's guidelines; in the individual case, the context of the listing decides. Where generated images are used in advertising, a legal assessment of the specific case is the right step.
Sources and legal references
- KI-Verordnung (EU) 2024/1689, Art. 3, 50, 99 und 113
- Digital-Omnibus-Verordnung zur KI (EU) 2026/1744
- Europäische Kommission: Leitlinien zu den Transparenzpflichten nach Art. 50
- Europäische Kommission: Praxisleitfaden zu KI-erzeugten Inhalten
- Europäische Kommission: FAQ zu Art. 50 KI-Verordnung
- Gesetz gegen den unlauteren Wettbewerb, § 5
- Gesetz gegen den unlauteren Wettbewerb, § 5a
- eBay: Grundsätze zu Artikelfotos
- eBay: Picture policy
- eBay : Règlement sur les photos
- tapinomahub API-Dokumentation
Frequently asked
Does a cut-out part image have to be labelled as AI-edited?
According to the Commission's guidelines, deleting backgrounds that are visible in the original file is standard editing, to which the marking obligation under Article 50(2) does not apply. As long as the part itself stays unchanged, the image still shows the item on offer. The assessment of the individual case remains with the business.
May I use an AI-generated image for a used part?
Not in place of the photo: it does not show the condition of the item on offer and can mislead about its nature. eBay prohibits stock photos for used items and photos that do not accurately represent the item. Whether a supplementary image clearly labelled as a depiction is permissible depends on the individual case and on the rules of the marketplace concerned.
Since when have the labelling obligations applied?
Article 50 has applied since 2 August 2026. Only providers of generative systems placed on the market before that date have until 2 December 2026 for machine-readable marking under paragraph 2. There is no transition period for disclosing deep fakes.
Who has to label – the provider of the image tool or the dealer?
Both, each something different. The provider of the AI system ensures machine-readable marking; the deployer discloses deep fakes. Whoever publishes the listing is primarily responsible for misleading content and compliance with marketplace rules.
How can I tell from the API response that an image was generated?
For a generation job, synthetic: true appears in the result and in every view, together with disclosure and basis. For POST /vision/part/composite and POST /vision/identifiers/redact, the provenance object of the result image holds the classification in synthetic and the notice text in requiredLabel. For cut-outs, sourcePixelsPreserved confirms that the pixels come from the photo.
What fines does the AI Act provide for?
Under Article 99(4), infringements of the transparency obligations in Article 50 are subject to fines of up to EUR 15 million or, for undertakings, up to 3 % of total worldwide annual turnover for the preceding financial year, whichever is higher. For SMEs and, since the amending regulation, for small mid-cap enterprises, the lower of the two amounts applies.
This article is general professional guidance and does not replace legal advice. The applicable statute and the conditions imposed by your competent authority prevail.