Since 2 August, new transparency obligations have applied to AI-generated content. For online shops, this is more than a formality: product images are the central selling point, and anyone who has them generated in whole or in part must in future disclose that a machine was involved.
At the same time, an experimental study suggests that precisely this disclosure has an effect – and not a desirable one. In the experiment, the AI label worsened the perception of the product. The notable secondary finding: the effect also occurred when the labelled photo had in reality been taken by a human. So it was not the image quality that was decisive, but the label.
Why the effect matters to shop operators
For operators of online shops, this creates an uncomfortable tension. On the one hand there is a legal obligation to be transparent, on the other a measurable influence on perception – and thus potentially on conversion, i.e. the proportion of visitors who actually buy.
The obvious but wrong reaction would be to avoid or hide the labelling as far as possible. That is legally risky and additionally undermines trust as soon as it is noticed. The more sensible question is: how can the disclosure be embedded in the design and information architecture in such a way that it fulfils its function without appearing like a warning?
Labelling as a design task
A disclosure about the use of AI is initially a piece of text at a particular place in the interface. How it is perceived depends on placement, wording and context. A few considerations that can be organised in shop projects:
- Wording: A neutral, explanatory sentence has a different effect from an isolated abbreviation. Anyone who describes what was generated – for example a background or a colour variant – gives the labelling a substantive core instead of a blanket suspicion.
- Placement: A disclosure directly above the price creates a different proximity from one in the product details area. Both can be permissible, but the effect differs.
- Differentiation: Not every image is equally AI-assisted. There is a difference between a fully generated motif and a retouched photo, and the communication can reflect this.
- Consistency: A uniform presentation across all products is less conspicuous than one that appears only on individual items.
The technical foundation
For labelling to work reliably at all, a clean data basis is needed. In practice, this often fails not at the frontend but due to missing documentation: no one knows any more how a given image was created.
It therefore makes sense to have a field in the content management or product information system that records the origin of every asset – fully generated, partially edited, classically photographed. If this field is maintained, the output in the shop can be automated instead of being updated manually. This reduces errors and turns later adjustments to changed requirements into a configuration question rather than an editorial project.
For TYPO3-based websites and connected shop systems, this means concretely: the metadata belongs in the central file management, not in individual content elements. This way the information is retained even when an image is output in several places.
Test instead of assume
The cited study is an experiment and describes an effect under controlled conditions. How strong it is in your own shop depends on the product range, target group and price level. For products that require explanation, transparency can even be read as a quality signal; in strongly visually driven categories, the effect is likely to be more pronounced.
Anyone who has the opportunity should compare variants of the labelling in live operation – for example different wordings or positions on an otherwise identical product page. This provides more reliable indications than any assumption about how customers react.
The strategic question behind it
The finding that the label alone depresses perception – regardless of whether the image was actually generated – points to a more fundamental issue. Evidently many people currently associate AI images with an expectation of lower authenticity. This attribution can shift as generated content becomes more commonplace.
In practice, this means: the use of generated product images should be a conscious trade-off and not a purely cost-based decision. Where real photos offer a recognisable added value, for example with materials, fits or details, the effort can still pay off – not least because no labelling is then required.
For web and shop projects, part of the work thus shifts forward: the question of the origin of image material belongs in the conception phase, not in the post-processing shortly before launch. Anyone who records metadata in a structured way from the outset meets transparency requirements without extra effort and retains design scope in implementation. And anyone who measures the effect of labelling instead of fearing it makes better decisions about where generated images should be used at all.