Lieferando has temporarily suspended the use of AI-generated food photos.The trigger was faulty images on menus that attracted public ridicule.The attempt to illustrate dishes automatically also failed to achieve the desired result at parent company Just Eat.

Taken on its own, this is a footnote from the delivery business.For everyone who uses generative AI in websites, online shops or content management systems, it is a cautionary tale: generated content is produced quickly, but it inherits neither care nor responsibility.Both must be established organisationally.

Why product images in particular are delicate

With texts, the audience forgives an unfortunate formulation.With images that are meant to show a specific product, the standard is different.A photo on a menu is a statement about what can be ordered.If the depiction recognisably deviates from the actual product, it creates not just an aesthetic problem but a breach of expectation.

Generative models produce plausible-looking surfaces, not verified facts.They do not know the specific dish of a specific restaurant.They produce a statistically probable variant of it.This is precisely where the errors that stand out publicly arise: wrong details, inconsistent quantities, depictions that do not exist in that form.

The same pattern applies to other industries.Anyone who has product visualisations, reference images or mood motifs generated should ask themselves the question: does this image claim something about a real product?If so, the risk rises significantly.

Scaling multiplies errors

The second factor is volume.A marketplace with many thousands of partner businesses cannot curate images individually if the process is to run automatically.That is precisely what makes generative methods attractive – and precisely what makes them risky.

In small pilot projects, an error rate of a few percent is hardly noticeable.In full operation across large inventories, the same rate means an absolute number of visible outliers that condense into a brand issue on social networks within hours.

The practical consequence: quality assurance must grow with the output volume.Anyone who generates automatically needs automated review steps plus human spot checks – not either or.

What a robust approval process must deliver

A viable workflow for generated content consists of several stages that can be mapped in most CMS and shop environments:

  1. Generation with a documented prompt, model version and timestamp.Without this metadata, it cannot be traced later how a piece of content came about.
  2. Automatic pre-check against technical and subject-specific minimum criteria, such as format, resolution, brand conformity or prohibited motifs.
  3. Editorial approval by a named person.Generated content does not go live unchecked but passes through the same workflow as editorial content.
  4. Publication with labelling, provided the content is AI-generated and this may be relevant to users.
  5. Feedback channel and emergency path: a way to report complaints and a switch that stops automated delivery at short notice.

The last point is regularly underestimated.When a function has to be switched off, what counts is how long that takes and how many systems are affected.A kill switch – that is, a central switch for immediate deactivation – should be planned in from the start, not improvised in a crisis.

Responsibility remains with the operator

Anyone who publishes content on their website is liable for it – regardless of whether a human or a model produced it.This applies to product descriptions just as much as to images.For marketplaces and portals, there is the additional factor that partner businesses are affected who did not create the content themselves but are associated with it.

This leads to a clear rule for platform projects: automatically generated content that concerns third parties should not be published without their knowledge and ideally not without their consent.An opt-in is slower, but it prevents a technical error from becoming a conflict with partners.

Labelling is not an admission

The concern that labelling generated content looks like a quality defect persists.In practice, the opposite is more likely: unlabelled content that is later exposed as AI-generated costs significantly more trust than a factual marking from the outset.

Technically, labelling can be cleanly mapped in CMS systems such as TYPO3 – as a field on the media object or on the content element that can be output in the frontend and evaluated in the backend.This means it remains possible at any time to evaluate what proportion of a site is generated.

Sensible areas of use remain

Withdrawing in one place does not mean that generative methods are unsuitable for the web.

They play to their strengths where no factual claim is made about a real product:

  • Drafts and variants in the design process that are subsequently replaced by real assets
  • Abstract backgrounds, icons and illustrations without product reference
  • Text suggestions for meta descriptions or teasers that are editorially reviewed
  • Structuring tasks such as tagging, translation drafts or summaries

The dividing line does not run between text and image, but between verifiable claims and decorative embellishment.

Classification for web and software projects

For ongoing projects this means: generative functions are a product decision with process consequences, not a purely technical question.Anyone introducing them should specify approval, labelling, logging and deactivability in the same step as the generation itself.The effort required for this is manageable, compared with the effort of subsequently removing faulty content from a large inventory and managing the communication about it.

Sources