Generated images have long since arrived on websites, shops and menus. Two current reports show that two questions are growing together as a result: how do such images affect the audience – and how well can they still be recognised as AI products at all?

Too perfect seems less appetising

A study has investigated why AI-generated food images make people less hungry than expected. Some restaurants rely on artificial intelligence for their menus to illustrate dishes. However, the result is ambivalent: some images seem rather off-putting than appealing.

According to the study, the reason lies in perfection itself. A flawlessly arranged, evenly lit presentation without any irregularity does not correspond to what guests expect from a real plate. The discrepancy between image and expected reality turns into scepticism – an effect that can fundamentally be transferred to other product categories, even if the study argues using the example of food.

For websites, this is a practical hint: visual language is not purely an aesthetic matter. If a product image seems too smooth, it can set expectations so high that disappointment is inevitable – or that trust is not established in the first place.

New models for detection

At the same time, Pangram Labs has presented the Image 1.0 and Image Flash 1.0 models. They are intended to reliably detect AI-generated images and videos. This creates a technical counterpart to the generators: what was created should also be identifiable again.

How robust the detection is in practice cannot be derived from the announcement alone. Detection models fundamentally work with probabilities, not certainties. Anyone who integrates such tools into editorial workflows should therefore treat them as an indication and not as a verdict.

Nevertheless, the development is relevant. As soon as detection services are more widely available, platforms, media or even competitors can check whether image material was generated. The silent use of AI visuals thus becomes riskier.

What both reports have in common

At first glance, an appetite study and a detection model have little to do with each other. Taken together, however, they describe the same shift: AI images are no longer just a question of production costs, but a question of impact and verifiability.

  • Impact side: Generated images can undermine the intended message if they are too perfect or too generic.
  • Verification side: The use of AI material can increasingly be checked from the outside.

Both together make the conscious handling of image material a quality feature. Transparency is not only an ethical position, but also risk prevention: anyone who discloses that a visual was generated cannot later be caught having concealed it.

Practical consequences for website projects

For companies that operate websites and shops, this results in several concrete starting points.

  1. Real material where expectations arise: Product images, menus, references and team photos are places where users expect a correspondence in reality. Here, real material is usually the safer choice.
  2. AI visuals for illustration instead of depiction: Abstract backgrounds, schematic representations or mood images are less susceptible to the described perfection effect than concrete product depictions.
  3. Plan for labelling: An image caption or a note in media management costs little and creates clarity. In TYPO3, this can be mapped via additional fields in file management, so that the information is attached to the asset and is not lost in the documentation.
  4. Document provenance: Anyone who records which image comes from which source can provide evidence instead of reconstructing it when questions arise.

Visual language as part of editorial work

The reflex to quickly fill image gaps with generated motifs is understandable – especially with large product catalogues or frequently changing content. The study suggests that this path does not automatically lead to the desired result. An imperfect but real photo can in doubt create more trust than a flawless generation.

A simple rule in the editorial process is therefore sensible: before every AI image, the question is whether the motif is intended to depict something or to illustrate something. Depictions need a reference to reality, illustrations do not. This distinction can be recorded in editorial guidelines and checked in approval processes – without the need for an elaborate set of rules.

Classification for web and software projects

For ongoing web and software projects, part of the attention thus shifts from image production to image governance: which types of images may be generated, how are they labelled, and where is the provenance stored? These questions belong in the content management system and in the editorial workflow, not in a separate Excel list.

Anyone who sets this up properly now will not have to maintain it later or explain why a notice is missing.

Sources