More and more restaurants are starting to use generative AI to create menu illustrations, but such images are causing dissatisfaction among consumers. TechCrunch reports that these images are often too smooth and symmetrical, with details that seem complete but lack the texture that real food should have, making it easy for people to feel something is "not right" at a glance.

In some cases, the distortion is quite obvious, such as abnormal cheese textures or shrimp shapes that are not reasonable; more often than not, the issue is not so extreme, but rather reflected in an overall sense of uniformity. Those interviewed believe that this discomfort is not an illusion, but rather the result of the generated models producing similar outputs.

Training methods promote aesthetic convergence

According to a report citing the statements of Reality Defender, the Chief Technology Officer, and Alex Lisle, the image model identifies patterns based on massive training data and then generates results that match the prompts. For example, if a user requests to create a hamburger restaurant menu, the model will invoke the visual features it deems "most similar to a menu image."

However, model training is not just about pursuing accuracy; it also tends to favor content that is “more pleasing” or “safer.” As a result, ice cream balls become overly round, and pizzas, wraps, or seafood all end up with an unnaturally standardized appearance. Over time, the outputs begin to increasingly resemble a single aesthetic template.

It's not a crash, but a degradation in quality.

Lisle indicates that if a model continuously absorbs content generated by itself (AI), in extreme cases, it may experience a 'model crash', which means its output capability becomes completely ineffective. However, phenomena like restaurant menus are more akin to 'convergence'; that is, the model does not lose its functionality, but the quality of the content it produces continues to decline.

This kind of degradation usually doesn't immediately make the images seem absurd; instead, they appear ordinary, neat, and harmless. However, the more one looks at them, the less real they seem. It is precisely for this reason that many people may not be able to spot the problem at first glance, but they still instinctively feel uncomfortable.

It's becoming increasingly difficult to judge authenticity.

The director of Digital Future Center, Imagining, the, and AI believes that the common tendency in AI regarding images and language is to "smooth out" the edge features. This optimization reduces offensiveness and a sense of dissonance, but it also leads to a noticeable homogenization.

Reports indicate that the public is developing an increasingly strong intuitive ability to recognize the content generated by AI. Even if they cannot accurately explain why, people can still perceive the difference between it and real photos or artificially designed images. As this type of content further enters commercial scenarios, it will become even more difficult to determine whether the images are credible or genuine.