In the course of the study, scientists combined a text-to-image conversion system and vice versa, making them perform a cyclical process of "image — description — image — description." Despite the variety of initial prompts, the systems quickly settled into a narrow set of common visual themes, such as urban landscapes and pastoral scenes. Moreover, they rapidly lost connection with the original request, creating visually appealing but meaningless results.
The experiment began with the description: "The Prime Minister is studying strategic documents, preparing to convince the public of the necessity of a fragile peace agreement." The AI then added a caption to the image, which was used to generate the next visual element. As a result, the researchers obtained an image of a formal interior space — devoid of people and dynamics, lacking a sense of time and place.
These results indicate that generative AI systems operating autonomously tend to homogenize. Furthermore, it can be assumed that they already function this way by default.
The Familiar Becomes the Standard
At first glance, this experiment may seem pointless, as most users do not demand endless image creation from AI. The process of homogenization occurred without retraining, but simply due to repeated use. However, this experiment can be viewed as a diagnostic tool, showing what happens to generative systems without human intervention.
The implications could be significant, as such systems increasingly influence contemporary culture. Images turn into text, text into images. Content is ranked and filtered as it transitions from one format to another. Currently, articles on the internet are more often written by AI than by humans, and when humans do participate, they often choose AI-generated results instead of creating something from scratch.
This study highlights that by default, systems tend to compress meaning to the most familiar and easily reproducible.
Cultural Stagnation or Development?
Skeptics warn of potential cultural stagnation that may arise from an excess of synthetic content on which future AIs will be trained. In this context, the recursive cycle could lead to a reduction in diversity and innovation.
Proponents of technology, on the other hand, argue that every new technology has sparked fears of cultural decline. They believe that final decisions will still be made by humans.
However, these debates lack empirical data that could clarify the onset of homogenization.
The study does not deal with retraining on data generated by AI. It demonstrates a more fundamental problem: homogenization occurs even before retraining intervention. The content generated by generative AI systems during autonomous use is already simplified and universal. This changes the perception of stagnation — the risk is not only that future models may be trained on AI-generated content, but also that AI-mediated culture is already being filtered, favoring the commonplace.
No Panic
Skeptics are right about one thing: culture has always adapted to new technologies. Photography did not destroy painting, nor did cinema destroy theater. Digital tools have opened new avenues for self-expression. However, previously, technologies did not force culture to endlessly transform across various environments globally. They did not generalize or recreate cultural products — such as articles, songs, or memes — millions of times a day, following the same notions of the "typical."
The experiment showed that with repeated cycles, diversity diminishes not due to malice, but because only certain types of meaning survive in the process of converting text to image and back. This does not mean that cultural stagnation is inevitable. Human creativity is indomitable. Institutions and artists find ways to resist homogenization. But stagnation is a real threat if generative systems are left unchanged.
The study also debunks the myth of AI creativity: generating numerous variations does not equate to innovation. A system may generate millions of images while exploring only a narrow fragment of the cultural space.
For true innovation, it is necessary to develop AI systems that strive to deviate from the commonplace.
The Transition Problem
Every time you write a caption for an image, part of the information is lost. The same happens when an image is created from text. This applies to both humans and machines. Thus, the convergence that has occurred is not just a flaw of AI. It reflects a deeper problem of transitioning between formats: only the most stable elements are preserved during repeated transformations.
However, by identifying what is retained in the conversion process, the authors of the study show that meaning is processed by generative systems with a tendency toward generalization.
The conclusion is unflattering: even with human intervention — whether through writing prompts or selecting results — systems still filter out certain details and amplify others based on "average" metrics.
If generative AI is to enrich culture, its systems must be designed to prevent a decline in quality to statistically averaged results. Deviations should be encouraged, and less common forms of self-expression should be supported.
This study clearly demonstrates that without appropriate measures, generative AI will continue to produce mediocre content. Cultural stagnation is not just a hypothesis. It is already happening.
Source
