This study shows that when generative AI systems operate without human intervention, they begin to exhibit a clear tendency towards uniformity. During the experiment, researchers combined text-to-image and image-to-text systems, creating a cycle of "image — description — image — description." Despite the diversity of initial prompts, the systems quickly began generating a limited set of template visual themes, such as urban landscapes and pastoral scenes. Moreover, they rapidly lost connection with the original prompts, producing visually appealing but shallow results.
The experiment started 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 captions to the images, which were used as prompts for creating the next image. As a result, after several iterations, the researchers ended up with a dull image of a formal interior, devoid of life and dynamism.
These results highlight that generative AI systems have a tendency towards homogenization if allowed to operate autonomously without external oversight. Furthermore, it can be assumed that they currently function this way by default.
Standardization as a Result
This experiment may seem irrelevant, as most users do not engage AI for endless generation of their images. However, the process of devolving into uniform standards occurred without retraining, but rather based on repeated usage. Nevertheless, the setup of the experiment can serve as a diagnostic tool, showing what happens when generative systems operate without human intervention.
These consequences can have serious implications, considering that such pipelines increasingly influence modern culture. Textual and visual formats are constantly transforming and intersecting, while AI-generated content is increasingly replacing human creations.
The findings of this research indicate that these systems by default compress meaning into the most familiar and easily reproducible content.
Where is this Development Leading?
In recent years, skepticism about generative AI leading to cultural stagnation has become more widespread. Critics argue that filling the internet with synthetic content will ultimately reduce diversity and innovation. However, technology advocates counter that every new technology raises concerns about cultural decline and assert that final decisions in creative matters will be made by humans.
These debates require empirical data that could demonstrate where homogenization truly begins.
The new research does not focus on retraining on AI-generated data. It reveals a deeper problem: homogenization occurs even before the retraining process intervenes. The content generated by these systems is already simplified and universal. This calls into question arguments about stagnation. The danger lies not only in the fact that future models may learn from AI content, but also that AI-mediated culture is already filtered in favor of the familiar and conventional.
No Need to Panic
Skeptics are correct that culture always adapts to new technologies. Photography did not destroy painting, nor did cinema destroy theater. Digital tools have opened new avenues for self-expression. However, previous technologies did not compel culture to change infinitely on a global scale, as AI does. They did not generalize and reprocess cultural products millions of times a day based on the same notions of the "typical."
The experiment shows that with repeated cycles, diversity diminishes not due to malicious intent or negligence, but because only certain types of meaning survive in the process of transforming text to image and back. This does not mean that cultural stagnation is inevitable. Human creativity possesses resilience. Institutions and artists have always found ways to resist homogenization. Nevertheless, stagnation poses a real threat if generative systems are left in their current form.
The research also debunks the myth of AI creativity: generating an infinite number of variations does not equate to creating innovations. A generative system can produce millions of images while exploring only a small part of the cultural space.
To create something new, it is essential to develop AI systems that strive to move away from the conventional in culture.
The Transition Problem
Each time you create a caption for an image, certain details are lost. The same happens when transforming text into an image, whether done by a person or a machine. In this context, the convergence that has occurred reflects a deeper characteristic — the transition from one medium to another. When meaning repeatedly passes through different formats, only the most stable elements remain.
However, by demonstrating what is preserved during transformation, the authors emphasize that meaning is processed in generative systems with a tendency towards generalization.
The conclusions are bleak: even with human involvement — whether in writing prompts or selecting results — these systems continue to cut off certain details and amplify others, focusing on "average" metrics.
If generative AI is truly to enrich culture rather than diminish it, systems must be designed to prevent degradation to statistically averaged results. Deviations can be encouraged, and less common forms of self-expression can be supported.
This study clearly shows: without these measures, generative AI will continue to produce mediocre and uninteresting content. Cultural stagnation is not just a threat; it is already happening.
Source
