Generative AI Begins to Degrade Due to Lack of New Data
Generative artificial intelligence relies on vast amounts of human content collected from open internet sources. However, scientists are increasingly asking an alarming question: what will happen when the original material begins to run out and models start training primarily on data created by other AIs?
Research already demonstrates a troubling pattern: neural networks are gradually "consuming" their own content, leading to a steady degradation in quality. With each new iteration, the generated results become more clichéd and distorted, losing diversity and expressiveness.
In a recent paper published in the journal Patterns, an international team of researchers simulated this process. They combined an image generator with an automatic image captioning system and ran them in a closed loop—without involving new human-created data. After several iterations, the system began to produce extremely bland, impersonal images devoid of individuality. The authors aptly dubbed them "visual elevator music"—predictable, inexpressive, and lacking creative spark.
This experiment revealed a fundamental phenomenon: even without additional training, autonomous feedback loops in AI naturally drift toward common attractors—averaged patterns that seem "safest" to the system from a statistical standpoint. Instead of evolving, the model gradually collapses into a zone of maximum predictability and minimum risk, losing its ability to generate original solutions.
Professor of Computer Science at Rutgers University, Ahmed Elgammal, views these results as alarming evidence of a growing "cultural stagnation" fueled by generative AI. What is particularly concerning is that the collapse in diversity occurred without any retraining: the model simply repeatedly reprocessed its own output data, inevitably sliding toward clichés.
The situation is complicated by the fact that algorithms are already actively promoting AI-generated content to the top of search and recommendation systems, displacing human creations. While proponents of the technology insist that humans will remain the "final arbiters of creativity," reality is proving more complex. The mediation of culture through AI increasingly filters the flow of content in favor of the familiar and predictable. The very mechanism of statistical models pushes them toward averaging: the system naturally chooses the most probable, and therefore the most clichéd, solutions. As a result, a self-reinforcing closed loop emerges, where each new generation of content becomes even more impersonal.
To prevent cultural stagnation, researchers emphasize the need for fundamental changes in the approach to AI development. A key condition for preserving creative diversity is the active involvement of humans in the generation process. It is precisely the collaboration between humans and machines that can counter the drift toward averaged results. At the same time, it is important to design AI models so that they do not merely reproduce statistical patterns but consciously deviate from them, preserving space for surprise and originality.
Furthermore, careful control of data sources is required: closed loops where AI is trained exclusively on synthetic content must not be allowed. It is necessary to develop comprehensive quality metrics that would consider not only the technical accuracy of generation but also the cultural value, emotional depth, and originality of the created content.
Without such measures, generative AI risks turning from a tool of creativity into a mechanism for its averaging. Instead of enriching the cultural landscape, it will reproduce and reinforce clichés, gradually eroding the diversity of human expression.