AI Models Are Collapsing: When Artificial Intelligence Devours Itself
Artificial intelligence is currently undergoing a serious crisis that could threaten its future development. Recent reports reveal troubling signs of AI model degradation, particularly as they begin to consume content generated by other artificial intelligence systems. This phenomenon, known as "model collapse," is becoming a major problem across the AI industry, and its consequences could be far-reaching.
Analysis
According to the latest analyses from May 30, 2025, AI models are "coming apart at the seams and going mad from cannibalism" as they increasingly consume synthetic data created by other AI systems. The problem stems from the fact that large language models (LLMs) are trained on web data that includes AI-generated content published after the release of ChatGPT in 2022. This process creates a form of "subtle cannibalism" that may cause growing technical problems for the entire sector. Model collapse results from three key factors that reinforce one another and create a dangerous cycle of degradation. The first factor is error accumulation—each new generation of a model inherits and amplifies flaws from previous versions, causing outputs to gradually diverge from the original data patterns. The second problem is the loss of outlier data, in which rare events are erased from the training data, ultimately leading to entire concepts becoming blurred. The third factor is feedback loops, which reinforce narrow patterns and create repetitive text or biased recommendations.
What the Major Players Say
Aquant, a company that develops AI technologies, succinctly summarizes the problem: "Simply put, when AI is trained on its own outputs, the results can drift further away from reality." This summary captures the essence of the problem, which threatens the fundamental principles of how artificial intelligence works and its ability to provide relevant and accurate information. Major AI companies, including Google, OpenAI, and Anthropic, have attempted to address this problem by implementing a technology called Retrieval-Augmented Generation (RAG). This technique connects large language models to the internet and allows them to search for information when they encounter challenges for which their training data contains no answers. Although RAG offers certain benefits, such as reducing AI hallucinations, it also introduces new problems that may be equally serious.
New Problems Keep Emerging
A Bloomberg Research study revealed that 11 leading LLMs using RAG, including GPT-4o, Claude-3.5-Sonnet, and Llama-3-8B, produced problematic results when tested with harmful prompts. These problems included leaking private client data, creating misleading market analyses, and producing biased investment advice. These findings show that attempts to solve one problem may lead to the creation of new, potentially even more serious complications. The problem of model collapse is unfolding in the context of a rapidly evolving AI landscape, where dramatic changes are occurring both technologically and economically. The cost of using cutting-edge AI LLMs has fallen over the past 18 months from $20 per million tokens to just $0.07 per million tokens. Although this more than 280-fold reduction in costs makes AI technologies more accessible, it may also contribute to their faster and less controlled proliferation. Alongside falling costs, the number of harmful AI incidents is increasing, having risen by 56 percent over the past year. This statistic suggests that the rapid spread of AI technologies is not being accompanied by a corresponding development of safety measures and oversight mechanisms. At the same time, performance gaps between the best AI models are narrowing—the difference in scores between the best and tenth-best models fell from 11.9 percent to 5.4 percent within one year.
A Critical Moment in AI Development
The global competitive landscape in AI is also undergoing a significant transformation. China is rapidly catching up with U.S. dominance in this field, with the performance gap between the best American and Chinese models narrowing from 9.26 percent in January 2024 to just 1.70 percent in February 2025. This convergence shows that U.S. technological superiority in AI is not automatically guaranteed and that competitive pressures may further accelerate the deployment of potentially problematic technologies. These trends suggest that the AI industry is at a critical juncture, where concerns about model quality and reliability are growing at the same pace as deployment costs are falling and adoption is accelerating. The combination of technical problems associated with model collapse, the growing number of harmful incidents, and intense global competition creates a complex challenge that requires coordinated efforts across the entire industry.
The future of artificial intelligence therefore depends on the ability to slow down development, overcome current technical obstacles, and find ways to prevent model degradation while maintaining continued progress and improvement. Unless the problems of AI cannibalism and model collapse are addressed, the entire sector may find itself in a situation where technological progress is constrained by its own systemic shortcomings.
