ChatGPT Models Show Progressive Self‑Convergence in New Experimental Study

The arXiv pre‑print 2603.12683v1 provides the first longitudinal analysis of recursive training on synthetic data for ChatGPT‑style large language models. The authors constructed a series of experiments in which successive public releases of ChatGPT were prompted to generate text at a temperature of 1.0—maximizing stochasticity—and then measured pairwise similarity using a standard cosine‑based text‑embedding metric. Over the span of three model generations, the average similarity score rose from 0.31 to 0.48, indicating that newer versions produce increasingly overlapping outputs even when the sampling temperature is set to its most diverse setting. The paper attributes this trend to “model self‑convergence,” a term the authors coin to describe the gradual homogenisation of outputs across versions as the training data pool becomes polluted with LLM‑generated text scraped from the web.

The study’s methodology isolates the effect of synthetic data infiltration by comparing two training regimes: (1) a baseline where each model is trained on a static corpus of human‑written text, and (2) a “recursive” regime where the corpus is augmented each iteration with the model’s own generated outputs. In the recursive condition, the decline in output diversity is statistically significant (p < 0.01) relative to the baseline, while the baseline shows no appreciable drift in similarity scores. This controlled design supports the authors’ claim that the observed convergence is not an artifact of model architecture changes alone but is driven by the feedback loop created when models ingest their own creations.

Beyond the raw similarity metrics, the authors examined semantic richness by measuring the variance of topic distributions across generated samples. Using Latent Dirichlet Allocation (LDA) on 10 000 generated paragraphs per model, they found a 22 % reduction in topic entropy from the earliest to the latest release. The narrowing of topical space aligns with the similarity increase and suggests that self‑convergence may erode the breadth of knowledge representation in future iterations if left unchecked. The paper warns that such erosion could manifest as “model collapse” in extreme cases—where output devolves into repetitive or meaningless text—but notes that the current trajectory remains within a functional envelope, merely indicating progressive alignment rather than outright failure.

The authors contextualise their findings against prior theoretical work that predicts catastrophic forgetting when LLMs are repeatedly fine‑tuned on narrow data streams. By contrast, their empirical evidence shows a more gradual, measurable shift toward homogeneity, which they argue is a “progressive self‑convergence” rather than a sudden collapse. They recommend mitigation strategies such as curating training pipelines to filter out synthetic content, injecting diverse human‑authored data at regular intervals, and employing regularisation techniques that preserve entropy in the model’s latent space. While the paper stops short of prescribing a concrete industry‑wide standard, it underscores the need for proactive data hygiene as LLMs become increasingly prolific content generators.

Finally, the authors acknowledge the broader implications of their work for the AI ecosystem. As more organisations deploy ChatGPT‑style models in consumer‑facing applications, the volume of AI‑generated text on the public internet is set to rise sharply, potentially accelerating the self‑convergence feedback loop described in the study. The arXiv authors conclude that “continuous monitoring of output diversity metrics is essential to prevent inadvertent degradation of model capabilities,” a recommendation that aligns with emerging best practices in responsible AI development. Their findings thus serve as both a diagnostic benchmark and a call to action for developers, researchers, and policy‑makers tasked with maintaining the health of next‑generation language models.