We discover that learning from data produced by other models causes model collapse – a degenerative process whereby, over time, models forget the true underlying data distribution, even in the absence of a shift in the distribution over time. We give examples of model collapse for Gaussian Mixture Models (GMMs), Variational Autoencoders (VAE) and Large Language models (LLMs). We show that over time we start losing information about the true distribution, which first starts with tails disappearing, and over the generations learned behaviours start converging to a point estimate with very small variance. Furthermore, we show that this process is inevitable, even for cases with almost ideal conditions for long-term learning i.e. no function estimation error.

Stable Diffusion revolutionised image creation from descriptive text. GPT-2, GPT-3(.5) and GPT-4 demonstrated astonishing performance across a variety of language tasks. ChatGPT introduced such language models to the general public. It is now clear that large language models (LLMs) are here to stay, and will bring about drastic change in the whole ecosystem of online text and images. In this paper we consider what the future might hold. What will happen to GPT-{n} once LLMs contribute much of the language found online? We find that use of model-generated content in training causes irreversible defects in the resulting models, where tails of the original content distribution disappear. We refer to this effect as model collapse1 and show that it can occur in Variational Autoencoders, Gaussian Mixture Models and LLMs. We build theoretical intuition behind the phenomenon and portray its ubiquity amongst all learned generative models. We demonstrate that it has to be taken seriously if we are to sustain the benefits of training from large-scale data scraped from the web. Indeed, the value of data collected about genuine human interactions with systems will be increasingly valuable in the presence of content generated by LLMs in data crawled from the Internet.

Und siehe da, es gibt auch den KI-Inzest, ein technisches Pendant. Eine Studie mit dem programmatischen Titel «Der Fluch der ewigen Wiederkehr» zeigt, dass KI-Modellen diese exponentielle Selbstverdauung nicht guttut. Sie kollabieren. Dieser Kollaps wird dadurch hervorgerufen, dass die Sprachmodelle im Verlauf der Zeit die Originaldaten unwiederbringlich vergessen, sodass die Modelle schlechter werden darin, gute und wirklichkeitsgetreue Inhalte zu produzieren, und immer mehr Fehler machen.