The ouroboros, an ancient symbol of a snake consuming its own tail, takes on a new and significant meaning in the age of AI. As AI language models like ChatGPT generate editorial content that increasingly populates the internet, it brings along a plethora of errors, causing concern among human editors working on these websites. This predicament is problematic because these language models are trained on the very material found online. Essentially, AI is devouring itself. This process, commonly referred to as "model collapse" by AI researchers, resembles a disheartening game of telephone. If AI trains on error-ridden synthetic data repeatedly, the desired output becomes nonsensical gibberish. One recent study published on the pre-print arXiv server, using the language model OPT-125m, generated text about English architecture. Upon repeated training, the 10th model's response was incomprehensible, showcasing an unusual preoccupation with jackrabbits. Similarly, another study on the pre-print arXiv server focused on AI image generators trained on AI art. After the AI's third attempt to generate a bird or flower using exclusively AI data, the results became blurry and unidentifiable.

While these instances may have relatively low stakes, this recursive feedback loop could amplify racial and gender biases, which could be devastating for marginalized communities. For instance, ChatGPT has been caught racially profiling Muslim individuals as "terrorists. " To effectively train new AI models, companies require uncorrupted data free from synthetic information. Alex Dimakis, co-director of the National AI Institute for Foundations of Machine Learning, stated that "filtering is a whole research area right now, " emphasizing its immense impact on model quality. Dimakis even suggests that a small collection of high-quality data can outperform a larger synthetic one. Although human data is not flawless—biases can be found in various sources—AI can aid in debiasing these data sets to create improved ones. Currently, engineers must diligently sift through data to ensure that AI isn't being trained on self-created synthetic data. Despite all the concerns surrounding AI's potential for replacing humans, it appears that these revolutionary language models still require the human touch. Residing in Portland, Darren, a writer/editor fascinated by science fiction and the intricacies of our world, resides with his cat. His prior works can be found at Gizmodo and Paste with some effort.