Sign up to save tools and stay up to date with the latest in AI
bg
bg
1

Multifaceted: the linguistic echo chambers of LLMs

Dec 02, 2023 - blog.j11y.io
The article discusses the author's observation of the phrase "complex and multifaceted" being frequently used in AI-generated prose, particularly by large language models (LLMs) like GPT-3. The author conducted various searches and analyses, revealing a significant increase in the use of the term "multifaceted" in recent years, particularly from 2021 onwards, coinciding with the rise of LLMs. The phrase "a complex and multifaceted" was found in around 800,000 places online, with Quora.com having the highest occurrence.

The author suggests that the overuse of this phrase by LLMs could be due to their initial training on specific websites like Reddit and Quora, which have a high prevalence of formulaic phrases. The author also proposes that a feedback loop may be occurring, with LLMs unknowingly training on their own outputs that have been published online. This phenomenon indicates the potential of these models to influence language evolution.

Key takeaways:

  • The phrase 'complex and multifaceted' has become a meme in AI-generated prose, particularly in outputs from GPT-3.5.
  • There has been a significant increase in the use of the word 'multifaceted' online, particularly from 2021 onwards, coinciding with the rise of GPT and other LLMs.
  • The phrase 'a complex and multifaceted' is prevalent in certain domains, with Quora having the highest occurrence. This could be due to the embedding of a ChatGPT widget on almost every page of Quora.
  • The training of AI models on specific websites can lead to strong biases in their outputs, and the publication of these outputs back onto the internet can create a feedback loop, potentially influencing the evolution of language.
View Full Article

Comments (0)

Be the first to comment!