Have you noticed certain words popping up more often in online articles and academic papers? Words like "delve," "intricate," and "underscore" seem to be everywhere these days. New research suggests there's a surprising reason for this linguistic shift: the rise of large language models (LLMs) like ChatGPT in writing. Researchers from Florida State University have developed a method to identify these overused words, which they call "focal words." Their analysis of scientific abstracts revealed 21 such words whose usage has spiked dramatically since the arrival of ChatGPT. But why does ChatGPT favor these specific words? The researchers explored several possibilities, including the model's training data, architecture, and the algorithms used in its development. Surprisingly, they found no evidence that these factors are primarily responsible. Instead, their research points to a more intriguing explanation: reinforcement learning from human feedback (RLHF). In RLHF, humans evaluate the quality of the model's output, and this feedback shapes the model's future responses. The researchers hypothesize that if responses containing certain words are consistently rated higher by human evaluators, the model learns to use those words more often. This theory is supported by tests on Meta's Llama LLM, which showed a correlation between RLHF and the overrepresentation of focal words. An online experiment further explored this connection. Participants were asked to choose between abstracts with and without focal words. Interestingly, participants showed a dislike for abstracts where "delve" appeared at the beginning, possibly indicating growing awareness of LLM-generated text. This research raises important questions about how LLMs are shaping language and the influence of human feedback on AI. The lack of transparency in LLM development makes further investigation challenging, but the researchers argue that understanding these linguistic quirks is crucial as AI-generated text becomes increasingly prevalent. One key concern is the potential decoupling of form and content. LLMs excel at creating fluent text, but if their word choices are driven by superficial preferences rather than meaning, this could erode our trust in eloquent language as a sign of quality thinking. As AI continues to evolve, it remains to be seen how these lexical patterns will develop. Will increased awareness of focal words lead to changes in how LLMs are trained? Or will today's AI-driven language become the norm for tomorrow, potentially even influencing the way we speak and write?