🤖 LLMs are starting to change the language we speak - and it's already happening.
The most well-known example is the English word "delve." After the launch of ChatGPT, this word began to appear in scientific articles on PubMed 10-100 times more often than before. Journalist Alex Hern suggested that the reason lies in RLHF: OpenAI and other companies delegate annotation to workers in Nigeria, where "delve" is a common word in formal English. They rated responses with it higher, the model began to prioritize it, and "delve" seeped back into human speech. An academic paper from the end of 2024 couldn't specifically confirm the "Nigerian" version, but agreed that RLHF as a mechanism aligns with the data.
In Russian, a similar path is being followed by "оптика" - "through the user's optics," "change optics," "interesting optics." The word was rarely used in ordinary speech, but now it's noticeably more frequent.
The most noticeable is the construction "it's not A, but B." LLMs love it and insert it into every other response. It's already visible in posts, reports, and presentations by real people who haven't annotated anything.
People complain that AI writes unnaturally, "not human-like." It's amusing that soon this complaint will disappear - not because AI will start writing differently, but because humans will gradually start writing like AI.