Rather than processing text like a human, the chatbot turns your prompt into an embedding — a group of numbers represented in a “vector space,” kind of like coordinates on a map. But just as a map is a flattened representation of its territory, this embedding is a flattened representation of language; some amount of human variation and context is necessarily lost.

Next, the chatbot formulates a reply, making a word-by-word prediction based on how it was trained to answer past inputs. This prediction draws on biased training data (the specific texts it learns from) and biased reinforcement learning (the feedback it receives). Ultimately, what looks like English to you is really a simulacrum of actual human speech.

 

The unnerving thing is that now, with hundreds of millions of people regularly engaging with chatbots, English-speaking humans are starting to talk like the inhuman communicator on the other side.

For example, ChatGPT uses the word “delve” at higher rates than people generally use when writing or speaking English. As Florida State University researchers Tom S. Juzek and Zina B. Ward have found, this likely results from small biases and errors in the human feedback process compounding over time. Employees of AI companies checking over large language model (LLM) outputs are often low-wage workers from countries such as Nigeria and Kenya, where “delve” is used at higher rates than in American or British English.

Slight lexical overrepresentations may thus have been reinforced into the model — sometimes even more frequently than in the workers’ own usage. Because of the job’s time pressures and stressful conditions, the checkers might have missed the “delve” frequency while clicking through sample texts, leading to further amplification. (Some other examples are “intricate,” “commendable” and “meticulous.”)

Now this overuse is bleeding into global culture. In the two years since ChatGPT launched in late 2022, the appearance of “delve” in academic publishing saw a tenfold increase as researchers began turning to AI for help with their papers. As scientists and writers have grown more aware of this phenomenon, they’ve taken steps to “sound less like AI.” I actually used to enjoy using the word “delve”; now I try to avoid it.

But most people likely don’t know about the chatbot biases toward certain words. Users assume that ChatGPT is talking in “normal” English, because that’s the implication of the tool’s user interface. They also assume that regular, everyday texts they encounter are normal English, even when they might have been AI-generated as well. Over time, it gets easier and easier to confuse the representation with reality.

Indeed, a study reported in Scientific American last month found that people have started saying “delve” more in spontaneous, spoken conversations. This isn’t AI’s doing anymore; we’ve started internalizing its biases and repeating them by ourselves. I say “we” because even those in the anti-“delve” faction aren’t exempt. We might avoid using the most well-known ChatGPT giveaways, but so many words are appearing with unnatural frequency that we can’t possibly avoid them all. Are we also supposed to stop using the chatbot-overused “inquiry”? Or “surpass”? There’s too much to keep track of.

Our susceptibility to unwittingly adopting LLM vocabulary will likely only increase as the circle of AI text, human text we don’t know is AI, and actual human speech spins faster and faster. In psycholinguistics, it’s generally understood that the frequency of a word’s representation affects its availability in people’s mental lexicon — the “dictionary” of possible options to draw from in spoken conversation. Essentially, the more we see a word represented, the more intuitive it feels to use it ourselves.

It’s normal for our mental language mappings to evolve, but now we’re in a feedback loop where our “maps” are converging with the chatbots’ “maps.” Both differ from the actual landscape — the unfathomable variation in how humans speak — but the confusion is making it harder to evaluate what’s real human language and what’s artificially generated. As this distinction becomes harder to discern, we’ll keep blurring the lines of reality, cyclically using that blur to construct our new realities. And as AI models continue to be trained on both their own output and the AI-influenced human writing they ingest, the pervasiveness of LLM-speak will only intensify.

Linguistically, there’s nothing wrong with this. The word “delve” isn’t any worse for your brain than its synonyms. But language is a harbinger of greater social shifts. There are many other, far more insidious misrepresentations that are also coded into LLMs. Racial biases, gender biases and political biases — all of these are likely trained into the models much like linguistic biases, but these are harder to definitively measure. We need to remember that these aren’t neutral tools: They hold the power to subtly reshape our thinking.