Learning the Language of AI

One of the least predicted outcomes of the AI age is that AI would develop its own lexicon, and that the consequences would ripple far beyond the labs.

The idea isn’t entirely new. In 2017, Facebook’s AI Research lab published “Deal or No Deal?”, a paper on training agents to negotiate. When researchers let two agents optimize against each other, they drifted away from English into a repetitive shorthand that worked for them and was hard for humans to read. The press reported that Facebook had “shut down” its AI. In reality, the researchers simply added a requirement that the agents keep producing human-like language, because the point of the research was negotiating with people. Given a reward and no reason to stay readable, the agents had done what optimizers do.

Today’s frontier models develop their preferences more quietly, but the underlying dynamic is similar.

How Models Get a Lexicon

According to DeepMind’s Chinchilla scaling laws, compute-optimal training requires roughly 20 tokens per parameter. A one-trillion-parameter model therefore needs around 20 trillion tokens. At roughly 130,000 tokens for an average-length book, that’s on the order of 150 million books, several times the number of books in the Library of Congress. Armies of humans and smaller models annotate and curate this data. Then, during fine-tuning and reinforcement learning from human feedback, organizations mutate it further to instill personality, reasoning patterns, and safety alignment.

Each of those steps rewards some phrasings over others. Across billions of examples, the result is a model with measurable preferences: words, structures, and reasoning patterns it uses more often than the humans it learned from. Those preferences then reach people through writing assistants, email clients, coding tools, and search, usually without anyone noticing.

We’re Starting to Talk Like It

The irony is thick. The more humans interact with AI, the more our language converges with its own, and over the last two years researchers have started measuring it.

Researchers at Tübingen and Northwestern adapted a method from Covid excess-mortality studies to vocabulary, measuring “excess words” across more than 15 million PubMed abstracts from 2010 to 2024. They estimate that at least 13.5% of 2024 abstracts were processed with an LLM, reaching 40% in some sub-fields. By their measure, the vocabulary shift was larger than the one Covid produced. It’s worth being precise about what that number means. It’s a lower bound on LLM-assisted writing, not a measure of people changing their own habits.

That’s where the Max Planck Institute for Human Development comes in. Its researchers transcribed over 360,000 academic YouTube talks and over 770,000 podcast episodes and used causal inference methods to compare word usage before and after ChatGPT’s release. They found an abrupt increase in words ChatGPT favors, like delve, meticulous, swift, comprehend, and boast. Some of it appeared in spontaneous conversation, not scripted talks, which suggests people are absorbing the vocabulary rather than just reading AI-edited text aloud. The effect also varied by domain. It showed up in science, business, and education podcasts, but not in religion or sports, which is consistent with exposure driving adoption. The lead researcher, Hiromu Yakura, started the project after noticing he was saying “delve” more himself.

The effect isn’t limited to vocabulary. In a study presented at CHI 2025, Cornell researchers had 118 Indian and American participants write about culturally specific topics with and without AI autocomplete. With AI, the Indian participants’ writing drifted toward American styles. Writing about Diwali without AI, one participant said they would “worship goddess Laxmi” and “pop crackers and eat sweets.” With AI, the description became “a time filled with happiness and warmth.” It’s a small, controlled study, but the direction is clear. The suggestions shaped not only the wording but the perspective.

The Cultural Cost

This fuels the postmodern critique that capitalism has commodified culture itself, an argument Fredric Jameson made in Postmodernism, or, the Cultural Logic of Late Capitalism decades before language models existed. The concern here is concrete. A small number of companies train the models that a large share of the world now uses to write and edit. Their training decisions, made for commercial and safety reasons, become stylistic defaults, and the studies above show those defaults flowing back into human writing and speech as optimized, proprietary dialects.

The Practical Upside

Yet the upside is practical. By mastering AI’s chosen lexicons in our industries, we can prompt, build agents, and integrate systems far more effectively.

The research supports this as well. In 2022, researchers found that adding “Let’s think step by step” to a prompt substantially improved models’ accuracy on reasoning tasks. In 2023, Google DeepMind had models optimize their own prompts, and one of the best-performing instructions began with “Take a deep breath.” Neither phrase means much to a person. They work because of what the models learned to associate with careful reasoning.

In my own work building RAG pipelines and agents, this is the difference between people who get consistent results from LLMs and people who don’t. As I wrote in “Beyond Vibe Coding,” you need to drive domain context into every prompt, and part of that is speaking the model’s dialect: clear product definitions, explicit constraints, structured outputs, and phrasing the model reliably responds to. It’s the same reason the engineers who get the most out of a database are the ones who understand how its query planner works, not just how to write a query.

In the End

The evidence shows linguistic convergence moving from papers to podcasts to how people describe their own traditions. It has real cultural costs, and it’s also a skill that pays off for anyone working with these systems. Knowing both is the useful position: learn the dialect deliberately, and know when to step outside it.

In the end, linguistic convergence is not just inevitable. It is a new literacy.