People are starting to sound like chatbots, and not just in the paranoid corners of the internet.
The suspicion has been around since ChatGPT went mainstream: spend enough time talking to language models, and you begin to pick up their habits. Researchers at the Max Planck Institute for Human Development think that might already be happening. They tracked YouTube transcripts for 18 months after ChatGPT’s launch and found measurable shifts in how people speak, with noticeable upticks in terms like "underscore," "comprehend," "bolster," and "meticulous."
Correlation is not proof, but the timing is hard to ignore. Something in the language environment has changed.
You can see the same pattern in messier, human corners of the internet. Moderators of large communities that revolve around personal drama and confession threads say their jobs have become much harder. The problem goes well beyond letting AI-written posts slip through. Now, human users have started to adopt the same polished, neutral, over-explaining style that AI tools produce by default.
The public examples are getting more obvious. In the U.K. Parliament, multiple MPs were caught leaning on ChatGPT to draft speeches. A distinctly American phrase, "I rise to speak," appeared 26 times in a single day, a level of repetition that would be bizarre in organic debate. Corporate communications have drifted in the same direction. Starbucks’ closure notices about "memories woven into daily rhythms" sounded less like a shift schedule update and more like a model trying to sound empathetic at scale. You do not need insider access to see the fingerprints.
For people learning AI and machine learning, this is a real problem. Evaluation has always been the hard part of AI work. If human writing, public speeches, and corporate statements are increasingly shaped by model outputs, then the benchmark you are trying to match is already contaminated. It becomes harder to say when a model is doing something impressive versus when it has simply blended into a landscape that was already adjusted to its tone and vocabulary.
That contamination flows in both directions. When AI systems are trained on text that already reflects earlier AI outputs, you get a feedback loop. The next generation of models learns from data that is a little flatter, a little more generic, a little more "AI-shaped" than the last. Over many cycles, that can degrade the subtle variation that makes human language so useful for testing edge cases, spotting failure modes, and understanding where a system genuinely struggles.
Anyone working with language models needs at least a basic grasp of natural language processing and the structure of training data, or you are flying blind around the very thing you are trying to measure. And measuring is something we adore here. That's one of the reasons we were interested in the Max Planck Institute study on the influence of LLMs on human speech.
Online communities watching this trend tend to swing between concern and dark humor. Some worry about the erosion of individual voice, others joke that we have collectively “cooked” our ability to sound like ourselves. Underneath the memes, there is a clear intuition: language is (of course) still shaped by our families, regions, or subcultures. But it is also being shaped by autocomplete, chatbots, and policy-tuned corporate models that all push toward the same safe, noncommittal center.
The interesting question now is what to do about it. One path accepts the emergence of a new default dialect, defined by our proximity to machine learning systems and optimized for clarity, risk reduction, and inoffensive corporate tone.
The other path treats that style as a tool, not a destiny, and pushes back by valuing writing and speech that sound distinctly human, even if that means rough edges and sharper opinions. If you want to understand where the technology is going, digging into how modern large language models are built and evaluated is a better use of time than doomscrolling about their side effects.
It's not likely that we're ever going to stop large language models from influencing how people talk. Now the question will be how to notice the feedback loop and design around it, or whether we're ok letting the defaults quietly decide how everyone sounds.
