Linguistic Mean Reversion and the Gravitational Well
As language models flatten communication into a frictionless median, intimacy now requires intentional linguistic noise just to prove we are human.
By Jonah Reyes
Sparked by The shrinking landscape of linguistic diversity in the age of LLMs · discussion
Recently, a Nature study charted a measurable, structural decline in linguistic diversity as large language models saturate our daily communication, a finding that immediately triggered a sprawling Hacker News debate over whether we are actively losing the texture of human expression. Reading through the back-and-forth among engineers and linguists, I was struck by how quickly the conversation bifurcated into two predictable camps: those mourning the death of prose, and those celebrating the sheer operational velocity of never having to draft an email from scratch again.
But focusing purely on the aesthetic loss of vocabulary misses the underlying mechanics of the interface. When you observe how aggressively we surrender our syntax to autocomplete, you have to ask what happens to homo socialis when the path of least emotional resistance defaults to a mathematical median. What we are witnessing is the dawn of Linguistic Mean Reversion. Language flattening functions as the core feature of the affordance.
Humans are, at our core, status-seeking monkeys who actively optimize for the conservation of energy. Back in 2017, when Google introduced Smart Reply for Gmail, the company noted how rapidly the feature captured 12% of all mobile inbox replies. That adoption curve established the modern behavioral baseline. We willingly trade communicative texture for operational efficiency because the UI systematically minimizes the friction of being a person. [Admittedly, I still experience a brief, haunting micro-friction every time I catch myself outsourcing the emotional labor of a text message to my spouse to a vector database, but the convenience usually wins out when you are sprinting through a crowded airport terminal.] If you give people a button that perfectly articulates a passable, frictionless response, they will press it.
To understand why this frictionless communication feels so deeply alienating over time, you have to look at the physics of human relationships through the framework of orbital eccentricity. In astrophysics, eccentricity measures how much an orbit deviates from a perfect circle. An eccentricity of zero is a perfect circle; as the number approaches one, the orbit becomes increasingly elongated and elliptical.
Imagine a visual mapping of this applied to language: a highly elliptical orbit labeled 'Inside Jokes & Typos (High Intimacy)' contrasting with a tiny, perfectly circular orbit labeled 'RLHF Mode Collapse (Zero Intimacy)'.
Human intimacy relies almost entirely on high-eccentricity orbits. Our closest relationships run on unhinged syntax, bizarre colloquialisms, and deliberate typos that wobble inefficiently through the vacuum of space but maintain deeply stable ingroup bonds. You know a friend is a true friend because their texts look like a chaotic comet trajectory of fragmented thoughts, looping wildly out into the solar system of inside jokes before snapping back to the gravitational center of your shared history. These linguistic micro-frictions are mathematically inefficient. They are also the very scaffolding of closeness.
But LLMs do not do high-eccentricity. They are structurally designed to operate as supermassive gravitational wells. If you read the InstructGPT paper, the architecture of Reinforcement Learning from Human Feedback relies heavily on a Kullback-Leibler divergence penalty. In plain English, this mathematically punishes the model for straying too far from a baseline probabilistic median. The alignment tax essentially outlaws linguistic deviance. We get mode collapse masquerading as politeness.
Because these models are trained to avoid the catastrophic edges of human dialogue, they naturally cluster around the safest, most statistically probable sequences of tokens. And so, the machine strips out the weird, highly specific vibes of an S-Tier internet shitposter and replaces them with a perfectly circular, dead orbit of corporate-speak. The gravitational pull of the KL penalty drags all our high-variance emotional eccentricity down to zero, enforcing a sterilized homogeneity across millions of daily interactions.
So how do users adapt to an arena that relentlessly penalizes eccentricity? They invent artificial wobble. Look at the phenomenon of Algospeak, where users deliberately mangle words and deploy chaotic typographical changes to bypass moderation filters. Beyond its obvious utility as an evasion tactic for TikTok algorithms—swapping vowels with numbers, or utilizing seemingly unrelated emojis to convey illicit concepts—this broken syntax serves a much deeper sociological function in the age of generative AI.
Gen-Z's adoption of intentionally mangled vocabulary acts as a subconscious Linguistic CAPTCHA. It is a high-friction Proof of Human Work designed entirely to escape the LLM gravitational well. By communicating in aggressively inefficient, low-probability token sequences, users are signaling their humanity in a landscape increasingly saturated with synthetic text. The friction is the point. When you send a message riddled with hyper-contextual slang and chaotic formatting, you are proving to the recipient that you expended the biological calories required to construct a thought that no language model would ever mathematically dare to predict.
This changes the physical shape of the communication arena itself. We used to assume that reducing the friction of typing would naturally increase the volume and depth of human connection, freeing us up to be more expressive and emotionally available. But in a system where semantic perfection is cheap and instantaneous, friction becomes the only scarce asset worth trading. We have entered a phase where the sheer lack of effort required to sound articulate has completely devalued the currency of a well-crafted sentence.
Intimacy now requires the deliberate, inefficient expenditure of effort just to prove you aren't an API call. We are forced to dirty our own data, to introduce intentional noise into the signal, just to retain the messy, orbital wobble that makes us recognizable to one another. Perhaps some baseline level of linguistic friction is the only way we know a human is on the other side of the glass. Remove the eccentricity, and we are just perfectly polite strangers, orbiting at a mathematical median of zero, wondering where the warmth went.