Humanity's Language Space Is Converging

Thu, 13 Aug 2026 00:00:00 +0000

In July 2025, Cloudflare published an analysis of data collected from June 19 to 26: the balance between how much content AI platforms crawled from websites and how much referral traffic they returned was breaking down. For Anthropic, based on the web referrals visible to Cloudflare, the platform crawled roughly 70,000 pages for every visit it sent back to a content site.¹ One year later, Cloudflare observed that non-human requests accounted for more than half of all requests on its network for the first time, and that 52% of crawler requests were used for AI training.²

Meanwhile, the rate at which humans actively contribute knowledge is falling. According to the public Stack Exchange API,³ Stack Overflow received 109,294 new questions in the month ChatGPT was released. By July 2026, that number had fallen to 1,437—a decline of 98.7%. The knowledge community that taught language models how to answer programming questions is being replaced by the models it taught.

These two trends lead to the question at the heart of this article: what will happen to humanity’s language space in the age of AI? I begin with the proposition that “language is intelligence,” define the stock and flow of language space, and then offer my view on the question.

I. Language Is Intelligence

In 1950, Alan Turing tried to answer the question “Can machines think?” in Computing Machinery and Intelligence.⁴ He quickly recognized, however, that “thinking” itself could not be defined rigorously. He therefore set the question aside and replaced it with an operational test: place a human interrogator in text-only conversations with both a machine and another human; if the interrogator cannot tell which is which, then the machine has exhibited intelligence.

Alan Turing as portrayed in The Imitation Game

The key move was this: Turing replaced the difficult question “Does it possess intelligence?” with the operationally testable question “Can it express intelligence through language?” This article adopts the same operational definition. It does not ask what intelligence ultimately is at the level of consciousness; it asks only whether intelligence can be expressed, transmitted, and observed.

For computers, this substitution is particularly elegant. Everything in a computer—text, images, audio, and video—can be represented in binary, and binary is itself a language in the broad sense. You cannot, of course, expect me to watch short videos by staring at a string of zeros and ones. Yet with a fixed symbol set and rules of composition, that string can indeed be reconstructed into any of these modalities. Any multimodal information in a computer can therefore be represented as a form of language.

As for memory, consciousness, and the other parts that resist clear description: they have either been expressed in some form, or they have never entered public language. As Wittgenstein wrote in the Tractatus Logico-Philosophicus, “Whereof one cannot speak, thereof one must be silent.” What cannot be expressed does not participate in transmission, accumulation, or training, and therefore lies outside the scope of this article.

Within this framework, then, language is intelligence that can be expressed and transmitted.

II. Language Space

Language Has Boundaries

The Tractatus contains another equally famous line: “The limits of my language mean the limits of my world.”

Our framework considers only intelligence that can be expressed and transmitted. If language has boundaries, then this part of intelligence has boundaries as well. We can therefore substitute the boundary of language for the boundary of intelligence and discuss its extent and size.

To make that discussion possible, I will introduce a model of language space that is not rigorous, but is useful for discussion.

Everything that can be expressed forms a high-dimensional space, which I call the total language space. The content that any individual entity—a human or a language model, for example—can express forms a language subspace within it. Under the definition above, the boundary of that subspace is also the boundary of the intelligence that the entity can manifest.

To describe how much humanity has already understood and expressed, define the subspace size $C(t)$:

$$ C(t)=\left|\Omega(t)\right|,\qquad \Omega(t)=\bigcup_{i=1}^{N(t)}R(i) $$

Here, $\Omega(t)$ denotes the human language subspace that has entered public expression by time $t$; $R(i)$ is the language subspace occupied by the $i$-th human entity; and $N(t)$ is the number of such entities. $C(t)$ is therefore the size of the human-known language subspace. In the two-dimensional projections below, it corresponds to the projected area jointly covered by all these subspaces.

Different entities have different language boundaries. To use a simple example, a senior professor generally knows more than a young child and therefore occupies a larger language subspace. Yet they still share a native language, everyday experience, and much else, so their subspaces overlap substantially. Likewise, a stronger language model will generally occupy a larger subspace than a weaker one.

With these concepts in place, we can project the model into two dimensions:

Language-subspace projections of humans and language models

The circles above represent two-dimensional projections of language subspaces. On the left are hundreds of human subspaces of varying sizes, many of them overlapping; isolated points near the edge correspond to obscure knowledge held by only a few people. On the right are only four language-model subspaces, but because each model has been trained at scale, each individual projection is large.

The contrast is striking: humans form a long tail through a large number of small subspaces, whereas language models cover common knowledge through a small number of large subspaces.

Language models also share large amounts of the same internet training data—sometimes even the same data providers—and are often distilled from one another. Their subspaces therefore overlap heavily. Adding more models does not expand language space in the same proportion.

The other side of the coin is that a language model can attain a very high degree of mastery over parts of its own subspace. Within a familiar region, its answers may be more accurate than those of most humans who occupy the same region.

Add a $z$-axis to the two-dimensional projection, using height to represent an entity’s degree of mastery at each location, and the result is the following three-dimensional terrain:

Projected area and degree of mastery

The dashed lines above mark the boundaries of the subspaces, while height represents mastery. To make the comparison clear, the two projected areas are drawn at the same size. On the left is a human: several domains contain modest peaks of accumulated expertise, separated by flat ground and depressions. On the right is a language model: one extremely high and steep peak rises within its familiar subspace, then drops rapidly beyond it.

Two concepts must therefore remain distinct: the boundary encloses the projected area, while terrain height represents mastery at a particular location. A language model may achieve very high mastery within an existing subspace without increasing the size of that subspace.

The Stock and Flow of Language Space

With subspace size defined, we can return to the original question: is internet content converging or diverging?

The amount of content on the internet can continue to grow while the human-known language subspace stops growing. The former counts the number of points; the latter measures the range jointly occupied by those points.

Because $C(t)$ changes over time, we can distinguish two quantities:

The stock is $C(t)$ itself: the size of the human-known language subspace.
The flow is the rate at which subspace size grows, $F(t)=C^{\prime}(t)$: the amount of language space newly added per unit time that no entity had previously reached.

Whether language space converges should therefore be determined by whether $C(t)$ has a finite limit:

$$ \text{convergence} \iff \exists\, C_{\mathrm{max}} \lt \infty,\quad \lim_{t \to \infty} C(t) = C_{\mathrm{max}} $$

In other words, if $C(t)$ eventually approaches a finite value, the human-known language space converges. In the smooth model used here, its flow $F(t)$ then gradually approaches zero. The converse does not automatically hold: flow approaching zero is not by itself sufficient for convergence. This article makes the additional assumption that the sum of all future additions to language space is finite.

One point is especially important: convergence does not mean collapse to zero. $C(t)$ is cumulative. It can stop growing, but it does not return to the origin.

Total flow has three sources—human entities, language models, and Agents—so I will examine how each affects the human-known language space:

$$ F(t)=F_{\mathrm{human}}(t)+F_{\mathrm{LM}}(t)+F_{\mathrm{Agent}}(t) $$

These terms classify new knowledge by its source. An output from a language model or Agent counts toward $\Omega(t)$ only after it has been confirmed and entered public expression.

III. Language Models and Agents Contribute Differently to Flow

A Model Sees a Projection of the Real World

By language model, I mean a model that does not call tools or receive feedback from an external environment, and instead predicts subsequent tokens solely from the context it has already received. It is trained by fitting human knowledge products. What it encounters is not the real world itself, but a projection of the real world onto human language.

A language model can combine and interpolate within an existing subspace, producing sentences that never appeared in its training set. But being novel at the surface level does not mean that an utterance has left the human-known language space at the semantic level.

Under the central premise of this article, the relation among the three spaces can be written as:

$$ \mathcal{L}_{\mathrm{LM}}(t) \subseteq \mathcal{L}_{\mathrm{human}}(t) \subseteq \mathcal{L}_{\mathrm{all}} $$

These three terms denote the language-model subspace, the human-known language space, and the total language space. A language model can add content within the existing range while leaving the boundary of the human-known language space approximately unchanged.

This requires distinguishing two concepts:

A language model can achieve greater mastery within an existing subspace: understanding the same class of problems more fully and answering them more accurately.
The human-known language space can be expanded: content beyond the old boundary can be understood and expressed for the first time.

Present-day language models mainly do the former.

Adding content versus expanding the boundary

The dark regions above represent new content. In the left panel, the boundary remains fixed and all new content falls within the existing range, so only the quantity of content increases. In the right panel, new content appears beyond the old boundary, expanding the human-known language space. Both cases may be described as “generating new content,” but only the latter increases subspace size.

The Trend in Subspace Size

Once stock and flow have been separated, and language models have been shown to add content primarily within an existing subspace, we can draw the trend projected by this article:

Projected evolution of the human-known language subspace

The vertical axis above is the human-known language subspace size $C(t)$, and the horizontal axis is time. The dot marks the transition from accelerating to decelerating growth; the dashed line shows what would happen if that transition never occurred.

This is a conceptual trend, not a statistical fit to historical data. In prehistory, durable written media were scarce, so expression could not readily accumulate across generations. Ancient societies began to accumulate knowledge, but the rate was limited by literacy and the cost of copying. Printing, the Industrial Revolution, and the information age then increased both the number of people able to express themselves and the speed of transmission, pushing subspace size into a period of accelerating growth.

In the future, once language models enter the picture—and again, I mean language models rather than tool-using Agents—if human-contributed flow continues to decline and Agents fail to replace the missing flow, a fourth stage begins: subspace size shifts from accelerating growth to decelerating growth, and the human-known language space shifts from accelerating expansion toward gradual convergence.

How Humans and Agents Contribute Flow

The human-known language space has expanded because humans interact with the real world.

Reinforcement learning models precisely this process. Humans act in an environment; the physical world and other humans provide feedback; that feedback changes what they know; and the resulting knowledge is externalized in language. A materials scientist, for example, may learn from a series of failed experiments and, in one experiment, finally discover an entirely new material and receive a reward. The boundary of human language—that is, human knowledge—thereby expands.

Human language has evolved through this process. By contrast, if a language model only statistically fits existing human knowledge products, it will struggle to move beyond the boundary already drawn by its training data. This is the generalization boundary at issue here.

The same argument reveals the fundamental difference between a language model and an Agent. A language model can process only what has entered its context; an Agent can call tools, take actions, and receive feedback from an environment, giving it the opportunity to obtain observations absent from existing human knowledge products.

Return to the laboratory. Suppose an Agent controls experimental equipment, varies the conditions automatically, and observes a previously unknown phenomenon. If the phenomenon can be reproduced and validated, then expressed as a new concept by either the Agent or a human, it has the same structure as a scientific discovery: the real world supplies new feedback, the feedback becomes new knowledge, and the knowledge expands the human-known language space. The fact that language models struggle to contribute flow does not mean that Agents cannot contribute it.

Real-world interaction creates new flow into language space

The image above renders this feedback loop as a laboratory: existing literature supplies the starting point, the environment provides new feedback, and a validated result is ultimately expressed as new knowledge.

Even when an Agent acts only inside a digital sandbox, the sandbox is not an absolute constraint. The rules of Go were designed by humans, yet AlphaGo produced moves that no human had played before. Conway designed the rules of the Game of Life, yet he did not foresee either the glider or the system’s Turing completeness. A sandbox constrains the axioms, not every conclusion that follows from them.

Whether an Agent Can Expand the Boundary Depends on Environment and Reward

Whether an Agent genuinely contributes flow depends on whether its environment supplies new information, and whether its reward can distinguish a discovery from an error. An Agent that merely searches existing webpages still primarily recombines human knowledge products. An Agent that obtains new observations, and receives rewards grounded in experimental replication, program verification, or physical outcomes, may expand the human-known language space.

For language models and Agents rewarded only by human preference, the original limitation remains. What counts as a better answer must ultimately be judged either by humans or by a reward model fitted to human preferences. To an existing evaluation system, an output genuinely outside the human-known language space is often indistinguishable from an error: both are unfamiliar, low-probability, and difficult to reward positively.

In other words:

If reward comes only from human preference, then asking a system to leave the language space already occupied by humans is in inherent tension with asking it to align strictly with human preference.

This leads to a further conclusion. Under the current training paradigm, a language model can attain very high mastery of knowledge within an existing subspace, but it struggles to expand the human-known language space reliably, so its language-space flow is approximately zero. An Agent that can engage the real world, obtain new observations, and submit them to objective validation contributes nonzero flow.

IV. Human-Contributed Flow Is Approaching Zero

Although Agents may become a new source of flow, interaction between humans and the real world remains the primary way language space expands today. Yet human-contributed flow is declining for two reasons: strong language models are changing people’s willingness to create, and they are weakening the need to document. At the same time, the volume of language-model output is reducing the share of original human content.

Willingness: Knowledge Is Becoming Cheap

By a strong language model, I mean one whose output quality already exceeds that of most humans. Before such models appeared, writing a long article with genuine insight required research, problem solving, and independent judgment.

Yet strong language models lower the cost of writing for everyone at once. A well-structured article with correct terminology and coherent logic no longer proves whether its author spent years mastering the field or encountered it only yesterday. The same ambiguity applies to this article—although it was in fact written by the author himself.

In my own field, many programmers once wrote blogs partly to demonstrate technical depth and exchange ideas with peers, and partly to summarize their work. When an article can no longer prove that its content came from the author, the incentive to use writing as evidence of one’s individuality, distinctive views, and problem-solving ability naturally declines.

Article-quality distributions before and after strong language models

The image above is conceptual. The dark curve represents authors who genuinely understand the field; the light curve represents authors without the same background. In the past, the quality distributions were far apart, so an article served as a signal of its author’s ability. With a strong language model, the two distributions overlap heavily. The writing remains fluent, but the signal loses its power to discriminate.

Creators respond in two opposite ways:

Some rush in. When production costs have already fallen but evaluation systems have not yet adjusted, the time lag creates an arbitrage opportunity. Academic papers, for example, have historically brought economic and reputational rewards. If language models sharply lower the writing barrier, creators are incentivized to increase output before the evaluation system catches up.
Others leave. For those who once used papers or blogs to demonstrate accumulated expertise, the return on public writing falls once an article can no longer establish that the ideas came from its nominal author. Luo Fuli, for example, has publicly said that she now rarely reads academic papers.

The second mechanism is that the need to document is weakening.

Many people once published articles because they had solved a problem only after reading a great deal and trying repeatedly, and wanted both to preserve the process and summarize the result. Writing was a byproduct of understanding: the author first had to understand the problem before explaining it. Now, more and more problems can be answered with a single prompt. When solving a problem no longer requires the user to understand it deeply, the need to record and share the process also declines.

Someone who solves a century-old mathematical problem will naturally publish a paper announcing the solution. Someone who completes an elementary arithmetic exercise has no reason to write an article about it. One role of language models and retrieval Agents is to turn many problems that once required dedicated documentation into questions that can be answered directly through autonomous search.

Capacity: Human Output Is Being Drowned Out

Whatever happens to incentives, some people will always choose to create. We must therefore consider a factor independent of psychology: the enormous gap between the content generation rates of humans and language models. The gap is not limited to text. Model-generated images, audio, and video also add content within existing subspaces.

This content-generation capacity is also moving from data centers onto every endpoint.

Some 7B–32B models already outperform early GPT-3.5 Turbo on several public benchmarks. As compression and training efficiency improve, language models will continue moving onto phones, personal computers, and embedded devices. Once those devices continually generate content and feed it back onto the internet, the aggregate content generation rate will far exceed human production.

There is an easily overlooked detail here. Model miniaturization often relies on distillation: a smaller model learns the output distribution of a larger teacher. The smaller model is therefore not a wholly independent linguistic entity; its subspace remains substantially constrained by its teacher.

In the language of our two-dimensional projection, miniaturization does not naturally create an independent new subspace. It copies an existing subspace onto billions of devices. The human-known language subspace changes little in size, while the number of content points grows rapidly.

This produces two further consequences.

First, the last throughput constraint on content generation disappears. Large-scale generation once depended on data centers, which imposed compute costs and left only a few organizations that could be held accountable. Once generation happens on every endpoint, total output is no longer constrained by any single provider.

Second, content provenance becomes harder to trace. Watermarking AI content requires the generator to cooperate. With open weights and local deployment, anyone can remove a marker or use a model that never adds one. Model-side watermarking alone is therefore unlikely to produce an enforceable universal standard.

Original Human Content Is Becoming the Long Tail

Here, the long tail no longer means minority languages, peripheral disciplines, or niche technology stacks. It describes the source distribution of future internet content. Because language models generate content much faster than humans, model-generated content will occupy the head by volume, while original human content gradually recedes into the tail.

Plot the content added during a given period by source and quantity, and the distribution looks roughly like this:

The source distribution of newly added internet content

Each point above represents one new piece of content. The enormous quantity of model-generated content forms a dense head, while the smaller amount of original human content forms a sparse, elongated tail. The long tail here describes only the relative quantities of the two sources.

Encountering original human content on the internet often feels like meeting an old friend in a foreign land.

V. The Inflection Point in Acceleration

Agents can contribute flow, but at least for the foreseeable future, they remain insufficient to offset the decline in human flow and have not formed a stable, sustainable source of new additions. There are two reasons:

Very few Agents interact with real environments rather than merely retrieving webpages, and each run produces only a limited amount of new knowledge that can be validated.
Real-world interaction requires tools, environments, time, and capital, making it far more expensive than an ordinary generation call.

Under this assumption, I still expect the total flow of humanity’s language space to approach zero, the sum of future additions to remain finite, and the stock $C(t)$ to approach a horizontal asymptote. Agents can slow this process for now, but cannot reverse it.

Plotting the derivative of the earlier subspace-size curve makes this change easier to see:

The changing flow of the human-known language space

The flow $F(t)=C^{\prime}(t)$ above is the amount of new language subspace added per unit time—the derivative of the subspace-size curve $C(t)$. Before the Industrial Revolution, flow remained low. It then rose rapidly and reached a peak, before declining in the future. Flow approaching zero does not mean that the existing subspace shrinks; it means only that the human-known language space ceases to expand.

Declining flow does not mean less internet content. Quite the opposite: there will be more content than ever, but most of it will consist of extracting, recombining, and republishing existing human knowledge products. Quantity keeps growing, while the boundary fails to expand at the same rate.

Put differently, most current language models and retrieval Agents consume the knowledge products accumulated by humans without expanding the boundary of language space at a comparable rate. Webpages, papers, and images do not disappear, but existing knowledge is processed repeatedly while knowledge beyond the boundary is not produced at the same pace.

The physical concept of heat death offers an analogy for this endpoint. Energy still exists, but without a temperature gradient it can no longer do work. Language space is similar: the stock remains, and content becomes more abundant, but the boundary no longer expands. Agents offer a theoretical exit, yet their present scale, knowledge output, and operating cost remain insufficient to change the outcome.

The heat death of language space—and a possible exit

In the image above, the shelves and terminals filling the room represent an abundant stock that still remains. The door to the real world represents observations that humans and Agents may still bring back.

VI. Three Implications for the Future

Implication One: Individual Gains and Collective Stagnation

Strong language models are plainly valuable to individuals. A knowledge framework that once required years of study can now be acquired through a few hours of conversation; writing, research, and organizing ideas all become faster.

Yet widespread individual benefit is compatible with a language space that has stopped expanding. Raising the average level does not expand the boundary; subspace size increases only when content lands outside the existing range. Strong language models raise everyone’s floor, but also pull more people toward similar answers. The mean rises; the boundary may not move.

Implication Two: Scarcity Shifts Back to People

When one factor of production becomes cheap, value shifts to complementary factors that remain scarce. Once access to knowledge is cheap, at least four things remain scarce: trust, distribution and attention, taste, and responsibility. A language model can provide an answer, but it cannot decide on a person’s behalf which questions are worth answering, nor can it assume responsibility for the consequences. Human connections therefore become more important.

Implication Three: Efficiency Becomes Stratified Again

Within the same group, people who use Agents proficiently will rapidly converge in productivity. At the same time, the efficiency gap between those who use Agents and those who do not will widen sharply.

Efficiency stratification created by Agent use

The vertical axis above represents work efficiency. In the left panel, five people occupy three tiers, with roughly equal gaps between adjacent tiers. In the right panel, the three Agent users rise together and move closer to one another, while the two non-users remain where they were. The result is a smaller gap within the first group and a larger gap between groups.

VII. Conclusion: From Explanation to Bearing Witness

If the flow of humanity’s language space comes from interaction with the real world, then the most important value of personal writing is no longer to repeat explanations of existing knowledge, but to record information that only you could obtain by being there.

A strong language model can write a clearer tutorial and assemble a more comprehensive survey. But it does not know the complete sequence of events from the first alert to the final diagnosis in that production incident; it does not know the bug that reproduces only on your machine; and it does not know why one set of experiments failed seventeen times in a row. Those facts have not yet been expressed, and therefore do not exist in the model’s training world.

Agents offer another possibility. An Agent that only searches webpages remains inside existing knowledge. An Agent connected to sensors, experimental equipment, and the physical environment may, like a human, bring new facts back into language space.

For now, however, a single Agent run still incurs substantial costs in time, capital, and access to a suitable environment, so it produces little verifiable new knowledge. Meanwhile, language models can generate content in bulk at almost no cost, further decoupling content quantity from boundary expansion. Agents can expand the boundary, but they still cannot change the trend of language space toward convergence.

In sum, the author believes that humanity’s language space is converging.

Cloudflare, The crawl before the fall… of referrals: understanding AI’s impact on content providers, 2025. ↩︎
Cloudflare, Content Independence Day, one year on: building the business model for the agentic Internet, 2026. ↩︎
Stack Exchange Questions API: November 2022 query; July 2026 query. The totals may change slightly as questions are deleted or restored. ↩︎
Turing, A. M. Computing Machinery and Intelligence. Mind, 1950, 59(236): 433–460. ↩︎

Internet | YuyaoGe's Website