In five years, traffic from AI on the internet will be a thousand times that from humans. Humans will no longer be the main users surfing the web—they will be little more than a rounding error in cyberspace. This is a prediction from Cloudflare in a recent earnings call, and it is not alarmist: right now, AI visits have already officially surpassed human internet users.
No one has seriously considered whether that pasture will still grow grass in the future.
According to estimates, the training of language models will exhaust all publicly available human text data between 2026 and 2032. The depletion of high‑quality language data may even arrive as early as 2026, two years earlier than the overall exhaustion of public text.
The Sinking of Stack Overflow
Stack Overflow was once the collective brain of the programming world.
From its launch in 2008 to its peak in 2014, it attracted over 200,000 new questions per month, building a vast archive of human intelligence.
Any engineer debugging at 3 a.m. knows the feeling: you search with an exact error message, and the first result is almost always a five‑year‑old Q&A on Stack Overflow, complete with dozens of votes and rebuttals.
But in the two years since ChatGPT appeared at the end of 2022, Stack Overflow's monthly question volume has dropped from over 200,000 to less than 50,000—back to roughly its 2009 level, when the platform had just launched.
The questions have dried up. People no longer go there because they ask AI instead.
According to Stack Overflow's own 2025 survey, 84% of developers use AI tools daily. At the same time, 46% of developers do not trust the accuracy of AI output; their biggest concern is not that AI gives wrong answers, but that it produces something "that looks correct, but is just a bit off"—sounding like an expert's answer, which easily lowers one's guard and places higher demands on the human reviewer who has to verify it.
About 35% of developers say that they sometimes visit Stack Overflow to fix, understand, or debug problems created by AI.
By analyzing the behavioral trajectories of 24,304 active contributors over the 17 months following ChatGPT's launch, researchers also found that those with the most advanced skills and best reputations on the platform were more likely to leave.
As a result, the proportion of unanswered questions on the platform rose from 19.9% to 25.7%, and overall real contribution volume dropped by about 35%.
Things are getting absurd: Stack Overflow's knowledge trained AI; AI took away Stack Overflow's questions; developers return to Stack Overflow to find fixes for the AI.
But without a steady stream of new questions, old answers will gradually become outdated. Answers from five years ago may still solve problems from five years ago, but they cannot know about frameworks released today, vulnerabilities discovered today, or a new bug no one has ever seen before.
Models can provide answers, but they cannot foresee the next question for humanity.
Free Riding, or the Tragedy of the Commons
Economics has a classic proposition called the "tragedy of the commons." In a shared pasture, each herder has an incentive to add one more cow, because the benefit is private while the cost is shared; eventually the pasture is overused and collapses.
In the past, people shared knowledge because of an implicit social contract: you contribute, others contribute, the whole community benefits, and you gain reputation, recognition, or learning feedback through your participation—a virtuous cycle. Stack Overflow's voting mechanism and answer ranking were designed to incentivize participation. Wikipedia's editing culture works the same way.
But AI is accelerating the departure of expert contributors. When a tool can produce a good enough answer at near‑zero cost, the marginal benefit of "sharing professional knowledge" shrinks rapidly. A database architect with 20 years of experience might spend three hours writing an in‑depth article on PostgreSQL tuning; anyone can now get a similar AI‑synthesized piece with just one question. So why should he still write?
You might say: because of passion, because of the desire to help others. And indeed, such people exist. But human time is limited. As feedback from sharing dwindles—fewer comments, fewer citations, fewer readers—the intrinsic drive needed to persist grows stronger, and those who can sustain such drive are always a minority in any group.
The result may be that valuable knowledge moves into corporate intranets, paid communities, and private chats; or that humans produce less knowledge, AI trains on even scarcer data, knowledge quality declines, and the entire system spirals downward.
A Computer‑Science Version of Inbreeding
Over the past thirty years, the internet has built a cathedral of human knowledge, with bricks laid by countless strangers who answered questions seriously on BBSs, volunteers who maintained Wikipedia entries tirelessly, and engineers and researchers who argued passionately on forums.
AI learned to speak from that cathedral.
Now, fewer visitors come to the cathedral, and fewer new bricks are laid. AI begins to fill the silence with its own echoes, to the point that researchers are already talking about the "inertialization" of knowledge: updates stop, AI consumes an aging inheritance, then regenerates the aging inheritance, ages again, regenerates again.
In July 2024, Ilia Shumailov, a research scientist formerly at Google DeepMind, and colleagues published a widely noted paper in Nature: "AI models collapse when trained on recursively generated data." The paper points out that large language models, variational autoencoders, and similar systems degrade in performance when repeatedly trained on content generated by themselves. In early stages, rare patterns disappear; in later stages, the overall output becomes monotonous, drifting toward averages, with odd outliers.
Some mainstream science media in the U.S. cited a vivid phrase: "Training AI on AI‑generated text is the computer‑science version of inbreeding."
As AI‑generated content floods the internet, and when that content is in turn used as training data, errors in the feedback loop accumulate continuously, producing increasing deviation between the model and the real‑world distribution.
Who Will Grow the Next Crop of Grass for AI?
Suppose 15 years from now, AI reaches "expert‑level" proficiency in most knowledge domains. By then, who will still have enough knowledge to judge whether AI's answers are correct?
Overseeing AI requires people who understand what AI is doing. But the process of cultivating such people has itself been taken over by AI. The next generation will rely on AI to learn, to work, to make judgments, and ultimately to oversee AI itself.
Some might say that no oversight is needed, just as we don't need to verify the results from a calculator. Yet we know that a calculator simply executes deterministic arithmetic. AI is not a calculator. When AI's complexity fully surpasses the cognitive limits of individual humans, our trust in it degenerates—from rational verification to blind faith.
A 2026 article by the World Economic Forum noted that the more capable AI becomes, the weaker effective human oversight becomes, because humans perform less and less cognitive work and thus lose first‑hand mastery of that work.
In the future, might we need to require AI users to periodically complete difficult cognitive tasks independently, much as pilots are required to fly manually from time to time, and establish some kind of capability‑audit mechanism to ensure that human judgment does not atrophy unnoticed?
Humans have a strong tendency: once a tool drastically lowers the barrier to a certain type of knowledge, new and higher‑level needs emerge. The addition of AI may force knowledge production toward depths that "cannot be replaced." But this requires a precondition: the value of cognitive labor must be seen, recognized, and returned to creators in some way.
We need to give answers more provenance, and more importantly, preserve for the next generation the opportunity to learn independently and make mistakes without relying on AI.
The collective wisdom of human civilization is not an already‑downloaded data package. It consists of collective memory, collective attention, and collective reasoning; it requires living people to continuously debate, revise, and supplement. AI can enhance parts of it, but if human participation itself shrinks, these three elements lose their living source, and what remains is an increasingly refined echo chamber.
AI can distill all past knowledge, but it cannot create out of thin air a group of people willing to take responsibility for the future.
We have long been saying that models need computing power and tokens, but there is another critical need: people who go out into the real world, discover new problems, and are willing to leave the answers publicly available.
The best knowledge base is not necessarily the one that provides the most authoritative and correct answers, but the one that always has people coming back to revise and update.
AI is racing ahead, but the knowledge pasture will not grow the next crop of grass just because the cattle run faster.
