When Humanity Knows More but Thinks More Alike

Artificial intelligence makes writing faster. That is the visible effect, but not the deepest one. The important effect is that AI changes the structure of knowledge production: what gets written, how similar the outputs are to one another, and, because future models will learn from those outputs, what the next generation of AI will know. This essay argues that the resulting convergence of knowledge is neither a disaster nor a free lunch. It is a trade. The danger is not that AI will make people ignorant; it is that it may make informed people more alike. Convergence delivers real efficiency gains, and those gains are worth taking. But the same process, pushed past a threshold, begins to consume the raw material that knowledge creation depends on. The design problem, if there is one, is not to stop convergence but to know where the threshold sits.

From Slow Content to Fast Content

Before AI, writing was slow content. An author searched, read, took notes, misremembered, interpreted, drafted, and revised through a long personal process. The slowness was inefficient, and it preserved something: divergence. Different authors read different books, followed different teachers, misunderstood different ideas, and carried different experiences into their work. The output was not just a polished statement of public knowledge. It was a trace of one individual's path through knowledge.

After AI, writing becomes fast content: a hybrid of the model's internal representation, retrieved references, the author's prompts, and the author's judgment. When the author stays deeply involved, the product still carries human experience and taste. But the human contribution now travels through a shared interface. If many authors use similar models, similar retrieval systems, and similar prompting habits, their outputs become correlated even when each author works honestly and alone.^[1]

This is the paradox of fast content. AI raises the floor of public writing while lowering the marginal uniqueness of each piece. A mediocre essay becomes competent. A scattered argument becomes coherent. But if thousands of people produce polished versions of the same AI-mediated synthesis, each new piece contains less independent signal. The content looks better; its incremental value to the knowledge system is smaller.

Three Compressions

The mechanism runs through compression, and it operates three times.

The first compression happens at training. A large language model is not a lossless archive of human knowledge. It is trained for performance: prediction, coherence, usefulness, alignment with human feedback. To achieve this, it reduces the historical distribution of knowledge, with its schools, failed theories, minority readings, local styles, and peripheral arguments, into a usable internal representation. That representation is powerful and selective. Peripheral versions are not deleted so much as down-weighted into near-invisibility.

The second compression happens at delivery. When a user asks a question, the model gives one answer. It may gesture at caveats, but by default it produces the coherent, balanced version that feels like the reasonable center of the field. The full distribution of disagreement is often available on request, but most users do not know when to request it. The interface compresses even where the weights do not.

The third compression is new, and it is the one that turns a static loss into a dynamic process. AI-mediated syntheses increasingly flow back into the public corpus from which future systems learn. One might object that serious labs filter machine-generated text out of their training data, and they do.^[2] But the filter catches the wrong object. What flows back is not raw model output; it is fast content: human-signed, human-judged, edited, published under real names, and correlated all the same. No ordinary filter can remove it, and no filter should try to remove all of it, because it is legitimate human writing. The correlation travels inside the human text. Each generation of models therefore learns from a corpus in which the center-seeking material occupies a growing share and the divergent, path-dependent material occupies a shrinking one. Compression compounds.

This recursion is what distinguishes the current moment from earlier information technologies, though the difference is one of degree becoming one of kind. Earlier media had feedback loops too: print fed on print, journalism quoted journalism, citation systems amplified whatever they had already amplified. What is new is that synthesis itself is automated, immediate, and universal, and that the synthesized output reenters the foundations of the next system within a single model generation. Books transmitted versions. Universities curated versions. Search engines ranked versions. AI synthesizes versions, and the syntheses now feed the next round of synthesis. The nearest analogy is a market where this period's prices are set partly by extrapolating last period's prices: the system can drift toward internal consistency rather than external truth, and the drift is invisible from inside because every local step looks like an improvement.

The Case for Convergence

Before treating this as a crisis, the gains deserve a full hearing, because they are large and they are real.

Most knowledge work is not discovery. It is transmission, application, and coordination, and all three benefit from standardization. Industrialization raised productivity by eliminating unnecessary variation: interchangeable parts, standard measures, common protocols. A factory does not want every worker inventing a new screw size; a hospital does not want every nurse improvising dosage conventions. Complex systems scale on standards.

Much of the written world has the same character. Documentation, contracts, curricula, routine analysis, standard literature summaries, and instructional material do not need endless reinvention. Variation here is mostly noise: idiosyncratic terminology, redundant explanation, avoidable error. When AI compresses this material toward a competent center, society genuinely gains. The floor rises. Bad versions die faster. A student in a small town gets an explanation as clear as one at an elite university. The waste that convergence eliminates was never romantic; it was waste.

There is also a subtler gain. A common interface creates a common language, and common language is what allows knowledge to accumulate at all. A field where everyone uses private terms and private methods cannot build on itself; debate becomes noise. Some convergence is not the enemy of knowledge creation. It is its precondition.

The honest version of the argument, then, is not that convergence is bad. It is that convergence has sharply diminishing and eventually negative returns, and that the recursive structure of AI training makes it hard to stop at the optimum.

Where the Trade Turns

Knowledge creation needs variation the way evolution needs mutation. Not maximum variation: most mutations are harmful, and most intellectual divergence is error, confusion, or low-quality repetition. What the system needs is an interior optimum, a diversity band.^[3] Too much variation and the field fragments; nothing accumulates. Too little and the field hardens; the dominant frame becomes too comfortable, and young scholars learn the answer before they learn the puzzle. Industrialization succeeds by eliminating unnecessary variation. Knowledge creation advances by preserving necessary variation. The whole problem is telling the two apart.

The Bayesian framing makes the cost precise. Fast content looks strictly better than slow content: the old knowledge stock, updated with new information, enriched by retrieval, shaped by judgment. An improved posterior. The flaw in this picture is correlated updating.^[4] If millions of users update through the same models and the same retrieval, they are no longer independent learners; they are partially shared learners drawing on a common compressed prior. When draws are correlated, adding documents adds little information. The corpus may grow while its effective sample size falls. Each individual answer is better on average, and the ensemble is worth less.

The recursive loop then converts this static cost into a dynamic one. In a single generation, correlated content is a modest problem: the classics still sit in the corpus, and the divergent tails still exist. Over repeated generations of models trained on the previous generation's textual environment, the tails thin. This is why classics may hold their training value better than fast content, not because they are old but because they were produced before the compression layer existed. They carry high-divergence human compression: independent paths through the material. A classic carries a distinct signal; a fast-content essay carries a correlated one. In that sense, classics become not only cultural objects but distributional anchors, fixed points of independent variation in a corpus whose new additions grow steadily more alike.

None of this requires anyone to act badly. No censor is needed. Convergence emerges from optimization: helpful answers, satisfied users, benchmark performance, demand for polish. The danger is not that AI gives bad answers. The danger is that good answers become too similar, and that the similarity feeds forward.

Not All Fields Are the Same

The threshold sits in different places in different fields, which suggests a rough taxonomy.

Mature, confined fields have stable, cumulative, widely agreed cores: basic mathematics, standard grammar, routine coding patterns, elementary physics, conventional accounting. Here convergence is close to pure gain. These fields sit far inside the safe region of the diversity band, and compression mostly eliminates error.

Contested frontier fields are domains where truth is genuinely unsettled and disagreement carries information: macroeconomic regimes, institutional development, the effects of social media, the future of labor markets, the nature of machine intelligence itself. Here premature convergence is expensive. Compressing a live debate into a smooth consensus answer does not resolve the debate; it hides it, and hides it precisely from the general reader who most needs to know it exists.

Pseudo-mature fields are the dangerous category: fields that look settled because one framework dominates, where the dominance reflects institutional incentives, publication patterns, data limitations, or historical accident rather than genuine closure.^[5] Here compression does not merely hide disagreement; it converts social consensus into apparent epistemic truth. The uncomfortable truth about this category is that it can only be identified with confidence in retrospect. Pre-1970s macroeconomics looked mature until stagflation broke it. Continental drift was a fringe theory for half a century while geology considered itself settled. Parts of nutritional science spent decades in confident consensus that later became far more contested. At the time, each looked like a mature field; the pseudo-maturity became visible only when the anomalies won. This retrospect problem is exactly why the risk is serious: an AI system that treats today's consensus as ground truth cannot distinguish a field that has converged on the truth from a field that has merely converged. Neither, honestly, can most humans. But the pre-AI system at least kept dissent alive in scattered journals, rival departments, and stubborn individuals, waiting for the anomaly. A compressed interface may not.

In mature fields, convergence is productive. In frontier fields, it is costly. In pseudo-mature fields, it is deceptive, and undetectably so until it is too late. The central design task is to build a compression layer that can tell these three apart; current systems generally cannot.

The Counterargument: Won't Model Plurality Save Us?

The strongest objection to this whole argument is that it assumes a monoculture. There are many labs, many models, open weights, fine-tunes, and users who deliberately prompt against the grain. Why should outputs correlate when the interfaces multiply?

The objection deserves a serious answer, and the answer is that plurality of vendors is not plurality of priors. The major models train on substantially overlapping corpora: the same public internet, the same books, the same code. They are tuned with similar human-feedback methods toward similar targets: helpfulness, coherence, inoffensiveness, the polished assistant register. They compete on the same benchmarks, which act as a shared selection pressure. Different companies, convergent optimization. The situation resembles index funds run by rival asset managers: fierce commercial competition, nearly identical portfolios, because they are all optimizing the same objective against the same data. Ten models drawn from one corpus and tuned toward one behavioral target are closer to one model than to ten.

There is also a genuine optimistic counterweight, and it should be conceded rather than buried. AI lowers the cost of heterodoxy as well as the cost of conformity. A researcher can now read outside her field cheaply, translate minority literatures, resurrect forgotten arguments, and stress-test a strange idea in an afternoon. The same interface that pulls users toward the center can, in deliberate hands, carry them to the periphery faster than any library ever did. The technology is not intrinsically convergent. The defaults are convergent.^[6] Almost everything in this essay is a claim about defaults, and defaults govern the mass of usage even when they bind no individual user.

The New Scarce Resource

Follow the economics of scarcity through the transition. After AI, information access is not scarce. Polished writing is not scarce. Competent synthesis is not scarce. What becomes scarce is independent signal: content whose information is not already implied by the common compressed prior.

Independent signal comes from private data, fieldwork, lived experience, new measurement, genuine theory, minority literatures, cross-domain analogy, and direct institutional participation. It also comes from a stance: the author's willingness to resist the model's default center rather than be absorbed by it. The future divide among knowledge workers will not run between those who have AI and those who do not; nearly everyone will have it. It will run between those who use AI to converge and those who use it to diverge productively. The first group produces competent fast content, and the market for competence is collapsing toward zero price. The second produces the input the whole system, including the training of future models, actually runs on.

This reframes what "human in the loop" should mean. Prompting for polish keeps a human in the loop and adds nothing the model could not reconstruct from the public corpus. The contribution that matters is orthogonal to the model: experience it has not ingested, judgment it cannot average into existence, risk it is tuned to avoid, a frame it has no reason to propose.

Conclusion: Managing the Band

The balanced conclusion is neither alarm nor complacency.

The gains are real and should be banked. Convergence in mature knowledge eliminates genuine waste, raises the public floor, and spreads competence more equally than any previous technology. Resisting this out of nostalgia for productive friction would be a mistake; most of the friction was never productive.

The risk is also real, and it is structural rather than sinister. Triple compression, at training, at delivery, and now recursively through the return of fast content to the corpus, pushes every field toward its center regardless of whether the field has earned a center. In mature fields this is fine. In contested fields it hides live debates. In pseudo-mature fields it hardens error, and no one inside the system can see it happening, because every local step looks like quality improvement.

The task, then, is not to choose between convergence and diversity but to manage the band: enough standardization to accumulate, enough disagreement to discover.^[7] Concretely, this means treating divergence as a measurable property of the knowledge system rather than a mood. One can ask of any field: how many serious schools survive, whether citation concentration is rising, whether new work adds independent signal or restates the dominant frame, whether minority theories still generate testable predictions. An AI layer that surfaced these measurements, that flagged contested fields as contested and refused to present pseudo-maturity as closure, would be acting as a thermostat rather than a compressor. Nothing in the technology forbids this. Nothing in the current incentives demands it.

Humanity will know more. That much is nearly certain. Whether it can know more without thinking too much alike depends on a design choice that is still open: whether the systems now being placed between people and the world's knowledge are optimized only for the smoothness of each answer, or also for the diversity of the whole.

Endnotes

[1] Early evidence points in this direction. Doshi and Hauser find that access to generative AI raises the judged quality of individual stories while reducing the collective diversity of the resulting pool: individual gains, ensemble losses. Anil R. Doshi and Oliver P. Hauser, "Generative AI Enhances Individual Creativity but Reduces the Collective Diversity of Novel Content," Science Advances 10, no. 28 (2024). Back

[2] The degenerate case, in which models train directly on unfiltered machine output, is analyzed in the model-collapse literature. Ilia Shumailov et al., "AI Models Collapse When Trained on Recursively Generated Data," Nature 631 (2024): 755–759. The argument here concerns a channel that survives the filtering which model collapse has prompted: human-authored content whose correlation is inherited from a shared AI interface rather than from direct machine generation. Back

[3] The trade-off between exploiting known solutions and exploring uncertain ones is a classic result in organizational learning: systems that over-exploit converge on competent mediocrity. James G. March, "Exploration and Exploitation in Organizational Learning," Organization Science 2, no. 1 (1991): 71–87. On why diverse problem-solvers can outperform abler but similar ones, see Lu Hong and Scott E. Page, "Groups of Diverse Problem Solvers Can Outperform Groups of High-Ability Problem Solvers," PNAS 101, no. 46 (2004): 16385–16389. Back

[4] The underlying logic is that of social learning and informational cascades: when agents observe common signals rather than independent ones, individual rationality is consistent with collective fragility. Sushil Bikhchandani, David Hirshleifer, and Ivo Welch, "A Theory of Fads, Fashion, Custom, and Cultural Change as Informational Cascades," Journal of Political Economy 100, no. 5 (1992): 992–1026; Morris H. DeGroot, "Reaching a Consensus," Journal of the American Statistical Association 69, no. 345 (1974): 118–121. Back

[5] The distinction between a paradigm that has closed a question and a paradigm that has merely suppressed its anomalies is Kuhn's. Thomas S. Kuhn, The Structure of Scientific Revolutions (Chicago: University of Chicago Press, 1962). Back

[6] The claim is behavioral, not architectural: default outputs cluster around a consensus register because that is what the tuning objective rewards, even though the underlying distribution remains queryable by a sufficiently deliberate user. Back

[7] The interior optimum has empirical counterparts. Studies of interdisciplinary research find an inverted-U relationship between the diversity of a paper's knowledge inputs and its citation impact, with both narrow and maximally scattered combinations underperforming intermediate ones: Alfredo Yegros-Yegros, Ismael Rafols, and Pablo D'Este, "Does Interdisciplinary Research Lead to Higher Citation Impact? The Different Effect of Proximal and Distal Interdisciplinarity," PLOS ONE 10, no. 8 (2015). Relatedly, the highest-impact science tends to combine deeply conventional knowledge with a small injection of atypical combinations, rather than novelty throughout: Brian Uzzi, Satyam Mukherjee, Michael Stringer, and Ben Jones, "Atypical Combinations and Scientific Impact," Science 342, no. 6157 (2013): 468–472. Back