The Ghost in China’s Machine

By /

(This article was originally published on Substack)

The case against China’s AI leadership is strong, well evidenced, and possibly irrelevant.

The Ghost in China's Machine

Here is a puzzle that does not resolve itself easily.

By almost every quantity you can count, China has already won the artificial intelligence race. It holds roughly three-quarters of the world’s AI patents. It leads the world in the volume of AI research papers published and in the citations those papers accumulate. Its best model, as of early 2026, trails the best American model by under three percentage points on the standard head-to-head benchmark — a gap that would have seemed impossible to close five years ago and is now, for practical purposes, closed. It installs more industrial robots than the rest of the world combined. Its researchers, trained in its own universities, produced DeepSeek — not a copy of an American system but a genuine efficiency breakthrough that forced every frontier lab on earth to recalculate what training a competitive model actually costs.

And yet.

Two numbers sit awkwardly beside all of that. The first comes from a 2026 working paper analyzing more than 1.5 million American and Chinese AI patents: Chinese inventors cite American frontier research far more intensively than American inventors cite Chinese work. The knowledge flows overwhelmingly in one direction. The country producing three-quarters of the world’s AI patents is still, in the most basic sense, a net importer of the ideas those patents build on.

The second number is stranger. Of the elite AI researchers educated in China — the ones publishing at the top conferences, the genuine frontier talent — only about eleven percent still work in China. That figure would be unremarkable if it were improving. It is not. It was sixteen percent in 2019. Over a period in which Chinese AI salaries rose sharply, domestic labs became genuinely world-class, national prestige in the field soared, and American immigration policy grew markedly more hostile to foreign researchers, China’s retention of its own best people got worse.

Something is happening here that money does not explain, and that chips do not explain either.

The standard accounts of China’s position in AI come in two flavors, and both are inadequate. The first says China is constrained by compute — that American export controls on advanced semiconductors impose a hard ceiling, and the story ends there. The second says China is unstoppable — that scale, state capital, and engineering talent will simply grind the gap to zero, and the story also ends there. Neither explains why a country with abundant talent, abundant capital, abundant data, and world-leading output volume keeps producing extraordinary execution while remaining dependent on ideas that originate elsewhere.

The explanation, I want to argue, has almost nothing to do with hardware. It has to do with a specific kind of freedom — one that is narrower than political liberty, cheaper than any semiconductor fab, and structurally very difficult for the Chinese state to grant.

To see why, we have to start somewhere that looks like a detour.

Table of Contents

What Creativity Actually Is

Most of what people believe about creative genius is wrong, and the errors matter for the argument that follows.

Start with the distinction between creativity and intelligence, which are not the same thing and are frequently confused. Intelligence, as psychometrics measures it, is largely convergent: working memory, processing speed, pattern recognition, the efficient arrival at a single correct answer. It is the ability to solve a well-posed problem. Creativity is largely divergent: generating many varied and loosely associated possibilities, and then recognizing which unusual one is worth pursuing. It is the ability to find a problem worth posing.

The relationship between them is captured by what researchers call threshold theory. Up to roughly an IQ of 120, intelligence and creative achievement correlate strongly — you need a certain amount of cognitive horsepower to do creative work in any serious domain. Above that threshold, the correlation largely disappears. Plenty of extremely intelligent people are not especially creative. Highly creative people are usually smart but rarely the smartest person in the room. The two capacities come apart.

This matters because they produce different things. Thomas Kuhn’s account of scientific revolutions turns on exactly this distinction. Normal science — the patient, rigorous extension of an existing framework — is intelligence-dominant work, and it is most of what science is. Paradigm shifts are something else. They are not reached by extending existing logic further; they require breaking the framework and seeing the same evidence through a fundamentally different structure. And Kuhn’s uncomfortable corollary is that deep expertise in the old paradigm actively impedes this. A very smart person thoroughly trained in an existing framework is often the last to see past it, because their intelligence has been optimized to solve problems the old way.

Three findings from the research on creative achievement will do heavy lifting later, and each is counterintuitive.

The first is about volume. Dean Simonton’s historiometric work — the statistical study of eminent careers — found that the single best predictor of producing a masterpiece in any given period is not skill peaking or inspiration striking. It is sheer output. Creators with more total work have proportionally more hits and more misses at every career stage. The hit rate does not improve with eminence; the number of attempts does. Picasso produced over 20,000 works. Bach wrote a cantata a week for years. The romantic image of the rare, perfect, effortless masterpiece is mostly survivorship bias: we remember the one percent and forget the ninety-nine percent that made it statistically possible.

The second is about constraint. Teresa Amabile’s decades of research at Harvard Business School on creativity inside organizations found an inverted-U relationship. Too little time or too few resources kills creative output through sheer stress. But too much comfort also reduces it. Some constraint sharpens creative problem-solving; the optimum is meaningful pressure without crushing pressure. Amabile also found that what matters most is not autonomy over goals but autonomy over method — people are most creative when handed a clear problem and left entirely free in how to attack it. Micromanaging process is among the fastest ways to destroy creative output, even when the goal itself is well chosen.

The third is about safety. Amy Edmondson’s research on psychological safety, later validated at scale inside Google, found that the strongest predictor of a team’s creative output was not the individual talent of its members. It was whether people felt safe voicing half-formed, possibly wrong ideas without social or professional punishment. This is the finding that matters most for what follows, and it is worth stating in its sharpest form: creative work requires the freedom to be publicly, expensively wrong. Not the freedom to criticize. Not the freedom to vote. The freedom to propose something that fails, in front of colleagues and superiors, repeatedly, without that failure ending you.

Two further points complete the picture. Creative individuals are disproportionately drawn from the socially marginal — immigrants, outsiders, people at the intersection of otherwise disconnected groups — not because marginality confers talent but because it forces the outsider’s eye that notices what insiders take for granted. And creativity is fundamentally recombinant: essentially every rigorous model of it converges on the same mechanism, which is the novel combination of existing elements rather than generation from nothing. Nobody creates from a void. The raw material is always imported from somewhere, which is why broad, eclectic exposure to ideas from outside one’s own field is one of the few interventions that reliably increases creative output.

Hold all of that. It is the measuring instrument.

What the AI Field Tells Us About Itself

Apply the instrument to artificial intelligence, and the field turns out to be an unusually clean case study — because its foundational moments are recent, well documented, and almost all of them were creative reframes rather than incremental improvements.

Convolutional neural networks, the architecture that made computer vision work, came from borrowing the structure of the biological visual cortex. Generative adversarial networks emerged in 2014 from reframing generative modeling as a two-player game between a forger and a detector — importing game theory wholesale into machine learning, reportedly conceived during an argument in a bar. Diffusion models, which power most modern image generation, were adapted directly from non-equilibrium statistical physics in a 2015 paper that took years to be recognized as practically important. Each of these is the signature creative move: taking a structure from a distant domain and discovering it solves a problem nobody had connected to it.

The transformer, the architecture underlying essentially every large language model in existence, is the purest example. The 2017 paper that introduced it was titled “Attention Is All You Need,” and its central move was subtractive. The field’s consensus held that sequence modeling required recurrence — processing text in order, carrying state forward. The transformer removed recurrence entirely and bet that attention alone would suffice. It was not an improvement to the existing framework. It was a deletion from it.

And the field’s own origin story is a textbook Kuhnian case. Geoffrey Hinton, Yann LeCun, and Yoshua Bengio spent roughly two decades pushing connectionist neural networks against a dominant symbolic-AI paradigm that regarded their approach as a dead end. They worked through the “AI winters” as marginal figures, funded thinly, publishing to limited interest, largely outside the mainstream of their own discipline. They were right, and the field eventually capitulated — but only after the evidence became overwhelming. That is what originating a paradigm shift actually looks like: prolonged public wrongness, professional marginality, and institutional hostility, sustained by conviction until the world catches up.

Now notice the pattern in what came after. The period from roughly 2018 to 2023 — GPT-2 through GPT-4 and their equivalents — was not primarily creative. It was the most impressive exercise in disciplined execution the technology industry has ever mounted: careful empirical work establishing scaling laws, meticulous data pipeline engineering, distributed systems work to train at unprecedented scale, rigorous ablation studies isolating what mattered. The architecture was largely fixed. The work was doing it properly, at enormous scale, with enormous capital. That is intelligence-dominant work, and it produced staggering results.

Richard Sutton’s 2019 essay “The Bitter Lesson” is the field theorizing about exactly this tension. His argument was that across AI’s history, hand-crafted human cleverness has repeatedly lost to simple general methods that scale with compute. It is a rare case of a discipline explicitly debating whether creativity or brute execution is its own primary engine.

And the current moment has swung back. As pure pretraining scaling showed diminishing returns, the frontier moved to reasoning models, test-time compute, and reinforcement-learning-based post-training — which is another conceptual reframe, shifting where computation happens rather than simply doing more of the same.

The pattern across the field’s history is consistent: creativity opens each new paradigm, intelligence exploits it to its limit, and diminishing returns eventually force another creative reframe. The two capacities alternate as the field’s dominant driver, on a cycle much faster than most sciences because AI iterates so quickly.

Which raises the question this essay exists to answer: where in that cycle does China sit, and can it move?

What China Has Built

It is worth being scrupulous here, because the argument I am building is easy to mistake for a familiar and largely discredited genre — the confident Western prediction that authoritarian systems cannot innovate, which has been made repeatedly since 1989 and has been wrong every single time.

China’s position is formidable, and the reasons for it are deep.

Start with the cultural infrastructure, which is not a recent construction. The imperial examination system, formalized under the Sui and Tang dynasties and matured under the Song, created something genuinely rare in premodern history: a bureaucratic elite selected by competitive examination rather than by birth or conquest. The culturally exemplary figure in Chinese society became the scholar-official, not the warrior-knight or the priest. That is roughly 1,300 years of institutionalized reverence for demonstrated intellectual achievement, and it did not disappear with the empire. Today’s gaokao intensity — 13.42 million students registered for the exam in 2024, a record — is the direct descendant of that tradition. When people describe Chinese academic pressure as culturally distinctive, they are describing a thirteen-century-old institution still running.

Then there is the state’s specific competence. China has been organizing massive coordinated infrastructure projects since the Yellow River flood-control systems of antiquity — a pattern of statecraft in which centralized mobilization toward a defined objective is the core skill. Applied to AI, this produces things no market-driven system can easily replicate: designating AI a national strategic priority in 2017 and sustaining that commitment across a decade; building data centers and power generation at a pace that makes Western permitting processes look ornamental; and financing all of it with state capital that is, as one analyst put it, confidence-insensitive and counter-cyclical — it does not require quarterly reassurance and can hold a losing position for a decade if the strategic logic holds.

That patience is a genuine advantage, and an underrated one. Western AI investment — $285.9 billion in private capital in 2025, against China’s $12.4 billion — is financed on equity valuations and debt markets, which means it is pro-cyclical. A demand disappointment tightens the loop and forces retrenchment precisely when competition intensifies. Chinese capital has no such reflex. It is twenty-three times smaller and structurally more durable.

And then there is DeepSeek, which deserves more weight than it usually gets in Western analysis. It was not a fast-follow. It was a genuine algorithmic innovation in training efficiency, achieved under hardware constraint, which changed the global cost structure of frontier model development. Crucially, nearly all the researchers behind its foundational papers were educated and trained in China — not the historical pattern of top Chinese talent doing its best work at Google or OpenAI. It is a direct counterexample to any simple claim that authoritarian systems cannot produce breakthrough technical creativity.

There is a second, subtler signal worth noting. In 2025, China overtook the United States in published research on responsible AI — alignment, interpretability, bias, governance — by 812 papers to 394, a sharp reversal from the year before. That domain is conceptually generative work, not routine engineering. It suggests the picture is not static.

And Simonton’s finding cuts in China’s favor over long horizons. A country producing three-quarters of the world’s AI patents is taking more shots on goal than anyone. At sufficient volume, some fraction of attempts will be transformative regardless of how unfavorable the surrounding conditions are. Creativity at scale can partially brute-force what it cannot culturally cultivate.

So the honest starting position is this: China is not behind. On volume, resources, patience, state coordination, and speed of real-economy adoption, it is ahead. The question is not whether China can compete. It is whether China can lead — and those turn out to be different problems requiring different capacities.

The Thing Money Cannot Buy

Run through the inputs. Talent: abundant, and the domestic pipeline is strengthening — 38% of researchers publishing at NeurIPS in 2024 did their undergraduate work in China, up from 29% five years earlier. Capital: smaller than America’s but patient and strategically directed. Data: vast. Energy: abundant. Expertise: world-class. Volume: unmatched. Compute: constrained, but the constraint is being actively engineered around, with domestic chips already 41% of the Chinese market and a coalition of over 2,000 firms targeting 70% semiconductor self-sufficiency by 2028.

Every ingredient the creativity research identifies is present, except one.

To see how the missing one operates, you need to understand that the Chinese censorship system is considerably more sophisticated than outsiders typically assume — and that its sophistication is exactly what makes its effects hard to see.

The landmark study here is by Gary King, Jennifer Pan, and Margaret Roberts, published in the American Political Science Review. They built a system to post real content across nearly 1,400 Chinese social media platforms and observe what survived. Their finding overturned the conventional understanding: posts containing negative, even vitriolic criticism of the state, its leaders, and its policies were not more likely to be censored. What was reliably removed was content with the potential to spur collective action — to help people gather, coordinate, or organize — regardless of whether that content was hostile to the government, supportive of it, or entirely unrelated to politics.

This is a crucial and genuinely clever design. The state permits enormous quantities of criticism, which serves as useful intelligence about local grievances and functions as a pressure valve. What it will not permit is the infrastructure of coordination. It does not suppress thought. It suppresses assembly.

At first glance, this should mean the system leaves creative research entirely alone. A researcher inventing a new attention mechanism is not coordinating anything. And to a significant extent, that is true — it is why domain separability works, and why we should expect excellent Chinese work in politically inert technical fields indefinitely.

But watch what happens when creative work becomes commercially successful and culturally influential.

Consider danmei — a genre of male-male romance and erotic fiction, overwhelmingly written and read by young women, which grew from a niche internet subculture into one of China’s largest literary phenomena. Dozens of titles topped bestseller lists; in 2021 alone, sixty were optioned for film or television; several major Chinese stars launched their careers in danmei adaptations. It is, by any reasonable measure, one of the most creatively generative cultural movements in recent Chinese history — original, genre-breaking, and organically popular.

The state’s response was prosecution. Television regulators banned the dramas outright in 2022. Writers migrated to a Taiwan-hosted platform accessible only through a VPN. Chinese police pursued them there. Since 2024, coordinated operations — one in Anhui province, another in Gansu — have arrested dozens of writers, mostly women in their twenties. The charge is producing or distributing obscene material for profit, under Article 363 of the criminal code.

Here is the detail that matters most. The sentencing under that statute scales to earnings. Minor offenses carry under three years; serious offenses three to ten; especially serious offenses ten years to life. One writer received four and a half years after thirty-seven of her thirty-eight novels were deemed obscene and her earnings were found to exceed 1.84 million yuan. Another, in an earlier case, got ten years for selling 7,000 copies.

Read that structure carefully. The more successful your creative work, the harsher your punishment — not because success made the content worse, but because the sentencing formula indexes to commercial scale. A system could hardly design a more precise instrument for teaching creative producers that the safe move is to stay small.

And the content asymmetry is documented. The relevant law explicitly targets explicit descriptions of gay sex and “other sexual perversions.” Heterosexual erotica receives markedly less scrutiny; a Nobel laureate’s novels containing graphic sexual content are published freely. One scholar interviewed about the crackdown suggested danmei was treated as especially subversive not only for its queerness but because it let women detach from gendered realities tied to marriage and motherhood — offering a fantasy space outside the roles the state increasingly wants women to occupy as it confronts its demographic crisis.

The danmei case is not about AI, obviously. It matters as a demonstration of mechanism: when genuinely original creative work in China becomes influential enough to matter, the system’s response has been to identify it, price its success as an aggravating factor, and punish it proportionally.

Now look at what a generation of young Chinese people did in response to a broader version of this pressure. The terms are neijuan — involution, the exhausting competition for diminishing returns — and tangping, lying flat: the deliberate withdrawal of effort, the strategic refusal to compete. Tangping became the internet word of the year in 2021 and has only grown since. State media denounced it as shameful defeatism, which is itself revealing.

Consider what tangping actually represents from a creativity standpoint. It is an enormously clever, culturally sophisticated, genuinely original collective response — drawing consciously on Daoist wu wei and Buddhist detachment, reframed for a hyper-competitive modern economy. It went viral, spawned an entire aesthetic, and reshaped how an entire generation talks about work. It is, in short, an impressive creative achievement. And what it produces is nothing. The most creative thing a cohort of young Chinese people has collectively generated in recent years is a philosophically grounded justification for opting out.

That is what happens to creative energy that cannot find a productive channel. It does not disappear. It reappears, transformed, as withdrawal.

Which brings us back to the eleven percent.

The retention figure is the tell, and it is the tell precisely because of what has happened to everything around it. Chinese AI compensation has risen dramatically — average monthly salaries for AI scientists exceeding 130,000 yuan, with individual packages reaching the equivalent of $14 million. Domestic labs have become genuinely world-class. National prestige in the field has soared. American immigration friction has intensified so severely that inbound AI researcher migration to the United States has fallen by 89% since 2017.

Every material and reputational factor has moved in China’s favor. And elite retention got worse.

That is not a money problem. It is not an infrastructure problem, or a prestige problem, or a visa problem. Something else is driving the most independently-minded researchers out — and the creativity literature tells us exactly what category of thing to look for. Not political liberty. Autonomy over method. Psychological safety. The freedom to be publicly, expensively wrong.

The Selection Problem

The question is not whether Chinese people are creative. That framing is both offensive and obviously false — the evidence against it runs from the Song dynasty through DeepSeek. The question is narrower and more tractable: what does the system select for, and what does it screen out?

Every institution has a selection filter. It is expressed in who gets promoted, who gets resources, who gets authority, and who absorbs the cost when something fails. The filter is rarely written down, but it is legible in outcomes, and over time it determines which cognitive profiles accumulate at the top of an organization and which drain out of it.

Run China’s filter against the four criteria from Section II.

Autonomy over method. The Chinese system grants autonomy over method selectively — within approved domains, to individuals who have already been vetted, subject to periodic review. This is not micromanagement in the crude sense; Chinese researchers are not being told which loss function to use. But Amabile’s finding concerns the reliability of that autonomy, not its average level. Autonomy that can be revoked is not autonomy; it is a permission, and permissions shape behavior by their revocability rather than by how often they are actually revoked. A researcher who knows their latitude is conditional will choose differently from one who knows it is not.

Tolerance for ambiguity. Institutions inherit this from their governing priorities, and China’s governing priority, stated openly and consistently, is stability. A system organized around the prevention of disorder is poorly configured to sit inside unresolved problems — its institutional reflex is to close ambiguity, to convert open questions into managed programs with milestones and deliverables. That reflex produces excellent five-year plans. It does not produce the long unresolved incubation from which reframes emerge.

Psychological safety. This is the sharpest failure, and the danmei prosecutions are the cleanest demonstration of the mechanism. Recall that the sentencing formula scales to commercial success. What that structure teaches is not “do not be creative” — it is “do not be creative at scale.” The lesson generalizes far beyond erotic fiction, because the underlying legal instruments are vague and the enforcement is opportunistic. Edmondson’s research is unambiguous that creative output depends on whether people can be wrong in public without that wrongness ending them. A system in which the penalty for influential work indexes to how influential it became inverts exactly that condition.

Willingness to defy authority. Here the selection filter is at its most direct, because this criterion is in tension with the system’s core promotion logic. Advancement in the Chinese party-state runs on demonstrated loyalty and factional cleanliness. That is a rational design for governing a country of 1.4 billion people. It is also, precisely, a filter that promotes those who have never publicly contradicted a superior and screens out those who have. And this is the criterion Kuhn identified as indispensable for paradigm shifts — the willingness to tell a field, at professional cost, that its framework is wrong. Hinton, LeCun, and Bengio were not merely correct; they were correct against institutional consensus for two decades. A selection system optimized for loyalty will not reliably produce that profile at the top, because that profile is what the system is built to filter.

Against this, the same filter performs superbly on the criteria that govern the exploitation phase: volume, absorption, patience, and coordinated execution at scale. That is not a coincidence or a consolation prize. It is the same filter producing exactly what it was designed to produce.

The filter, measured

The difficulty with selection arguments is that they are hard to verify. “The system screens out nonconformists” is unfalsifiable if nonconformity is unrecorded — nobody publishes statistics on temperamental independence.

But one dimension of the filter is recorded, comprehensively and over time, which makes it the best available proxy for how the mechanism behaves.

No woman has served on China’s Politburo since March 2023. No woman has ever served on the Politburo Standing Committee in the history of the People’s Republic. Eleven of 205 Central Committee members are women. At the provincial level, women hold roughly three percent of leadership positions — five percent at the city level, nine percent at the county level. Note the gradient: representation shrinks the closer one gets to actual decision-making authority, which is the signature of a filter rather than a pipeline problem.

This is not offered as a metaphor for creative exclusion. It is offered as a measurement of it. Women in China are not less educated than men — the educational gender gap is essentially closed, and women are a majority of university students. They are not absent from the workforce; female labor participation exceeds that of the United States or Japan. What happens to them happens at the point of promotion into authority, and it happens more severely the higher the authority. That is the filter operating on the one variable we can count.

And when the excluded have attempted to organize — the #MeToo movement, the five feminist activists arrested in 2015 for planning an anti-harassment protest, the journalist sentenced to five years in 2024 for reporting on sexual abuse — the response has been suppression, for the structural reason identified in Section V: organizing is what the censorship system is specifically engineered to prevent. Criticism is survivable. Coordination is not.

The broader point is that creative achievement is disproportionately produced by the socially marginal — not because marginality confers talent, but because it forces the outsider’s eye that notices what insiders take for granted. China’s promotion system is, by explicit design, a machine for producing insiders: vetted, loyalty-screened, factionally clean, and advanced precisely in proportion to how thoroughly they have internalized the existing framework. It is very good at this. It is producing exactly the people it intends to produce.

The configuration problem

Research on creativity’s cognitive substrate converges on a specific pairing: creative achievement peaks when loosened associative filtering is combined with strong executive control — enough looseness to generate unusual connections, enough control to build something coherent out of them. Either capacity alone underperforms. Loose association without control produces noise; control without looseness produces competent extension of what already exists.

Read institutionally rather than psychologically, this is the whole problem in one sentence. The Chinese system has executive control at a level no democratic state can match — planning horizons, resource mobilization, coordination capacity, sustained strategic commitment across decades. What it screens out, through the promotion mechanics described above, is the other half of the pairing. And the two are not merely separate. The mechanism that produces the first is the same mechanism that excludes the second.

Why This Might Be Wrong

An argument this tidy should be stress-tested, and there are three serious objections.

The first is domain separability, and it is the strongest. The Soviet Union produced world-class mathematics, physics, and aerospace engineering — Sputnik, Kolmogorov, Landau — under severe political repression, by walling off strategically valuable technical fields from ideological interference while crushing genetics, literature, and dissenting social thought. The pattern of “protect what is useful, suppress what is threatening” is a recurring and demonstrably workable feature of authoritarian systems that take science seriously.

And the King-Pan-Roberts finding makes this more plausible for China, not less. If the censorship apparatus targets coordination rather than thought, it may simply not bind where architectural AI research happens. A researcher proposing a novel training objective is not assembling a crowd. The walled garden — a handful of elite labs with real internal freedom, insulated from the broader political environment — is a genuinely viable structure, and it may already exist in places we cannot observe from outside.

The second objection is the track record, which is embarrassing for arguments of this shape. The claim that technological and economic advancement will eventually force authoritarian political change — the “dictator’s dilemma” — has been confidently applied to China since 1989 and has failed every time. Markets were supposed to do it. The internet was supposed to do it; Bill Clinton famously compared controlling it to nailing Jell-O to the wall. Private wealth was supposed to do it. In each case the system absorbed the supposedly fatal input and emerged more capable, not less.

This is not an accident, and it connects to something much older. The recurring pattern of Chinese history is absorption: conquered repeatedly by steppe powers — the Mongols, the Manchus — and in each case metabolizing the conquerors into its own administrative and cultural logic rather than being replaced by them. That is two thousand years of demonstrated competence at digesting things predicted to destroy it. Anyone arguing that a new input will prove indigestible is betting against a very long record.

The third objection is the most interesting, and it is genuinely new: AI itself may substitute for the human originators the system screens out. My entire argument assumes that the paradigm-generating step must be performed by a person — someone with the temperament to defy consensus and the institutional latitude to survive doing so. But if increasingly capable AI systems perform more of the search, recombination, and hypothesis generation themselves, a state could obtain the fruits of creative cognition without ever employing the kind of person its promotion filter excludes. That would be the first genuine escape hatch from the dictator’s dilemma in its history — relocating the disruptive cognitive function into a system that does not organize, does not resent, and does not emigrate.

It is also recursive. Better AI means more automated origination, which means less dependence on human mavericks, which means the institutional cost of screening them out falls rather than rises. This is the solution a stability-optimizing system would find most attractive, and it is the one development that would most cleanly invalidate the argument of this essay.

There is one more complication, and it is a real paradox rather than an objection.

Recall Amabile’s inverted U: moderate constraint sharpens creativity, comfort dulls it. DeepSeek’s efficiency breakthroughs happened because of hardware scarcity — the team could not brute-force the problem, so they had to think differently about it. Which raises an uncomfortable possibility: American export controls may be manufacturing exactly the creative capability they were designed to deny. And its corollary is stranger still. If China achieves compute abundance, it loses the scarcity pressure that produced its most genuinely original work. The optimal configuration for Chinese AI creativity might be permanent moderate constraint — enough compute to compete, never enough to relax.

What Would Have to Change

If the diagnosis is right, the remedy is narrower than “democratize,” and there is a precedent for exactly this kind of narrowness.

In 1978, Deng Xiaoping faced a structurally identical problem. He needed market dynamism without political liberalization, and the prevailing view — in Beijing and in Washington — was that the two were inseparable. His solution was geographic fencing: Special Economic Zones where capitalist rules applied inside a bounded perimeter, while the political system outside remained untouched. It was widely predicted either to fail or to metastasize into political change. It did neither. It worked for forty years.

The analogous move is cognitive fencing. Not liberalization, but bounded enclaves — a handful of designated labs — where research autonomy, tolerance for failed bets, psychological safety, and genuine freedom to challenge technical orthodoxy are guaranteed by central authority, precisely because they are walled off from anything touching coordination or political legitimacy.

This is the minimum viable version of what the selection filter currently screens out. It does not require the Chinese state to reform its promotion mechanics, which are load-bearing for everything else it does. It requires it to do what it did once before: identify the specific freedom that generates the needed output, grant exactly that freedom, fence it, and — this is the hard part — defend the fence against its own enforcement apparatus.

That last requirement is where it gets difficult, and the danmei case shows why. Those prosecutions were not purely ideological. Reporting indicates that cash-strapped local police departments were financially incentivized, extracting settlement payments from writers and their families, with fines calibrated to earnings. The system’s creative producers were punished not only by central policy but by opportunistic local predation operating under the cover of a vague law. Any protected enclave would have to be defended against that dynamic — which means building an institution whose explicit purpose is to protect people from the government that built it.

That is a rare thing for any state to do. It is not impossible: Deng proved it. But nothing in the current trajectory suggests it is happening. The direction of travel on research autonomy, civil society, private-sector champions, and queer digital space has been tightening, not fencing.

Where This Leaves Us

Return to the puzzle we started with. A country producing three-quarters of the world’s AI patents, still net-importing the ideas underneath them. A country whose elite talent retention is falling while everything that should improve it improves.

The answer, I think, is this: China is likely to arrive at paradigm shifts through sheer volume and relentless exploitation, rather than to originate them through the loose, unsafe, authority-defying exploration that has historically produced them.

That is not a prediction of failure. Simonton’s finding means enough shots on goal eventually produce something transformative regardless of conditions — creativity at sufficient scale can partially brute-force what it cannot cultivate. China will produce extraordinary work. It already has. It will very likely lead the world in AI deployment, in cost efficiency, in embedding the technology into the real economy, and in the sheer volume of capable systems shipped. On several of those dimensions it is already ahead, and the American lead in capital and compute is not a permanent possession.

But arriving at a frontier and defining one are different acts, and they require different institutional capacities. The thing worth noticing is that holding a lead is structurally harder than chasing one, and for a precise reason: the follower needs absorption and execution, while the leader has nothing left to absorb and must originate. The citation asymmetry is the signature of a superb follower. The moment China is the frontier, that competence stops paying — and the competence that has to replace it is the one its own promotion system is built to screen out.

So watch for four things, in roughly this order of diagnostic value.

Does a Chinese lab ship something genuinely subtractive — a reframe that discards received architecture rather than optimizing it, the way the transformer discarded recurrence? That would be the clearest technical signal.

Does elite talent retention inflect upward? Eleven percent, rising, would mean something has changed that money did not change.

Does research in the conceptually generative, politically adjacent domains keep growing — alignment, interpretability, the questions about what systems should want and whom they should obey? The 812-to-394 reversal in responsible-AI publishing was a real signal, and worth tracking.

And finally, the institutional tell: is anyone ever granted real, protected, publicly announced authority despite failing the loyalty filter — a lab with guaranteed latitude, a researcher with standing permission to be wrong, an enclave where expensive failure is officially survivable? That would be the Special Economic Zone moment for cognition, and it would be visible when it happened.

As of now, nothing suggests it is. And the reason is the difficulty at the heart of the whole problem: the filter that screens out the originators is the same filter that produces the coordination, the patience, and the executional scale that make China formidable in the first place. It cannot be selectively disabled for the people one would have wanted to keep. A system does not get to run its selection mechanism only on the candidates it is happy to lose.

A Note on Process: This piece was written with the assistance of Claude, which aided in research, structuring, and drafting. All arguments, interpretations, and final editorial judgments remain entirely the author’s. All referenced sources have been verified and documented below.

Sources

All figures current as of September 2026.

Data and primary research

The danmei prosecutions

Tangping and neijuan

Political representation

Semiconductors and compute

On artificial intelligence

On creativity

  • Kuhn, Thomas S. (1962), The Structure of Scientific Revolutions. University of Chicago Press
  • Csikszentmihalyi, Mihaly (1996), Creativity: Flow and the Psychology of Discovery and Invention. HarperCollins
  • Getzels, Jacob & Csikszentmihalyi, Mihaly (1976), The Creative Vision. Wiley
  • Amabile, Teresa & Kramer, Steven (2011), The Progress Principle. Harvard Business Review Press
  • Edmondson, Amy (1999), “Psychological Safety and Learning Behavior in Work Teams,” Administrative Science Quarterly 44(2)
  • Carson, Peterson & Higgins (2003), “Decreased Latent Inhibition Is Associated with Increased Creative Achievement in High-Functioning Individuals,” Journal of Personality and Social Psychology 85(3)
  • Simonton, Dean Keith (1999), Origins of Genius: Darwinian Perspectives on Creativity. Oxford University Press

Discover more from Numerons

Subscribe to get the latest posts sent to your email.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top