I. A Question Left Over from the Previous Article
In the previous article, “Rational Infringement” in the Safe Harbor, we did the arithmetic: under the combination of the “safe harbor rule plus make-whole damages,” a leading platform's expected cost of infringement is divided by a probability far below 1, and “not buying the copyright” becomes the dominant strategy. The conclusion fit in a single sentence—pirate broadcasting at scale is not a moral problem; it is a math problem.
But that article left a more unsettling question outside the door.
If the arithmetic is this simple and the mechanism can be explained on a single page, then the question is: why has this bug gone unfixed for nearly thirty years? A rule enacted in the United States in 1998, transplanted by China in 2006, written into the Tort Liability Law (侵权责任法), then written into the Civil Code (民法典)—carried straight across at every step, utterly unchanged. That a rule so plainly broken has persisted for so long is itself a phenomenon in need of explanation.
There are two popular explanations, and I intend to use neither.
The first is the test-of-time theory: a rule that has run smoothly for nearly thirty years, without any major disaster, has evidently stood the test. This sounds reassuring, but the trouble is that it treats “being alive” as evidence of “deserving to be alive”—a rule's longevity proves only that no one could afford the cost of replacing it; it cannot prove that the rule still holds. Moreover, “why has it lived this long” is precisely the question we need to explain; to explain survival by survival is to say nothing at all.
The second is the inertia theory: legislative resources are scarce, and this amendment never gets a place in the queue. That much is true, but its explanatory power is too weak—it cannot explain why it is precisely this rule, of all rules, that remains guarded here like an impregnable fortress even after the country that invented it has itself admitted error.
This article offers a third explanation. To explain it clearly, we must first leave copyright law, go back four hundred years to the Ming dynasty, and look at a tax bill. One word in passing: the institution behind that tax bill ran smoothly for more than two hundred years.
II. First, a Story from the Ming Dynasty
The Ming dynasty had an institution called the scholar-gentry exemption (士绅优免): men of letters who had passed the examinations and earned degrees could be relieved of a portion of the corvée. Note how restrained the text was—what was waived was labor service, not the land tax, and the exemption was capped by acreage. The rationale was perfectly legitimate: the court had to nurture scholars, and scholars had to devote themselves to learning and moral cultivation; they could hardly be sent out to dig river channels and build roads.
On the day the rule was born, there was nothing wrong with it.
The problem lay in its operation. The acreage cap gradually became a dead letter, and evasion techniques sprang up to meet the demand—guiji (fraudulent entrustment of land to another's name), feisa ("flying sprinkling" of liabilities onto others' accounts), huafen ("flowery division" of holdings into multiple registrations). The names were legion; the effect was uniform: those who held degrees, as a practical matter, no longer paid taxes. More ingenious still, a derivative product grew out of the market, called touxian (投献), or "commendation": independent peasant proprietors would take the initiative to register their own fields under the name of some Master Juren—a provincial examination graduate—not only giving the land away outright, but every year paying an additional sum of "land rent" on top.
Giving away land and paying money for the privilege sounds like a swindle, but it was in fact arbitrage. The arithmetic was simple: once the land was attached to the degree-holder's name, the entire family slipped out of the tax-and-corvée system, and the taxes and services saved far exceeded the rent paid in reverse. The absurdity was not the peasants'; it belonged to the institution that made this arithmetic come out right.
So why did the authorities not investigate? This touches the crux of the matter: Ming tax administration operated on the principle that "the government does not go down to the countryside"—the assessment of the money-and-grain levies and the dunning of arrears were handled by the local gentry themselves. In other words, the probability that tax evasion would be investigated and punished was controlled by the very beneficiaries of that evasion. With tax inspection moonlighted by the largest taxpayers, the probability of enforcement naturally approached zero. Friends who read the previous essay will smile in recognition: there is that p again. Wherever p is institutionally pinned to zero, illegality is priced at zero; and illegality priced at zero will inevitably be produced at scale.
The story ends grimly. Those who could escape escaped ever more; those who could not were loaded ever heavier. The war in Liaodong demanded additional military pay, and more than five million taels of silver were apportioned across the acreage—nominally everyone's share, in reality pressed entirely onto the backs of independent peasants who had no degree and no avenue of attachment. The tax base shrank ever narrower, the surtaxes grew ever harsher, and the positive feedback loop cycled all the way to the dynasty's collapse.
Now pay attention to the most intriguing point in this story.
The qingyi—the "pure discourse" of the literati—that in those years opposed the cadastral resurvey of farmland and opposed the taxation of commerce argued entirely in terms of public reasons: "do not compete with the people for profit," "benevolent government shows compassion for the merchants." And—this is the key—most of the people who said these things were sincere. They genuinely believed they were defending all living souls under heaven. A system can channel benefits to a particular stratum without anyone having to lie at any point in the process: the text is neutral, the reasons are respectable, the participants are sincere—while the actual incidence of the tax burden keeps shifting toward the side least able to resist.
I call this condition "structural interest camouflage." Its definition takes only one sentence: a rule's justificatory argument froze on the day the rule was born, while the cost structure of the world had long since turned upside down; and so the argument is still the same argument, but the beneficiary is no longer the beneficiary the argument described.
The camouflage artist is no person. It is time.
Remember the four components in this story; the discussion below will match each to its counterpart, one by one: the text is neutral; the costs are hidden inside the procedure; labels automatically allocate the burdens; the premise has died, but the rule lives on.
III. The Safe Harbor's Birth Certificate
Let us return to the safe harbor. Before anything else, let us do something that may surprise some readers: say a few fair words on its behalf. This is not courtesy; it is logical necessity — only by first proving that it was right at birth can we prove that it is now past its expiration date. If it had been a bad rule from the very beginning, the diagnosis of "expiration" would have been attached to the wrong patient.
In 1998, when the United States Congress enacted Section 512 of the DMCA, what did the "online service provider" before its eyes look like? A warehouse. A conduit. Users deposited content; the platform stored it exactly as it came in and transmitted it exactly as it was stored — it did not select, did not rank, did not recommend. Working from this picture, the legislature accepted four premises, and every one of them was true at the time:
First, the platform was a passive conduit. A warehouse keeper does not know — and does not need to know — the rights status of every item of goods in storage.
Second, the costs of identification were asymmetric. Given the technology of the day, requiring platforms to review massive volumes of content item by item would have cost an astronomical sum, while the right holder could recognize its own works at the lowest cost of anyone. Hence the arrangement: the right holder discovers the infringement and sends a notice; the platform takes the content down. This was a correct application of the economic principle that responsibility should fall on the party who can prevent the harm at the lowest cost — the "least-cost avoider" principle.
Third, the industry was still an infant. The Internet had just emerged from its swaddling clothes; a flood of liability would have crushed it dead at a single stroke. To grant a newborn industry a limitation of liability in exchange for its cooperation in governance was a bargain well worth striking.
Fourth, infringement was a marginal event. Infringement in that era was imagined as the occasional misconduct of individual users — it was the noise of the operation, not part of the operation. For noise, a procedure of "notice received, takedown on a case-by-case basis" was perfectly proportionate.
All four premises were true in 1998; and the rule, in 1998, was entirely good.
The safe harbor's problem has never been its birth, but its refusal to die.
IV. Four Death Certificates
Four premises, and today not one of them remains. Time for the autopsy, one certificate at a time.
The passive conduit—killed by algorithmic recommendation. Today, the core asset of the leading platforms is not storage but distribution: for every piece of content, for every user, the algorithm selects, ranks, weights, and pushes, and it monetizes according to how well that distribution performs. What is this? It is the industrialization of editorial judgment—a traditional editor passes judgment on a work once; the algorithm does so tens of billions of times a day. The warehouse keeper has not only begun deciding which shop window each item is displayed in and which customer it is pushed to; he also takes a commission on sales. To wield an editor's power while invoking a conduit's immunity—that bridge collapsed long ago.
Cost asymmetry—killed by recognition technology. Back then, the claim was that platforms simply could not review everything. Today, audio and video fingerprinting has turned "identifying a re-uploaded copy of a known work" into an engineering task requiring a one-time investment at near-zero marginal cost. This point needs no demonstration by outsiders: the platforms themselves use the very same technology every day for content moderation, for advertising attribution, and for blocking rival products, so its technical feasibility is proven by their own daily operations. And on the right holders' side? The same work is cut into clips, speed-shifted, mirrored, and screen-recorded, scattered across millions of accounts—delete one copy and another is uploaded, while monitoring costs rise with the platform's scale. The side on which the lower cost of prevention sits has now completely flipped; yet the rules still assign the work according to the old 1998 cost table.
The infant industry—killed by the industry's maturity. The previous installment used a metaphor: swaddling clothes designed for a newborn, wrapped around a full-grown strongman. This installment adds one more thrust of the knife: the swaddling clothes not only wrap the strongman—they also crowd out the cradle. The content ecosystem that the incumbent giants built up at zero copyright cost is precisely the threshold that keeps compliant new entrants out of the market. To protect giants in the name of protecting infants protects not the existence of the industry, but the incumbents' cost structure.
The marginality of infringement—killed by economies of scale. Pirated broadcasts of marquee sporting events and clip-cut versions of hit dramas have demonstrable traffic-driving power and complete engagement data; they enter the training of the recommendation system and enter the monetization chain of advertising and tipping. Infringing content is no longer noise in the operation: it has entered the production function and become a raw material. To a routine raw material on the assembly line, the rules apply a piece-by-piece rework procedure designed for occasional defects—and the very shape of that procedure dictates that it can always keep pace with the occasional, and can never keep pace with scale.
All four premises are dead, and the justification for the rules hangs in midair, wholly unsupported. Yet the rules themselves remain without a scratch. How was this pulled off? Now it is time for the most ingenious link in the entire story to make its entrance.
V. The Most Ingenious Link: The Labeling Game
In the world of law there is a fact so plain that it is easily overlooked: what you are called determines what you are liable for.
Copyright law keeps two boxes ready for conduct. One is labeled "direct infringement"—personally carrying out the act of making a work available to the public; the other is labeled "indirect infringement"—merely providing the conditions for someone else's act. The safe harbor is erected entirely on top of the second box: once conduct passes through the "indirect" door, liability is automatically capped at "takedown upon notice."
This dichotomy presupposes a clean dividing line: the actor is the actor, and the helper is the helper. But algorithmic distribution happens to sit astride that very line—and it has snapped the line under its weight. The same conduct can be described in two vocabularies. In the technical vocabulary: the upload button is pressed by the user, and the file sits in the user's storage space—the platform is a mere warehouse, so into the "indirect" box it goes. In the functional vocabulary: to which users a work is presented, in what position, and with what weight are all decided by the platform's algorithm, and the money, too, is collected by the platform—is that not "making available" itself? It belongs in the "direct" box.
The academy has been quarreling over this for more than a decade—the "server test" against the "substantial substitution test"—and what the quarrel is about is precisely which vocabulary should be used.
But pay attention to the quiet mechanism behind the quarrel: however the quarrel ends, so long as the conduct is ultimately stuffed into the "indirect" box, the safe harbor descends over it as a whole and liability is automatically capped—while the act of "stuffing into the box" presents itself, from beginning to end, in the guise of a "finding of fact." It says that it is merely answering the question "whose act is this?", and never breathes a word about the fact that it has simultaneously answered "who bears no liability." The substantive judgment that allocates liability has been packaged as a technical exercise in fact-finding.
Set this against the Ming dynasty story and you will find a difference worth pondering. The Ming allocated burdens by means of status labels (gentry 士 / commoners 民), and at least it did so openly—the different treatment of gentry and commoner was written on the face of the system. Today we allocate liability by means of conduct labels (direct / indirect), but the labels wear the cloak of "objective description." Judges find facts in all sincerity, scholars parse concepts in all sincerity; not a single person is lying—yet the cap on liability is already complete the moment the label settles into place.
This is the complete recipe for the "disguise in which no one lies": the label makes the decision on the person's behalf, and the person making the decision believes he is merely affixing a label.
Legal methodology has a respectable formulation for this: every concept has premises; the premises die, yet the concept goes on allocating consequences by inertia—the category outlives its premises. The Ming "gentry" went on drawing their preferential exemptions long after the scholar-gentry had ceased to shoulder the work of moral edification and governance; today's "indirect" goes on holding the safe harbor over platforms long after platforms have ceased to be passive. The same disease.
VI. The Two Strongest Cards on the Other Side
At this point, readers who know this field will surely have been storing up objections. Scholarly integrity means crossing swords with the strongest versions of the opposing arguments, so let us take the two toughest cards in the other side's hand and lay them face-up on the table.
The first card: the chilling effect. "Make platforms filter proactively, and the platform's rational choice is to over-block rather than risk missing anything—and what gets wrongly deleted is lawful secondary creation, commentary, and parody. What you are after is copyright; what you kill is expression."
This card's framing is sound—this really is a trade-off between two types of error—but it books the error cost of the status quo at zero, and that is where it goes wrong. How tolerant are the current rules of the error of "infringement remaining online"? Close to one hundred percent: taken down, then re-uploaded, restored to its place at zero cost. The trade-off has never been properly framed as "filtering or freedom"; the right question is "if the errors on both sides are priced, where does total error cost bottom out?" Besides, the proposal can be narrowed: the obligation would attach only to identical re-uploads of works already identified by notice—entire sporting events pirated in full-length illicit streams, whole episodes of series sliced into clips. Machine recognition of this kind of content is simply not in the same order of difficulty as deciding whether a given parody amounts to fair use. As for the genuinely hard transformative-use judgments? An honest accounting: machines cannot make them, and machines should not be asked to make them—that ground is conceded exactly as it stands, and left to human beings and to courts. Surrender what machines cannot judge, and the case for what machines can judge becomes unanswerable.
The second card: technological neutrality. "The Sony case settled this long ago: so long as a technology has substantial lawful uses, its provider cannot be pursued merely because it can be turned to infringement—algorithms are general-purpose technology, and to touch them is to betray innovation."
This card is aimed at the wrong target. What Sony protected was technological design—the VCR's birthright as a product. Once Sony had sold the machine, it neither controlled what users recorded nor took a cut recording by recording. But every act of content distribution by a platform is a present business decision coupled with a present instance of profit-taking. The later Grokster case drew the line cleanly: lawful uses can shield product design; they cannot shield business conduct that induces infringement. In one sentence—technological neutrality can shield the knife-maker; it cannot shield the one who hands the knife to others and takes a commission keyed to the results.
VII. Even the Rule's Birthplace Has Issued a Diagnosis
If everything above still reads like one person's opinion, then consider what the inventors of the rule themselves have said.
In May 2020, the U.S. Copyright Office released a report on its years-long study of Section 512, and its conclusion employed a vivid turn of phrase: the balance Congress had sought in 1998 had gone "tilted askew"—tipped out of true—systematically skewed toward the platform side, with the burden on right holders growing ever heavier. The European Union went further: Article 17 of the 2019 Copyright Directive directly imposed on content-sharing platforms an obligation to "make best efforts to ensure that notified works do not reappear." Poland, objecting that the provision harmed freedom of expression, brought an action before the Court of Justice of the European Union; in 2022 the Court ruled—the provision is valid and, when paired with safeguards for users, compatible with fundamental rights.
The birthplace has owned up; the transplant site is still reciting an instruction manual from thirty years ago. The most valuable thing about this contrast is that it points the answer toward two perfectly plain mechanisms—explaining all of this requires no interest-based narrative whatsoever. Legal transplantation inherently carries a time lag: what we imported in 2006 was the text of the rule; we could not import the consciousness of the 1998 premises that lay behind the text. Doctrine has a natural inertia: textbooks, case teaching, and adjudicative reasoning replicate themselves around established concepts. For a concept to persist, no one needs to invest any interest in it; all it takes is that no one be willing to pay the intellectual cost of reconstructing the concept. The structural mechanisms already explain everything cleanly; any further conjecture added on top would be drawing legs on a snake—a superfluous addition.
The sincerity of the defenders and the failure of the rule are fully compatible. To prove the latter, there is no need to doubt the former.
VIII. What Is Outdated Is Not the Text but the Commentary
So where does the way out lie? Does the statute need amending?
No. This is the most surprising — and the most encouraging — part of the entire diagnosis: the text of Chinese law is not outdated in the slightest. What is outdated is the commentary — the 1998-style reading that interpreters have imposed on open-textured provisions.
Article 1197 of the Civil Code (民法典) says "knows or should have known," and Article 1195 says "necessary measures." The former is an authorization for the negligence inquiry; the latter, an authorization for the proportionality inquiry. Neither phrase is hard-wired to any particular era's cost schedule. Better still, the Supreme People's Court's 2012 judicial interpretation on the right of communication through information networks had long since planted the hook: engaging in the "selection, editing, modification, or recommendation" of works is a factor in assessing constructive knowledge; setting up rankings or creating indexes that recommend hit works can support a finding of constructive knowledge; and one who places advertisements against specific content and takes remuneration for it bears a "higher duty of care." These provisions were written in the portal era of human editors, yet their normative logic — the deeper the editorial intervention and the greater the direct profit, the higher the duty of care — dovetails seamlessly with the algorithmic age. What is algorithmic recommendation? It is "editing and recommending" at scale. What is traffic monetization? It is "directly obtaining economic benefit," refined. Scale and refinement do not change the nature of the conduct; they only magnify its weight.
Once the commentary is upgraded, the path emerges on its own: the intensity of "necessary measures" tracks the morphology of the infringement. For an occasional, isolated infringement, deletion suffices; where a work already subject to notice keeps flooding back, the measures should escalate from "deletion" to "blocking" — deletion cleans up the past, while blocking cuts off the future. The provision writes the two words side by side; it has been waiting for this day all along. Pair this with the yardstick discussed in the previous installment — punitive damages whose multiplier is fixed at the reciprocal of the probability of detection and enforcement — and the two ends, the establishment of liability and the scope of liability, are both joined.
Let me add just one Ming echo, to cool the correctionists' ardor: in the early-Qing zouxiao purge (奏销案) that rectified tax arrears in Jiangnan, the man who owed a single li and the man who owed a thousand shi were stripped of rank for the same offense, and contemporaries jeered that "a tanhua laureate — a third-place graduate of the palace examination — is not worth a single copper cash." And the result? Since owing more brought no added punishment, the rational actor simply defaulted to the hilt — once punishment breaks free of proportionality, deterrence falls to zero at the other end. A rule designed to correct a failure must itself speak law and economics; otherwise it becomes the next failure.
IX. A Final Word
Put the two essays together, and the picture is complete.
The previous essay answered the question "how does the equilibrium take shape": inverted imputation drove down the probability of detection and punishment, and zero pricing produced infringement at scale—that was a problem of arithmetic. This essay has answered the question "why does the equilibrium never break": the labeling game made that arithmetic disappear into the grammar of the law—that is a problem of structure. The solution to the arithmetic problem is to multiply the multiplier back in; the solution to the structural problem is to turn the annotation to today's page.
As for the concept that runs through both essays, let me state its full name one last time: structural interest disguise is using yesterday's proof of efficiency to underwrite today's distribution of benefits. The Ming dynasty began in the name of nurturing scholars and ended in the collapse of the tax base; the safe harbor began in the name of fostering an industry and continues on as a zero-priced equilibrium. The text remained neutral throughout, the justifications remained respectable throughout, the participants were by and large sincere—while the imputation remained persistently skewed.
That is why exposing such a disguise has never required finding a villain. It requires only the production of that expired proof of efficiency.
An equilibrium is not broken by eloquence. It can be broken only by three things: a freshly recalculated cost ledger, a newly recalibrated yardstick, and a company of interpreters willing to turn the annotation to today's page.
The first two have already been supplied, in the previous essay and in this one.
The third is the reason these two essays exist.
——
This essay is a sequel to "'Rational Infringement' in the Safe Harbor" (《避风港里的"理性侵权"》), developing the argument from the angle of institutional structure; the rules and sources discussed are not annotated here one by one. Readers who wish to pursue these questions further may consult the U.S. Copyright Office report Section 512 of Title 17 (2020), Article 17 of the EU Directive on Copyright in the Digital Single Market, and the judgment of the Court of Justice of the European Union in Case C-401/19 (2022). At the methodological level, the discussion of the openness of types in Larenz's Methodology of Law (《法学方法论》) bears on Part V of this essay. On the Chinese-law side, the basic normative authorities are Articles 1195 through 1197 of the Civil Code (民法典), Articles 9 through 12 of the Supreme People's Court Judicial Interpretation Fa Shi 〔2012〕 No. 20 (as amended in 2020), and Fa Shi 〔2026〕 No. 7.