—The Overlooked "Dark Figure of Infringement," and the True Function of Punitive Damages
Let me begin with an oddity from a real case.
A leading livestreaming platform presented a thoroughly respectable figure in court: for the infringing links reported by the right holder, its "disposal rate" was as high as over 96%. It sounded almost unimpeachable.
Yet another set of figures from the same case reads as follows: over one year and seven months, the right holder kept sending complaint letters, cumulatively reporting more than 3,600 infringing links, while the pirated livestreams on the platform kept streaming right up to the final day of that season—and their scale, rather than falling, grew. On the last match day, the right holder sent six more letters on that day alone.
How can a 96% disposal rate and the fact that "the pirated streams did not stop for a single day" both be true at the same time?
This was no occasional slip by one platform; it is a structural problem in platform-based online infringement. Explaining it clearly pulls into view an institution we have long misunderstood—punitive damages for intellectual property infringement.
The core of what I want to say comes down to a single sentence:
Punitive damages are, in essence, not a "moral punishment machine" but a "deterrence calibrator." Their multiplier is not there to express how angry we are at the infringer; it is there to calibrate "the price the infringer actually has to pay" back to "the harm he has actually caused."
In what follows, I will start from that oddity and work my way, step by step, to the newly effective judicial interpretation—Fa Shi [2026] No. 7 (法释〔2026〕7号).
I. "Whack-a-Mole": Why Notice-and-Takedown Fails
First, let me explain the trick behind that 96%.
The "action rate" the platform cites has as its denominator "the links reported to it." Out of 100 links it is told about, it deals with 96 — which sounds responsible. But the problem is this: the infringement actually occurring on the platform goes far beyond what it has been told about.
Live sports streaming is an especially typical case. A match runs 90 minutes; pirate livestream rooms open, close, and open again; the moment a streamer's account is banned, he comes back under a highly similar name — in industry slang, a "sockpuppet account" (马甲号, literally "vest account"). In one case, a single "football livestream" series ran a dozen or more nearly identically named accounts, which took turns pirate-streaming over forty matches.
Against an adversary of this kind, the unit on which "notice-and-takedown" operates is "one link" or "one livestream," not "one person." Delete one link, and nothing stops him from streaming on under a new one. The granularity of the rule (aimed at links) is badly mismatched with the degree of organization of the infringement (aimed at actors operating at scale). The result: the more you delete, the more slips through.
More fatal still is the time-sensitivity of livestreaming. A match lasts only two hours; once deletion comes later than the end of the match, the harm has already occurred. By that platform's own admission, the proportion of cases in which it "stopped the infringement in real time within the duration of the match" was under 2%. In other words, the vast majority of that so-called 96% "action" consists of going back to "deal with" the user long after the match has finished — which, so far as actually stopping the infringement is concerned, is close to meaningless.
That is why rights holders say what they say: the notice-and-takedown rule has become a time-consuming, labor-draining game of "whack-a-mole."
Viewed through law and economics, what lies behind this is a mismatch over "who shoulders the cost." The platform holds the moderation technology, the user data, the recommendation algorithms, and the enforcement backend; it is the "least-cost avoider" when it comes to blocking infringement. The rights holder, by contrast, must discover each infringement, gather evidence on each, and send a letter for each — at extremely high cost, rising linearly with the scale of the infringement. Yet "notice-and-takedown" shifts the entire job of identifying and locating infringement onto the rights holder — the party for whom it costs the most — while letting the platform lie back in the immunity of "I never received notice."
This is the first failure.
II. Why “Making Whole” Cannot Cure Pirate Streaming on Platforms
The second failure is more insidious. It hides inside the principle we know best — damages are awarded on the principle of making good the actual loss.
The make-whole principle carries an implicit premise: infringement will sooner or later be discovered and pursued, so "make good whatever loss was caused" is enough to restore the status quo ante.
But in platform infringement, that premise simply does not hold. The overwhelming majority of infringements never even make it into court.
We can see this with the simplest possible arithmetic. Suppose:
- a single act of infringement causes a loss of L;
- the probability that this infringement will ultimately be discovered and result in a damages award is p (say, one in ten — p = 0.1).
What, then, is the expected damages the infringer faces? It is p × L. Whenever p is less than 1, p × L is less than L — the expected price of infringing is lower than the harm it causes, and lower still than the money the infringer earns from it.
What does that mean? To an infringer who can do the arithmetic, infringement is a business that turns a guaranteed profit. At that point, make-whole damages not only fail to "make whole" — they effectively encourage infringement, because they hand infringement a discount.
Pirate streaming on platforms is precisely the kind of setting in which p is systematically low: content is vast in volume, livestreams are fleeting, users are anonymous, and evidence is expensive to gather. Later, I will use a concrete method to show you that, in that case, the infringements the right holder was able to document may have amounted to only about 13% of the infringements that actually occurred.
When p sinks this low, the collapse of make-whole damages is not a failure of enforcement — it is a failure of the institution itself in this setting.
III. So What Exactly Do Punitive Damages "Punish"?
Let us now return to that misunderstood institution.
Many people—including quite a few judges—instinctively understand punitive damages as a form of moral retribution: because you acted "intentionally and with serious circumstances," you are ordered to pay a "multiple" of the loss, by way of punishment. In this understanding, the multiplier is the gauge of culpability: the worse the conduct, the higher the multiplier.
That understanding points in the wrong direction.
Try a different vantage point. Remember the equation above? To make the expected cost facing the tortfeasor once again equal to the harm he has caused, the optimal award should be:
Damages = Loss ÷ Probability of Being Held Liable, i.e., D = L / p.
This is a classic result in law and economics (the work of scholars such as Becker, Polinsky, and Shavell). What it says is this: if only one in ten torts will ever be pursued to liability (p = 0.1), then in the one case that does get caught, the award must reach ten times the loss, so that "the sum of the expected damages across all torts" equals "the sum of the harm caused by all torts."
Do you see it now?
The "punitive" multiplier in punitive damages is, functionally, not an extra punishment at all; it is a compensatory backfilling of "the losses in all the cases that were never caught."
The multiplier is not the product of outrage; it is the product of 1/p, a calibration coefficient. When p = 1 (every tort is pursued to liability), 1/p = 1, and no multiplier is needed at all—full compensation suffices. It is precisely because p is less than 1 that a multiplier is needed. The size of the multiplier is determined, at bottom, by "how low the probability of liability is," not by "how wicked the tortfeasor is."
This is what I mean by a "calibrator." What it calibrates is the expected cost facing the tortfeasor, bringing it back to the true social cost of his conduct.
(Two clarifications, to forestall the quibbles: First, the calibration account does not deny the threshold of "intent"—only a deliberate tortfeasor who does the arithmetic can be steered by incentives; deterrence has no purchase on negligent torts. Second, the calibration account does not claim "the more damages, the better"—awards far exceeding L/p would produce overdeterrence, which is why the law sets a fivefold cap. It is a theory with boundaries, not a slogan.)
IV. We Used a "Fish-Counting" Method to Estimate the Platform's True Scale of Pirate Streaming
The calibration thesis sounds plausible, but it has one fatal practical problem: p is simply invisible. How much infringement actually occurs on the platform, how much of it is discovered, and how much of it is pursued? No one can count it. Judges can only go by feel and pluck a number at whim somewhere "between one and five times." The arbitrariness of the multiplier is rooted precisely in the fact that p is unknowable.
So can p actually be calculated?
Yes, it can. We borrowed an old technique that ecologists use to count fish and wild animal populations — the mark–recapture method.
The principle can be explained in a single sentence: To find out how many fish are in a pond, you don't have to catch them all. You cast the net twice, see what proportion of the two hauls consists of "repeat fish," and you can work backward to the population of the entire pond. The fewer fish are caught both times, the bigger the pond and the more fish have slipped through the net.
Criminologists use the same method to estimate the "dark figure of crime" (the true volume of crime that is never recorded), and census bureaus use it to correct for undercounts. The "dark figure" of platform infringement is structurally the very same thing.
That case happened to provide a natural "two casts of the net":
The rightsholder's anti-infringement work was carried out by two teams. One team handled notarized evidence collection — they had to click into each livestream room, record the screen, and then back out to the directory: slow, with few captures per unit of time. The other team handled sending notice letters — mainly refreshing the livestream directory and grabbing links: fast, with many captures. On roughly the same match days, the two teams swept the same "popularity-first" livestream list.
The results were as follows:
- The evidence-collection team caught 2280 infringing livestream rooms;
- The notice team caught 3621;
- The overlap between the two teams was only 467.
Plugging these three numbers into the mark–recapture formula (N = evidence captures × notice captures ÷ overlap):
N = 2280 × 3621 ÷ 467 ≈ 17,700
In other words, the pirate livestream rooms that actually occurred on the platform that season numbered roughly 17,700. Yet what the rightsholder painstakingly captured as evidence — the figure used to calculate damages — was only 2280.
Capture rate = 2280 ÷ 17,700 ≈ 13%.
Put differently: the number used to calculate damages may have covered only slightly more than one-tenth of the actual infringement. The remaining nearly ninety percent never surfaced.
Is this figure solid? I must candidly lay out two points (which also serves to muzzle the opposing side in advance):
First, there is a lower bound that requires almost no assumptions. The distinct livestream rooms directly observed by the two teams, pooled together, add up to 5434 (2280+3621−467). The true number cannot be smaller than the number actually observed, so it is at least 5434. Even if you completely reject the 17,700 estimate, the 2280 rooms used to calculate damages still account for only 42% of the infringement that was already visible.
Second, the estimate is more likely too low than too high. Because both teams fished from the same "popularity-first" list and both favored the popular matches, this "shared preference" statistically inflates the overlap and depresses the estimate — which means the true figure is very likely higher than 17,700. If the opposing side raises the objection that "the two teams were not independent," that objection itself helps us make the argument that "there was more infringement than 17,700."
By this point, the 96% trick has been thoroughly exposed:
The platform says it handled 96% — but that is the portion it was told about. The notified portion plus the evidence-captured portion, taken together, still amount to only about thirty percent of the actual infringement. Even if it handled one hundred percent of the notified portion, roughly seventy percent of the actual infringement was never touched at all, from start to finish.
The 96% figure is genuine, but it is calculated against a denominator that was quietly shrunk. This is the "denominator trick."
V. The New Rules Are Already Applying the Logic of "Calibration"
At this point, we must discuss the rules that have just taken effect.
Fa Shi [2026] No. 7 (法释〔2026〕7号)—the Interpretation of the Supreme People's Court on the Application of Punitive Damages in the Trial of Civil Cases Involving Infringement of Intellectual Property Rights (《最高人民法院关于审理侵害知识产权民事纠纷案件适用惩罚性赔偿的解释》)—was issued in April 2026 and took effect on May 1. It comprehensively revises the old 2021 interpretation. One provision in particular virtually writes into the text the logic of "calibration, not punishment."
That is Article 13:
Where an administrative fine or a criminal fine has already been imposed for the same infringing act and has been fully executed, the people's court shall take that into account when determining the multiplier for punitive damages.
Consider the logic of this provision. If punitive damages were moral retribution, then public-law fines and private-law punitive damages would address different relationships and different values, and ought properly to be calculated separately and stacked on top of each other—culpability does not lessen merely because an administrative penalty has already been suffered. Yet the new rules specifically require that, in calculating the civil multiplier, fines already executed be taken into account (that is, offset against one another).
Only one explanation makes sense of this: what is being calibrated is the "total quantum of deterrence" facing the infringer (public-law sanctions plus private-law damages), and the goal is to make the total cost commensurate with the social cost of the infringement ("the punishment fits the offense" (过罚相当)), rather than to heap repeated moral condemnation on the same culpability.
This amounts to the Supreme People's Court expressly acknowledging, in the text of the institution itself, the orientation of "calibration, not retribution."
Several other provisions of the new rules likewise pave the way for "calibration":
- Article 6 itemizes the circumstances constituting "intent," the first of which is "continuing to infringe after receiving an effective notice"—a provision seemingly tailor-made for platform-style "knowing yet deliberately letting it continue";
- Article 7 lists "making a business of infringing intellectual property rights," "enormous scale," and "severe damage to the right holder's market share" as serious circumstances—aimed squarely at professional pirate broadcasters and platforms' scaled-up profits;
- Articles 8 through 10 straighten out the base amount: it may be calculated according to actual losses, unlawful gains, or profits from infringement, and where all of these are difficult to calculate, by reference to license fees; in particular, Article 10 provides that where the defendant, without justifiable reason, refuses to provide its back-end account-book data, the base may be determined directly according to the plaintiff's claim—a sharp instrument for piercing the black box of platform data;
- Article 11 makes clear that the multiplier is determined comprehensively on the basis of subjective fault and the circumstances, and that it "need not be a whole number";
- Article 12: total compensation is capped at five times the base.
Note Article 11's phrase "need not be a whole number." It means courts no longer have to round crudely among "one times," "two times," or "three times," but may set a precise calibration multiplier based on the estimated probability of detection and prosecution—say, 3.3 times.
Now connect this with the earlier "fish-counting" result: the capture rate is roughly 13%–31%, and "being discovered" still does not equal "being held accountable" (the real-time blocking rate is under 2%), so the true p is even lower than the capture rate. On that basis, the calibration multiplier 1/p needed to restore deterrence must be at least 3.3 times, approaching 7.7 times. That is a floor, not a ceiling.
Article 11's tool of "need not be a whole number," together with the p derived from "counting fish," for the first time gives "what the multiplier should be" a calculable, verifiable answer—no longer a matter of a judge's momentary whim.
Sidebar: The Two Questions That Most Often Get "Derailed"
Now that we have come this far, there are two questions that will almost certainly be asked, so let me get them clear first.
Question One: Can the platform's advertising revenue and membership fees be used directly as the base for damages?
No—because what we have here are really two different markets.
The rights holder (the copyright owner) operates in the licensing market—its revenue comes from selling the broadcasting rights to sporting events to platforms; it does not broadcast the events itself, and it was never going to earn advertising or membership fees in the first place. The platform operates in the broadcasting market—it monetizes content through advertising, memberships, and viewer tipping. Through infringement, the platform obtained the position of a "de facto licensee" without ever paying the license fee—and it is precisely this unpaid license fee that constitutes the rights holder's loss. To take the platform's advertising revenue from the broadcasting market and back-calculate the rights holder's loss from it is to jumble the two markets together. That is why the base in cases of this kind is usually "the license fee (times a multiple)"—which corresponds exactly to the ordering in Article 8 of the New Rules, under which "where the amount is difficult to calculate, reference may be made to the license fee."
(Of course, the two markets are not completely sealed off from each other: the platform's pirated streams substitute for users' viewing of the legitimate feed, depress platforms' willingness to pay rights holders, and scare off potential buyers—and in this way the harm is transmitted back from the broadcasting market into the licensing market. But that is a topic for another article.)
Question Two: Will a court accept a number "estimated" in this way?
This is the most down-to-earth question of all, and the answer is: used in the right place, it stands up.
The key is to keep the "base" and the "multiple" distinct. We have never proposed taking the estimate of 17,700 and using it as the damages base—the base should still be worked out the honest way, from the verified number of matches and the license fee. The estimate is put to work in calibrating the multiple and in establishing the circumstances and the infringer's intent. And the multiple, to begin with, is a matter of the judge's discretion within a statutory range; Article 11 of the New Rules even states expressly that it "need not be a whole number." To take a mathematically grounded probability of detection and enforcement as a point of reference when fixing the multiple—surely that is more judicious than plucking "double" or "triple" out of thin air?
What is more, the administrative and criminal spheres have long had precedents for using sampling and extrapolation to establish the scale of unlawful conduct, and reasonable estimation in civil damages is likewise permitted by law. If ecology can use this method to count fish, and criminology can use it to estimate the dark figure of crime, then the dark figure of platform infringement is structurally the same problem—it has simply changed into legal dress; the statistical core is unchanged.
VI. Where the Problem Lies: The Multiplier Has Been Treated as a "Moral Outrage Meter"
The Supreme People's Court's legislation is proactive and on point. The problem lies in adjudicatory implementation.
In practice, one can clearly observe that when applying punitive damages to platform defendants, lower courts generally set the multiplier low — either picking the low end of the one-to-five-times range, or applying a steep discount on the grounds of the platform's "technological neutrality" or the "massive volume of content that cannot be dealt with in real time," or even sidestepping punitive damages altogether and falling back on statutory damages.
Why does this happen? Because judges, in their subconscious, treat the multiplier as a "moral outrage meter": the platform is not, after all, the one directly pirating the broadcast; its subjective culpability falls short of the pirate broadcaster's; and, well, "objectively speaking, a massive volume of content really is hard to handle item by item" — so the punishment it "deserves" is lighter, and the multiplier lower.
But switch to the "calibrator" perspective, and the conclusion is exactly the opposite.
Platform-type infringement calls for a higher, not lower, multiplier — not because platforms are worse, but because: in the platform setting, the probability of detection and prosecution p is systematically lower (in this case, roughly 13%–31%), so the calibration coefficient 1/p is inherently larger; at the same time, the platform, as the least-cost avoider, adopts a strategy of "hands-off tolerance plus token enforcement," which is itself a rational choice in a low-p environment — unless a sufficiently high multiplier changes its cost-benefit structure, it has no incentive whatsoever to govern in earnest.
A simplified ledger makes the point: suppose the harm in a single case of a given type of infringement is L, and the probability of being held liable is p=0.2; restoring deterrence requires an award of about 5L. If the court awards only one times (L) or two times (2L), the expected damages facing the infringer are only 0.2L and 0.4L — a mere twenty percent and forty percent of the harm he caused. At that expected price, any infringement whose return exceeds twenty to forty percent of the harm remains worthwhile. A low multiplier not only fails to deter; it amounts to putting infringement on sale at a discount with a "clearly marked price."
Courts' excessive indulgence of platforms' "technological neutrality" and "massive content" narratives amounts, in substance, to once again pushing the governance burden that should be borne by the least-cost avoider back onto the right holder, the party whose costs are highest — which runs exactly counter to the original purpose of the safe harbor rules.
A word in passing: in that case, the platform was anything but "neutral" — it had placed advertisements on search engines along the lines of "Free HD Premier League live stream — download a certain app to watch for free," actively using pirated content to funnel traffic to itself. Running infringing content as a traffic business on the one hand, while telling the court that it is "merely a neutral technological storage space" on the other — these two faces should not both be credited.
VII. A One-Sentence Close
To compress this entire article into a single sentence:
Punitive damages are not a moral-outrage meter; they are a deterrence calibrator. The multiplier = 1 ÷ the probability of being held liable. In a setting like platform piracy—where for every one infringer caught, nine slip through—a low multiplier is not "punishment too light"; it is "no calibration at all." It never reaches the tipping point at which a platform flips from "inaction pays" to "governance pays."
The Supreme People's Court has already written the logic of "calibration" into the new rules (most notably the offset in Article 13 and the non-integer multipliers in Article 11). What comes next is whether adjudication can keep pace.
But to turn "calibration" from a slogan into a number in a judgment, what we have actually been missing is something long thought to be "incalculable"—the probability p of being held liable.
The point of this article is: it can be calculated. With a net cast twice, and with an old method of counting fish, we can estimate just how large the iceberg of infringement below the surface really is.
Once p can be estimated, the punitive damages multiplier ceases to be a matter of a judge's momentary inclination and becomes instead an empirically grounded result of calibration.
This step is worth taking.
(This article is offered for public legal education and scholarly discussion; the cases involved have been anonymized. As to the facts and the application of law in any specific case, the effective judgment shall control.)