Author: Nebula Walker Date: 06AUG2026 MYTHOGEN ENGINE (mythogenengine.com)
Approved, Then Unseen
I. An Afternoon That Left No Trace
The post was submitted for review. It was approved. The admin even liked it.
You scroll into the group. The post is there — formatting intact, timestamp normal. Three hours later, the reach counter stalls at twenty-four. The page status says no issues. The account has no restrictions. The violation record reads zero.
So you arrive at the only possible conclusion: you didn't write well enough.
Until you borrow someone else's account and log in. They scroll through the entire group. Your post is nowhere to be found.
Up to that moment, the only "evidence" you had was the absence of response. And absence of response, in any appeals process, does not constitute evidence — it doesn't even constitute a problem. Because the system's default explanation is always ready, always reasonable, and always points back at you.
What I want to discuss is not how demoralising this is. What I want to discuss is this: this state of affairs is not an accident, nor a side-effect of the algorithm. It is an invention — with a number, an applicant, and an approval date.
II. Only a Matter of Degree
First, let's dismantle a false premise.
Every argument about "whether or not censorship occurred" assumes that censorship is a discrete event: blocked or not, deleted or not, yes or no. People argue about whether the switch was flipped.
But the platform doesn't hold a switch. It holds a dial — a continuous one.
Blocking an account is turning the dial to zero. Rejecting a post is turning it to ten percent. Approving but not distributing is turning it to twenty percent. Doing nothing and letting the ranking system handle it is turning it to some number that nobody needs to be accountable for.
The effects belong to the same category. They differ only in degree.
And this is precisely the advantage of a continuous dial: no event means no evidence.
A discrete action produces a timestamp, a notification, an identifiable subject. Continuous attenuation produces nothing. Same outcome — one requires accountability, the other doesn't.
So the more precise term isn't censorship. It's attenuation. This is not a loophole. It's a chosen design.
Deletion produces three things: a point in time, a notification that can be screenshotted, and a narrative that can be told.
Demotion produces none.
No point in time, so you don't know when it began. No notification, so there's no entity to appeal to. No narrative, so you can't explain to anyone what happened — all you can say is "not many people seem to have seen it," and that sounds like a complaint, not an accusation.
A person who's been deleted can say, "I was deleted." That sentence has a subject, a verb, and an object. It can be verified; it can be refuted. A person who's been demoted can't even form a complete sentence.
So deletion is more merciful. At least it acknowledges you existed.
III. Patent 10356024
Filed in February 2015. Granted by the United States Patent and Trademark Office in July 2019. Title: "Moderating content in an online forum." Applicant: Facebook.
In plain language, the abstract says this: the system can receive a list of prohibited content and block it by reducing the distribution of those comments to other viewers, while still displaying the blocked content to the user who posted it, so that the user will not realise they have been blocked.
Three actions, composing a single mechanism:
- Reduce distribution to others
- Maintain normal display to the author
- Stated purpose: to prevent the author from knowing
The first two are engineering. The third is the patent's actual claimed invention — not "how to restrict distribution," but how to make the restricted person feel as though they are not restricted. This statement was written into the application as a technical feature and was granted.
The patent also extends this capability to "forum creators or moderators," and it can be executed automatically by text sentiment analysis or machine-learning classifiers. In other words, from its inception, this was never just a platform-level anti-spam tool. It is a power designed to be delegated to community administrators.
To be clear: a granted patent does not mean deployment. Nor does it mean that every instance of silence you encounter traces back to this code. I will not make that accusation. But a patent records intent and method — ten years ago, a group of engineers sat down and defined "making the user unaware" as a technical problem worth filing and worth protecting. That fact alone says a great deal.
This logic was not confined to the distribution layer. When the same principle — no announcement, no record, no notification — was carried to the criteria layer — deciding what counts as good content, what counts as machine-written — the detector ended up manufacturing the very thing it claimed to detect. That is the subject of Volume II of this series (The Least AI-Like Article Was Actually Written by AI — A Real-Time Experiment in Cognitive Blind Spots).
And the real intersection of the two layers is this: they operate independently, yet they penalise the same behaviour. Careful revision produces clean, consistent text — which is precisely what the detector flags as machine-generated. Careful revision also produces long-form writing, outbound links, and low emotional arousal — which are precisely the features that the content distribution guidelines explicitly list for demotion. Two machines that have never met, yet they always land on the same person.
There is also a question of timing worth noting separately.
This patent was filed in 2015 and granted in 2019. Generative AI entered the public consciousness from late 2022 onward.
That is to say, "reducing distribution while keeping the author unaware" predates the technology now being scapegoated by at least seven years. It is not a model error. It is not training-data bias. It is not anything that can be explained away as "the technology isn't mature yet." It was conceived, documented, and paid to be protected before deep learning was widely deployed for content ranking.
Attributing it to algorithmic misjudgement is a very convenient mistake. Because algorithms have no motive, and mistakes can be fixed — as long as the problem is framed as a technical issue, it automatically becomes something that will eventually be resolved, and no one needs to be held responsible.
But a patent is not a misjudgement. A patent is something that someone sat down, thought through, wrote into a document, and paid to protect. That is a decision, and the nature of the decision is commercial.
In this matter, the algorithm is the executor, not the decision-maker. Putting the executor on trial lets the real defendant carry on with their meetings.
IV. The More Common Version: No One Needs to Press a Button
The truly everyday form is far milder than the patent, and far harder to address.
Meta's own transparency pages explain that each group's individual News Feed is an AI-based ranking system: it scores each post, ranks them by score, and factors in your typical interaction patterns with that group. The same page also states that group administrators can set a default sorting method that applies to all members and visitors.
This creates a perfectly legal, perfectly traceless operation: approve your post, even like it, then do nothing.
Approval only makes the post exist in the database. Existence does not equal distribution.
If the administrator runs courses, cheat sheets, or tool lists, they don't need to suppress anyone. They only need to keep the default sort on "Most Relevant" and ensure that their inner circle's content gets early engagement — pinning, cross-promotion, opening comments — and outsiders will sink to the bottom automatically. Every single action, viewed in isolation, is legitimate community management.
Stacked together, the effect is identical to the patent. The only difference: no one did anything wrong.
This is the most refined aspect of the system. It doesn't need a villain.
And therefore, questioning intent is pointless. I cannot prove any administrator's motive, nor do I need to — because in this structure, the outcome does not depend on intent. Self-interested management behaviour and deliberate exclusion output the same News Feed.
V. The Third Layer: Dissent Appears as "Non-Existence"
Everything discussed so far has been about "voices being turned down." But that is only the first-order effect.
A group administrator effectively controls three things:
- Who gets to speak — review, approval, delay
- Which content rises to the top — default sorting, pinning, early engagement
- Which opinion appears to be everyone's opinion
The third layer requires no additional action. It is the automatic product of the first two layers combined — and it is the real output.
Because after suppressing ten dissenters, what remains is not "ninety supporters plus ten silent people." It is a room that looks entirely unanimous. Every person in that room calibrates their own judgement against "no one disagrees."
Dissent does not appear as "dissent." It appears as "non-existence."
And non-existence reads as agreement. You cannot be suspicious of an empty chair.
So the mechanism's output is not silencing. It is consensus. And consensus is something that can be manufactured, traded, and is extremely valuable.
VI. This Is Why Scams Move Faster
The investment fraud methods published by Taiwan's Criminal Investigation Bureau and local police departments follow a fixed structure: social media ads attract users to join a group; group members play multiple roles; "plants" post profit screenshots to create an atmosphere where everyone is winning; early small withdrawals build trust; after larger investments, withdrawal is refused on grounds of tax or bond payments; anyone who starts to question is immediately kicked out.
Map each of these actions back to the three layers of control from the previous section. They match perfectly.
The difference is not in the technology. It's in the disclosure.
What a fraud ring needs — a trust machine — is simply a controlled group: controlling who speaks, controlling what rises, controlling what looks like consensus. They don't need to build this machine themselves — the platform ships it as a standard feature to every moderator.
So a course-selling administrator and a fraud-running syndicate use the same default settings. The former sells anxiety relief; the latter sells the evaporation of capital. Same architecture, different price tier.
This also explains something that puzzled me for a long time: why structural analysis fails to cool the frenzy inside these groups. Because those people aren't information-deprived — the room they're in has already been calibrated for unanimity. In a space where all opposing views appear as "non-existence," an article saying "that's not how it works" doesn't read as a correction. It reads as noise.
VII. What Meta Itself Wrote
The EU Digital Services Act requires platforms to submit transparency reports. Meta's version includes a note, stated very clearly: demoted content does not necessarily violate Community Standards; therefore, demotion actions are classified under "content distribution guidelines," not under violation records.
This single sentence explains the seemingly contradictory picture:
- Page status: no issues
- Account restrictions: none
- Violation record: zero
- Average reach: twenty-four
There's no contradiction. Because demotion doesn't go through the violation system at all. It writes to no panel you can see, so your account stays permanently clean while your voice never arrives.
You thought you were reading a health report. You were actually reading a health report that doesn't include that particular test.
The Digital Services Act, Article 17, stipulates: for any visibility restriction — and the text explicitly includes demotion — the platform must provide the affected user with a clear and specific statement of reasons.
Same company, same system, same algorithm. In the EU, they must inform you. In Taiwan, in Hong Kong — they don't have to.
That sentence is worth pausing on.
It's not that they can't. It's that they don't have to.
Every claim of "we can't distinguish," "the scale is too large," "it's technically difficult" is instantly invalidated by an enforceable regulation. Technical feasibility was never the issue. It never was.
VIII. The Company That Can Demote with Precision Claims It Can't Identify Content Farms
Place this back in a commercial context to see the full picture.
In November 2025, Reuters reported, based on a batch of internal documents: Meta internally estimated that approximately ten percent of its 2024 revenue — roughly USD 16 billion — came from scam and contraband advertising; the platform was pushing approximately 15 billion high-risk scam ads daily, annualising to about USD 7 billion in revenue.
The enforcement threshold was designed as follows: the automated system required 95% confidence to block a suspected scam advertiser; below that threshold, the suspected party was not blocked but charged a higher advertising fee. In the first half of 2025, the ad integrity team was instructed that no enforcement action could cut more than 0.15% of projected revenue — a maximum of USD 135 million out of USD 90 billion. A separate internal document estimated the regulatory fine ceiling at approximately USD 1 billion, noting it was far lower than scam-ad revenue.
Meta subsequently responded that the ten-percent revenue figure was an internal one-off rough estimate using overly broad criteria and should not be treated as an actual figure. This rebuttal should be included, because it doesn't actually undermine the argument — it only undermines the total amount.
The two truly fatal items are not estimates. They are directives: the 95% blocking threshold, and the enforcement cap at 0.15% of revenue. Estimates can be debated on methodology; directives have no methodology to debate.
And the same company's content distribution guidelines list, in black and white, the types of content subject to demotion: content with limited originality that mainly repackages from other sources; clickbait and engagement bait; links to ad-heavy, slow-loading websites.
That description is the definition of a content farm. The rule exists but doesn't trigger.
So we get a clean comparison:
- For a long-form article: demotion is precise, instant, requires no reason, no notification, no record.
- For a scam ad: requires 95% confidence, and enforcement has a budget cap.
A company capable of the former, claiming it cannot do the latter — that statement is logically untenable.
It's not a capability problem. It's an allocation problem. The capacity for precision demotion has always existed. It just isn't applied to the paying side.
IX. The Shelf-Fee Analogy, and Why It Doesn't Hold
At this point, the most common defence will appear: isn't this just shelf fees in a supermarket?
Good placement costs money. Eye-level shelves cost more than the bottom row. Endcaps cost more than mid-aisle facings. Those who can pay get the front; those who can't get the corner. This has been normal commercial practice for decades, and no one thinks it calls for outrage.
This analogy deserves serious engagement, because it isn't frivolous — retailers do sell visibility, and platforms do too. But there are three differences, and every one of them lands in the same place.
First, shelves are visible.
A supplier walks into the store and immediately knows they've been placed on the bottom shelf. They don't need to borrow someone else's membership card, control for variables, or compare screenshots from two devices. They look down and they know. Demotion is visible to the naked eye, occurring in broad daylight.
Second, there is a price list and a receipt.
Shelf fees are negotiated. There is a quote, a contract, and a clear exchange: you pay this amount, you get this placement. Afterward, both parties know what happened, because it was a documented transaction.
On the platform side, no such document exists. No one quotes you "pay this and your post will be visible to group members." When you're moved to a lower tier, no document is generated. This isn't a cheaper version — this is a version where the transaction doesn't even exist.
Third, the rejected party knows they were rejected.
A supplier who fails to secure shelf space will know. They can count, they can negotiate, they can file a complaint, they can try a different chain, they can tell others about it — because they have an event they can narrate.
And this is exactly what the preceding sections have repeatedly described: a demoted person does not have this awareness. They don't even know which group they belong to.
There is also a commonly overlooked difference on the buyer's side.
The customer can see the shelf too. They know the bottom row exists; if they want to find something, they squat down and they can. The entire store's inventory is in front of them — the only difference is ease of access.
A News Feed has no bottom shelf. You can't squat down, because that space simply doesn't exist — undistributed content, to the reader, is not hard to reach. It's non-existent.
So the real difference isn't whether you charge. It's whether visibility itself is visible.
A supermarket sells good placement, and all placements are visible. A platform sells visibility, and the allocation of visibility is itself invisible.
One is a market where fairness can be debated. The other doesn't even generate the information required for that debate.
X. The Cruellest Part Is Not Silencing — It's Exploitation
What truly stings is not having your article deleted. It's having someone praise your content while knowing that most people will never see it.
In this structure, an administrator doesn't need to refute you, block you, or even dislike you. They only need to let you keep writing.
And by continuing to write, you provide more than articles. You provide content density for the group, activity metrics for the platform, training signals for the recommendation system, indexing material for search engines, and the brand value of "this group has many quality creators." Only your reach — your influence — does not increase.
This produces a peculiar state: the system still needs you to create. It just doesn't need anyone to see you.
Here I must again hold the line: I cannot claim that anyone knowingly exploited you while aware of your suppression. I don't have that evidence, and this article will not pretend to.
But even without that evidence, this remains true — because the structure makes the following scenario possible: even if an administrator genuinely admires your content and even likes your posts, as long as they don't actively change the distribution settings, you will remain a creator who contributes value long-term yet remains invisible long-term.
Asymmetry doesn't need malice to sustain itself. It only needs no one to make a change.
And once this asymmetry stabilises, the relationship among the group operator, the platform, and the creator becomes one of unilateral benefit — not because anyone decided to exploit anyone, but because changing it offers no advantage to the only party with the power to do so.
XI. Two Things I Cannot Claim
At this point, I must apply the brakes myself. An article about the burden of proof cannot be supported by accusations that cannot be proven.
First, I cannot claim any specific person did anything. All the mechanisms above have public documentation: the patent, transparency pages, statutory provisions, transparency reports. But who pressed what in which group — I have no evidence, and this article will not pretend to. The strength of structural analysis lies precisely in not needing to name names — when a system makes self-interest and malice produce the same outcome, asking which one occurred is itself a designed detour.
Second, I cannot claim that "if distribution were fair, things would be different." There are actually two independent constraints here, and conflating them leads to a false conclusion.
One is distribution attenuation — real, verifiable, and the platform's responsibility.
The other is demand mismatch: an article asking readers to admit they can't judge and an ad promising readers they don't need to judge are simply not competing in the same market. The former sells cognitive cost; the latter sells cognitive exemption. Even if the channel were perfectly fair, the latter would still win that room.
Separating the two explains why "extremely high completion rate" and "extremely low reach" can coexist — because they aren't even the same cohort.
XII. You're Not the Only One Who Won't Be Counted
There's one more thing I realised very late.
I learned that "many people have already stopped using Facebook" not from any data, but from someone mentioning it in an offline class.
This wasn't luck. It's structure.
No one posts "I'm quitting." Even if they did, the algorithm wouldn't surface it for you. And an account that has stopped posting appears exactly the same as one that never existed, in your News Feed. The system you inhabit is physically incapable of telling you how many people have left.
From this, a rule can be derived: information about platform attrition can only propagate through non-platform channels. Those who remain, no matter how they count, can never count those who have already gone.
This point is two sides of the same coin as all the mechanisms above. The system doesn't just make your voice invisible. It also makes others' departure invisible. You're left in a room that still looks busy — and the busyness itself is the final layer of manufactured consensus.
And the demoted person's situation is exactly the same.
Deletion can be counted. Each deletion has a subject, a timestamp, a notification — so it can be tallied, aggregated, written into reports, and eventually compressed into a number. And numbers create pressure.
Demotion has none of these. No one receives a notification, so no one knows which group they belong to. And a group whose members don't know they're in it will never organise, never speak out, and never become an issue.
Victims who cannot be counted are, institutionally, non-existent.
Not denied — simply ineligible for counting. Counting requires a list, and this mechanism's design purpose is precisely to produce no list.
No list means no group. No group means no demands. No demands means no policy pressure. The chain breaks at the very first link.
This is also why compressing a society's public discourse today no longer requires any censorship apparatus. All it takes is ensuring that the affected people cannot recognise each other.
XIII. If You Want to Self-Verify
A few practical words for those who still want to document what's happening.
The second account must already be a group member, using a different device and network, and must not be friends with you — the friend relationship independently boosts ranking. Open both simultaneously. Switch both to "Newest Posts" (chronological order) to eliminate ranking as a variable. If chronological shows the post but "Most Relevant" doesn't, the restriction is at the ranking layer. If neither shows it, the restriction is at the distribution layer. Screenshots must include the clock and scroll position; timestamps on both sides must match. Simultaneously post a control — short, no outbound links, non-controversial — and compare performance over the same period.
After all this, you'll have a record that can convince yourself.
Then you must accept one thing: you can prove "others can't see it," but you cannot prove "who made them unable to see it."
XIV. The Inversion of the Burden of Proof
And this is where the real anger should be directed.
Not at being suppressed. Suppression can be identified, resisted, and written about. The real design is making suppression and indifference evidentially indistinguishable.
In this design, the burden of proof is thoroughly inverted. The platform never claims your content is problematic, so it never needs to prove anything. But you must prove you've been suppressed — using a system deliberately designed to produce no evidence. You're being asked to submit a document that the system was designed never to generate.
And "unprovability" is not a side-effect. It is a specification. It was written into that 2015 application as the core feature of the invention: "so that the user will not realise."
I once wrote that the system doesn't need to deny you — it only needs to file you.
At the time, I thought that was a metaphor.
Sources
- US Patent 10356024, "Moderating content in an online forum," filed February 2015, granted July 2019
- Meta Transparency Center: "Reducing the Distribution of Problematic Content," "Types of Content We Demote," "How Facebook Group Feed Ranking Works"
- Meta's EU DSA Transparency Report (Facebook)
- EU Digital Services Act, Article 17
- Reuters (Jeff Horwitz, Engen Tham), Meta scam advertising internal documents series, November 2025
This Series
Attenuation: How Visibility Is Allocated — Three Volumes
- Volume I Approved, Then Unseen — How a Patent Made Silencing Leave No Evidence (this article)
- Volume II The Least AI-Like Article Was Actually Written by AI — A Real-Time Experiment in Cognitive Blind Spots
- Volume III One Like, One Point; One Flame War, Thirty Points — After the Forums Disappeared, Complexity Lost Its Place
- Criminal Investigation Bureau, Ministry of the Interior (Taiwan); Taiwan High Prosecutors Office: Investment fraud methods disclosure