The algorithm deciding who sees your speech

San Francisco — The message that ended Priya Nandakumar's reach never arrived. There was no notification, no email, no strike against her account. Over four days last autumn, the small-business owner watched the views on her cookware tutorials collapse from a reliable forty thousand to fewer than nine hundred, then to almost nothing. Her account remained, in every formal sense, in good standing. She had simply, and silently, stopped being shown to people. When she searched the platform's help centre for an explanation, she found a single sentence: that distribution is "not guaranteed" and reflects "many signals." There was no button to press, no form to file, no human to ask. The machine had decided, and the machine did not explain itself.

Nandakumar's experience is not a malfunction. It is the system working as designed. Across the largest social platforms, the consequential decision is no longer whether a post is removed but how far it travels — and that question is answered by ranking and demotion systems that operate with none of the procedural scaffolding built around outright deletion. There is no notice, no stated reason, and, crucially, no appeal.

The quiet penalty

Engineers call it "downranking," "deboosting," or, in the candid shorthand of internal documents reviewed by Meridian, simply "the throttle." Where a content takedown is a discrete, auditable act, demotion is a continuous adjustment to a score — a multiplier applied to a post's predicted reach. A piece of content judged "borderline" by a classifier might have its distribution cut by sixty, eighty, or ninety per cent. The post stays up. It is visible to the person who wrote it, and to almost no one else.

This design solves a real problem for platforms. Removals generate complaints, press coverage, and regulatory attention; they create a record. Demotion generates none of these, because the affected user typically cannot tell it has happened. The penalty is invisible to its target by construction, which is precisely what makes it attractive — and what makes it unaccountable.

A removal is an argument the user can have. A demotion is a verdict they are never told was reached.

Signals without meaning

The systems that assign these scores are not, in the main, reading posts for meaning. They are predicting engagement and risk from thousands of proxy features: how a post resembles others previously flagged, the behaviour of accounts that interact with it, the velocity of its early shares, the presence of certain links or phrases. A former trust-and-safety engineer who worked on one such ranking pipeline described the result as a model that is confident and illegible in equal measure.

You can ask the model why a post scored low and get a list of feature weights, but that is not a reason a person could act on. We could not have written an appeal letter on the user's behalf, because we did not know, in any human sense, what they had done wrong.Former trust-and-safety engineer, large social platform

That illegibility is not incidental. Modern ranking stacks chain together dozens of models whose outputs feed one another, retrained continuously on fresh data. A post demoted today might be scored differently tomorrow by a system that has, in the interim, learned something no one specifically taught it. There is often no single threshold a moderator set, no rule written in a policy document — only an emergent boundary that shifts beneath users who cannot see it.

Cases on the record

Nandakumar is one of eleven users and four current or former platform staff who described, in interviews over three months, a consistent pattern of consequential reach reduction without recourse. The cases span hobbyists, journalists, and small commercial accounts. What unites them is not the content but the procedural void around its suppression.

In none of these cases could the user obtain a stated reason. Several reported reaching support agents who were themselves unable to see whether a demotion had been applied, because the relevant scores sit in systems support staff cannot query. The decision exists, affects livelihoods, and is legible to no one positioned to defend it.

The engineering culture

Why build it this way? The answer the engineers offered was less about malice than about incentives. Reach is a continuous quantity, easy to dial; due process is discrete and expensive. Every notification invites a dispute, every dispute needs a reviewer, every reviewer costs money and introduces latency. A system that quietly adjusts a multiplier scales infinitely; a system that explains itself to each affected user does not.

• Demotion is treated as a "ranking decision," exempting it from the notice-and-appeal commitments that govern removals.
• Soft actions are deliberately excluded from appeal flows to suppress appeal volume.
• Support staff frequently lack tooling to see whether an account has been demoted.
• Models are retrained continuously, so no fixed, citable rule underlies a given penalty.

There is a tidy logic to it that troubled even those who built it. The same property that makes demotion humane in theory — it is reversible, proportionate, less punitive than deletion — makes it unaccountable in practice, because nothing about it is ever disclosed.

What remains open

Regulators are beginning to notice the gap. New transparency rules in several jurisdictions require platforms to explain "restrictions" on content, and lawyers are now testing whether demotion counts. The platforms argue that ranking is editorial discretion, not restriction — that no one is owed a particular audience. That argument may hold legally even as it fails users like Nandakumar, who never sought a guaranteed audience, only the one she had built and an account of why it vanished. The deeper question the technology forces is not whether platforms may rank speech, which they plainly must, but whether a decision that materially silences a person can be allowed to remain, by design, a thing no one is ever told about. For now, the throttle turns quietly, and the people it acts on are left to guess.