When Marta opens a video app at night, she is not looking for a claim about health. She wants to switch off after caring for her father. Within minutes, the app offers a clip saying that a vaccine has been “withdrawn in secret.” The video remains online because it does not breach a rule requiring immediate removal. Yet it is not merely stored somewhere: it follows videos about fatigue and chronic illness, carries a caption promising what “they do not want you to hear,” and sits beside ads for supplements. The next day Marta sends it to her sister and buys a product advertised alongside it. Saying that the video was not taken down describes only a very small part of what happened.
Public debate often turns this situation into one question: should it have been censored or not? That question matters when content is removed, but it cannot explain all of a platform's power. Publishing, amplifying, monetizing, and recommending are separate acts, performed through different decisions and systems, with different risks and remedies. A service can allow someone to post an opinion without having to turn it into a tailored invitation, a revenue stream, or a priority answer for a vulnerable viewer. Naming those layers does not weaken protection for expression. It makes that protection more exact.
Four verbs, four kinds of responsibility#
Publishing makes content accessible. It may follow a person's choice, community rules, or a legal obligation. This layer calls for understandable standards, notice when a limit is applied, and a route to appeal. If a post stays online, the platform is saying at least that it may remain on its infrastructure under its rules. It is not necessarily saying that the post is reliable, timely, or suitable for every audience.
Publishing makes content accessible. It may follow a person's choice, community rules, or a legal obligation.
Amplifying makes it more likely that someone encounters content without seeking it out. Trending lists, notifications, autoplay, suggested shares, striking labels, and prominent search placement all amplify. They turn one item among millions into an item with reach. It is therefore not neutral to invoke the speaker's freedom as a reason for every distribution choice. The speaker has a legitimate interest in being heard; the company that builds the megaphone decides who is offered it, when, and with which signs of authority.
Monetizing turns attention into money, whether by paying a creator, placing advertising, selling a subscription, or using the content to assemble commercially valuable audiences. That layer raises an uncomfortable question. If a company finds a message too unreliable, manipulative, or harmful to endorse openly, why does it treat the message as advertising inventory? Removing an incentive does not erase anyone's words. It prevents a commercial structure from rewarding a clip precisely because it generates fear, anger, or alarm.
Recommending actively selects content for a person or group using signals such as viewing history, rough location, watch time, or patterns associated with other accounts. Recommendation can be useful: it helps people find music, communities, and relevant explanations. It can also exploit a moment of vulnerability. In Marta's case, the important fact is not only that the clip exists. The system connects it with someone who has watched material about symptoms and may be looking for a quick answer. That connection is a product choice even if software carries it out.
A proportionate response follows the pathway of harm#
Assume the clip contains misleading claims but no immediate instruction to cause harm. A proportionate response may leave it accessible with clear context while slowing its automatic circulation during review. The platform can remove it from health recommendations, keep it out of sensitive searches, reject paid promotion, and pause earnings tied to its performance. Each measure targets a distinct lever: induced exposure or financial incentive, rather than punishment for an unpopular opinion.
Proportionality requires an account of both the likely harm and the mechanism producing it. When the danger is a confusing claim reaching people who did not ask for it, the first answer need not be a global deletion. It can be lower recommended distribution, links to accountable health information, and an explanation of why the topic receives special care. If the creator attempts to buy ads aimed at people with particular illnesses, stopping that campaign responds to a specific commercial relationship. If advertising repeatedly funds content that distorts evidence, withdrawing that monetization protects the coherence of the advertising system itself.
This does not make every label or reduction fair. A platform can be wrong, enforce rules unevenly, or hide a commercial preference behind the language of safety. A proportionate intervention therefore needs a limited scope, recorded reasons, a review date, and a meaningful appeal. The person who posted should know what changed: whether the video remains visible, whether it stopped being recommended, what revenue was paused, and what evidence was considered. Without that separation, public suspicion is understandable.
Transparency should make argument possible#
A useful transparency panel does not have to disclose all source code or every user's private data. It should show which kinds of signals tend to raise or lower a recommendation, which categories receive special protections, and which economic actions attach to content. In plain language, a platform could say that health videos without verifiable sources do not receive targeted promotion, do not enter recommendations after symptom searches, and cannot carry personalized advertising until review is complete. It should also report how many cases received that treatment, how long it lasted, and how many appeals changed the outcome.
The explanation must reach the person receiving the suggestion. Beside the video, Marta might see: “You are seeing this because you followed wellness accounts; this claim is disputed and has not been promoted as a health source.” That does not replace Marta's judgment or treat her as a child. It makes clear that the screen is not a spontaneous display. She can leave that pathway, request fewer recommendations on the subject, or open sources that state their evidence standard.
Information is not the same as an automatic excuse. “The algorithm decided” explains nothing. A responsible organization identifies the rule, the team that approved it, and the process for examining unintended effects. If the measure disproportionately reduces the visibility of small communities, independent reporters, or speakers in a minority language, that consequence belongs in the evaluation. Transparency that only lists large removal totals cannot show whether recommendation power was exercised with care.
Two shortcuts to avoid#
The first shortcut calls every distribution change censorship. The word can describe real abuses, especially when an authority silences criticism or blocks access to ideas. Using it for everything erases a crucial difference between banning a voice and declining to pay for its induced reach. A platform does not automatically become an arbiter of truth when it stops recommending an item; it does, however, owe accountability because it organizes public attention.
The opposite shortcut treats every intervention as technical housekeeping and ignores that reduced visibility can shape a conversation. It may harm a creator, misread irony, or overlook new evidence. The remedy is not to pretend that power is absent. It is to place power under contestable rules: notice before income is removed where possible, human review for complex cases, an appeal that does not require legal expertise, and an aggregate record that can expose bias.
Public institutions have a particular role here. Rather than demand the deletion of every disputed post, they can require comparable reports on promotion, recommendation, and advertising; support media literacy that teaches the four layers; and protect independent research into system effects. Advertisers can also decide where their money goes and insist that their campaigns do not fuel profitable deception. No single action solves misinformation, but each one reveals who is paying for a claim to travel farther.
A practical question for every lever#
Before acting, decision-makers should ask four separate questions. May this content remain available? Should we actively carry it to people who did not request it? Should we turn attention to it into revenue? Why is it being offered to this audience at this moment? The answers may align, but they should never be assumed. A video can remain published and cease to be recommended; it can be available through a deliberate search yet unsuitable for a paid campaign; it can receive context without treating its author as an enemy.
For Marta, a fairer response is neither a blank screen nor a moral lecture. It is being able to see that the video reached her through a recommendation decision, knowing it is not being financed through targeted ads, and finding reliable vaccine information one click away. For the platform, the duty is not to promise infallibility. It is to explain each layer, measure its effects, and repair mistakes. Moving beyond the censorship-freedom binary does not reduce freedom. It stops a merely formal freedom from concealing who purchases, ranks, and accelerates other people's attention.
