Censorship of social media: a working reference

Censorship of social media is the removal or suppression of what a person has posted, or of the account itself, by an authority the person did not choose. Three kinds of authority do this, and they are often confused with one another.

States censor by law and by force. They order posts removed, they block or throttle services, they shut networks down, and they prosecute people for what they wrote. Platforms censor by contract. Their terms of service say what may not be posted, classifiers and moderators enforce those terms, and the result is a label, reduced reach, a removal or a ban. Between the two lies a grey area in which governments ask and companies agree, with no order that anyone could appeal.

The right at stake is old and international. Article 19 of the Universal Declaration of Human Rights and of the International Covenant on Civil and Political Rights, Article 10 of the European Convention on Human Rights and the First Amendment of the United States Constitution all protect expression. Each of them binds governments. None of them binds a private platform directly, which is why the same deleted post is a question of rights in one setting and a question of contract in another.

The notes that follow describe how states censor, how platforms censor, the laws now written for both, and what censorship does to the people it touches. Numbers are the sources' own, and dates are matters of record.

A state that wants a service silenced has a range of tools, and the tool it chooses says as much as the decision to use it.

The mildest is the order. A law or an official directs the platform to remove posts, to hand over data or to restrict accounts, and the platform decides whether to comply. Turkey, India and Russia each have a statute for this and a record of using it. Next come the laws that require a company to keep a representative in the country. That person can be fined or arrested when the company does not comply, which gives an order a weight the company cannot ignore.

The technical measures follow. Domain names are removed from the national resolvers, addresses are dropped at the border, and connections are throttled until a service becomes unusable. These measures can be observed from outside the country. Projects such as OONI, NetBlocks and IODA run probes that show what a network inside the country can and cannot reach.

The strongest measure is the shutdown: the whole network, or all of social media, switched off around an election, an examination or a protest. The #KeepItOn coalition, run by Access Now, keeps the annual count.

Two cases from 2024 show the range in use. A justice of the Supreme Court of Brazil ordered X blocked across the country after the company refused to name a legal representative, and the block held for more than a month before the company complied. In Russia, YouTube slowed to the point of being unusable while officials attributed the slowdown to Google's own equipment.

What changes a state's choice is the cost of a block to its own economy, the attention of other governments, and the willingness of the company to be blocked rather than comply. On the record, that willingness has gone both ways.

A platform's censorship begins with a document. The terms of service and the community guidelines state what may not be posted. Enforcement is done first by classifiers, which are models trained to flag nudity, violence, spam and known illegal images, and then by moderators who review what the classifiers and other users report.

The outcomes form their own scale. A label or a warning screen is the mildest. Reduced reach comes next: the post stays up but is shown to fewer people. Users call this shadow banning and platforms call it demotion. Then come the removal of the post, the suspension of the account and the permanent ban. Reduced reach is the hardest outcome to see and the least often appealed, because the person affected is rarely told.

The record of all this is uneven. Most large platforms publish transparency reports twice a year, counting removals by category and government requests by country. In the European Union the Digital Services Act made the record itemised. Every restriction imposed on a user in the Union must come with a statement of reasons, and every statement is submitted to the Commission's public Transparency Database, which has received hundreds of millions of them since it opened in 2023. Meta's Oversight Board, an outside body funded by the company, hears a small number of appeals and publishes decisions that bind the company in each case.

What changes a platform's rules is pressure from advertisers, press coverage of a decision that went wrong in either direction, the cost of moderating a language the company has few staff for, and the law of the largest market the platform serves.

Three legal regimes now shape what social media removes, and they pull in different directions.

The European Union's Digital Services Act, Regulation (EU) 2022/2065, has applied in full since February 2024. Every hosting service must provide a way to report illegal content and must explain each restriction it imposes. The largest platforms must also assess and reduce systemic risks, including risks to elections, to public health and to fundamental rights such as freedom of expression, under the supervision of the Commission and with fines of up to six per cent of worldwide turnover. The Act names over-removal as a risk in its own right.

The United Kingdom's Online Safety Act 2023 places duties on any service with users in the United Kingdom, however small. Services must assess the risk of illegal content and remove it once they know of it, and must protect children from a defined list of harms, under codes of practice written by Ofcom. The duties on illegal content came into force in March 2025 and the duties on child safety in July 2025. Some small community sites closed rather than bear the cost of compliance.

In the United States, Section 230 of the Communications Act protects platforms from liability both for what users post and for removing it in good faith, and the First Amendment limits what a state may require of them. In July 2024 the Supreme Court, in Moody v. NetChoice, treated a platform's curation of its feed as expressive activity protected by the First Amendment and returned the Texas and Florida laws against so-called viewpoint moderation to the lower courts.

Germany's Network Enforcement Act of 2017, India's Information Technology Rules of 2021 and Brazil's Marco Civil of 2014, narrowed by the Supreme Federal Court in 2025, complete the picture. The same post can be compulsory to remove in one jurisdiction and unlawful to remove in another.

The first consequence of censorship leaves no record: the post that was never written. Surveys of journalists and activists in countries with active blocking find that self-censorship is the usual response, and the people who stop posting appear in no transparency report.

The second is displacement. A blocked or hostile platform sends its users elsewhere. Brazil's block of X in 2024 moved millions of accounts to Bluesky within days, and each wave of restrictions on the larger networks since 2022 has enlarged the federated ones. Audiences, archives and adversaries move with the users.

The third is circumvention, and its price. Virtual private networks, Tor and mirror sites keep a blocked service reachable for those who know how to use them, which is why blocking states then act against the VPNs, and the traffic that remains is easier to watch. Shutdowns have a cost that the economy pays whether or not anyone circumvents them. NetBlocks and several industry tallies put the yearly figure in the billions of dollars.

The fourth is the record itself. Most censorship is invisible from inside the country and is not announced by the platform, so what is known comes from measurement: the probes of OONI, the outage reports of NetBlocks, the statements of reasons in the Digital Services Act database, and the takedown notices archived by the Lumen database. This site follows those sources so that the record stays public.