A Threat Stays Up When the Right Account Posts It

In May 2020, Donald Trump published the same message on Facebook and Twitter during protests following the murder of George Floyd. It called protesters “THUGS,” threatened military intervention and included the line “when the looting starts, the shooting starts.” The phrase carried its own file: Miami police chief Walter Headley had used it in 1967 while announcing an aggressive campaign in Black neighborhoods.

Twitter covered the post with a notice saying it violated rules against glorifying violence but would remain accessible in the public interest. Users had to click through, and ordinary engagement was restricted. Facebook left its copy untouched. Mark Zuckerberg defended that decision under the company’s policies at the time, which distinguished warnings about state force from calls for private violence.

The text had not changed. Its author had not changed. What changed was the route through each company’s enforcement system.

That post remains the cleanest specimen of the public-interest exception because it strips away the comforting idea that moderation begins and ends with reading words. One platform treated the message as a violation worth preserving behind friction. Another treated it as permissible political speech. Had an unknown user posted the same sentence at a local activist, automated systems and frontline reviewers would have encountered a threat with none of the civic importance attached to a president’s account.

Public interest sounds like a principle. Operationally, it is a queue.

Four platforms, four escape hatches

Meta’s current rules allow a “newsworthiness allowance,” an exception that can keep violating content online when its public-interest value outweighs the risk of harm. The company says the allowance is rare and that politicians no longer receive an automatic presumption of newsworthiness. Reviewers are supposed to consider the speaker’s role, the subject’s political significance and the severity of the violation.

That balancing test contains the whole problem. Harm and news value are not measured with the same instruments. A credible threat may produce identifiable targets, while public value is inferred from status, reach and whether journalists are already paying attention. The first side describes danger.

The second describes importance, which platforms routinely confuse with prominence because prominence is available as account data.

X retains a documented public-interest exception for some posts by elected officials, government figures and candidates. Eligibility also turns on account characteristics, including reach and authenticity. A violating post may receive a notice while remaining viewable, with its circulation and engagement limited. The company thereby creates a second product out of prohibited speech: content too harmful for normal distribution but too politically consequential, by its own estimate, to delete.

YouTube uses a broader contextual route. Its EDSA exception, short for educational, documentary, scientific or artistic context, can preserve material that would otherwise breach policy when the surrounding video gives viewers enough information to understand why it is shown. Context often must appear in the video or audio itself rather than being tucked into metadata. A journalist who uploads footage of a threat as evidence may therefore survive review, while a clipped repost carrying the same footage can disappear.

TikTok also permits public-interest exceptions for documentary material, counterspeech and other contextual uses. Depending on the policy involved, the platform can attach warnings or make a post ineligible for the For You feed, its main recommendation surface. That last option matters. A video can remain technically available while losing the distribution that made publication meaningful in the first place.

These policies are often discussed as variations on free speech. They are closer to distribution licenses. The platform decides whether a post may exist, whether strangers can encounter it, whether users can repeat it and whether recommendation systems can turn it into an event.

The velvet rope is called escalation

Most users meet moderation through automation or a contractor working under time pressure. High-reach accounts can meet policy specialists, communications teams and designated escalation systems. The distinction is less cinematic than a secret command center. It is more durable too: one report enters an ordinary queue, while another arrives carrying enough institutional risk to summon people with authority to interpret the rules.

Meta’s Cross Check system made that hierarchy unusually visible. Public reporting based on internal documents showed that prominent users, including politicians and journalists, could receive additional review before enforcement. Meta described the system as protection against mistakes with severe consequences. Its Oversight Board later criticized the program’s opacity and unequal treatment, while recommending clearer criteria and deadlines.

Extra review is defensible. Removing documentation from a war zone or suspending a head of government can alter the public record. Yet the same safeguard also delays action against accounts whose reach increases the cost of delay. Cross Check did not merely protect speech from careless moderation.

It protected Meta from the reputational and political consequences of careless moderation against people capable of making trouble for Meta.

The incentive is plain. A mistaken removal affecting an ordinary user usually ends with an ignored appeal. A mistaken removal affecting a politician can trigger hearings, regulatory threats and days of hostile coverage. A journalist may have an editor, a platform contact or an audience able to publicize the error.

High-reach creators and advertisers can bring commercial leverage. Platforms call this risk management because “status-based due process” would look poor in the transparency report.

X writes part of the hierarchy into its public-interest eligibility rules. Meta embeds it in escalation and secondary review. YouTube and TikTok emphasize contextual exceptions rather than a general celebrity license, but professional publishers still have more capacity to package context correctly, contact platform representatives and make a removal embarrassing. Formal equality at the rulebook does not produce equal access to the person authorized to interpret it.

Return to the May 2020 post. Twitter’s notice did more than classify the words. It classified the account as historically significant enough that the violation itself became an artifact. Facebook reached a different policy outcome, but it performed the same status calculation before deciding that the post should remain.

An unknown account would struggle even to trigger that level of deliberation.

Evidence needs a costume

The hardest cases are reposts made to document abuse. A death threat in a journalist’s report may be evidence. The original is abuse. A clipped version pushed by supporters may become intimidation again.

Hash-matching systems, which identify copies by their digital signature, can detect similarity but cannot reliably determine which social act each copy performs.

Platforms try to recover intent from captions, account history and presentation. YouTube’s insistence that context appear inside the video shows how literal this can become. The uploader must supply the evidentiary costume before enforcement arrives. A newsroom with an editing desk can add narration, blur identifying details and preserve the source file.

A person recording police violence on a cheap phone may upload the raw clip because speed matters and storage is limited.

The public-interest exception therefore rewards legibility to institutions. Journalists know how to frame material as journalism. Politicians arrive preclassified as political. Large accounts generate internal consequences when moderators act.

Everyone else must hope that a caption survives automated scanning long enough for a human to understand it.

Even preservation can injure the people named in a threat. Warning screens reduce casual exposure, but they do not erase identifying details or stop committed users from taking screenshots. Removing recommendation eligibility can curb amplification, though it leaves the post available to followers already primed to act. Deletion cuts distribution fastest and can also destroy evidence needed by reporters, researchers or investigators.

No single switch resolves all four problems.

A better system would separate preservation from publication. Platforms could remove a credible threat from public circulation while retaining a secured copy, recording the policy basis and offering access through lawful or vetted research channels. Public figures would still receive rapid review, but the same deadline and explanation would apply to anyone facing a severe enforcement action. Exceptions would appear in searchable transparency records with identifying details redacted where safety requires it.

That costs money. It requires trained reviewers, durable archives and appeals that return before the news cycle has eaten the case. The cheaper system uses account status as a proxy for public importance and lets warning labels perform institutional seriousness.

The May 2020 post survived in two forms: one behind a warning and one in full view. Its afterlife explains why platforms preserve some threats. They are maintaining the public record, but they are also maintaining relationships with officials whose speech brings attention, regulatory danger and traffic. Newsworthiness is the language of the decision.

Power supplies the routing instructions.

Questions people ask

What is a public-interest exception in content moderation?

It is a policy that allows some rule-breaking content to remain available because a platform judges its civic, documentary or educational value greater than its likely harm. The platform may add a warning, restrict engagement or remove the post from recommendation feeds rather than delete it.

Do politicians get different moderation rules?

Some platforms formally consider political office when deciding whether violating content should remain accessible. Others deny blanket exemptions but route prominent accounts through additional review. Either way, politicians often receive more human attention and a more consequential appeal path than ordinary users.

Why can a journalist post material that another user cannot?

Context can change how a platform classifies identical footage or language. Reporting that identifies a threat, explains its significance and limits unnecessary exposure may qualify for an exception, while an unframed repost can look like praise, intimidation or renewed distribution of the abuse.

Does a warning label stop a threatening post from causing harm?

A label adds friction and may reduce casual sharing, but the material remains available to determined viewers and can still travel through screenshots. In the May 2020 case, Twitter’s notice constrained the post without removing it; Facebook’s copy remained fully visible.