The Twitter Files Showed a Process, Not a Plot, and That Is Worse

A conspiracy has a wonderful property: it can be ended. Find the conspirators, expose the plan, fire the people, and the thing stops. Every thriller ever written runs on that premise. The villain is identified, the lights come on, justice is done by the final reel.

A process is a different animal. A process has no villain. It has a policy manual, a ticketing system, a weekly meeting and a hundred reasonable people each making a reasonable-looking decision. Nobody planned the outcome, so nobody can be fired to stop it. The outcome simply emerges, week after week, from the machinery. That is why I argue that the internal Twitter documents released beginning in December 2022 are more troubling when read as evidence of a process than they would have been as evidence of a plot.

Let me define my terms, state the claim as a numbered argument, give the other side's best case its due and then show where it fails.

What was released, and what was not claimed

In early December 2022, after Elon Musk's purchase of the company, a series of internal Twitter documents and communications were released to selected independent journalists, among them Matt Taibbi, Bari Weiss and Michael Shellenberger, who published their findings in long threads on the platform, now called X. The releases became known as the Twitter Files.

I will be careful about what I claim, because this subject attracts exaggeration from both directions. The releases were selective: the journalists were given access under conditions set by the company's new owner, and they described what they were shown, which is not the same as the whole record. They did not, in my reading, reveal a single secret order from a government official to silence a named person. Some commentators claimed more than that, and some critics responded by dismissing the whole exercise. Both reactions miss what was actually documented.

What was documented, on the accounts published, falls into three categories.

1. Decisions about content. In October 2020 the company limited the sharing of a New York newspaper's report about the contents of a laptop attributed to Hunter Biden, blocking links and locking the newspaper's account, citing its policy on hacked materials. Its chief executive, Jack Dorsey, later called the decision a mistake, and the company revised the policy. I discuss the surrounding story in The Laptop Story Twitter Blocked and the Letter That Gave It Cover.
2. Visibility filtering. The documents described internal tools by which accounts could be tagged so that their posts were less visible without the account holder being told: removal from search suggestions, exclusion from trending topics, limits on amplification. The company had long said that it did not "shadow ban" in the sense of hiding accounts entirely, and the documents showed that it did use a family of visibility controls. According to the published threads, accounts carrying such labels included a prominent medical academic who had criticized pandemic policy and several political commentators, though critics noted that the threads did not show how many accounts were affected or what their politics were.
3. Contacts with government. The threads described frequent emails and meetings between Twitter's trust-and-safety staff and federal agencies, including the FBI, about possible foreign interference and about particular accounts and posts that officials flagged. The FBI said the correspondence reflected routine engagement.

Those are the materials. Now the argument.

The thesis in five premises

Here is the argument, point-by-point, in numbered steps.

1. Content moderation at scale is machinery, not a series of individual choices. A platform with hundreds of millions of users cannot review each post by hand. It builds classifiers, labels, queues and escalation paths, and it writes policies to guide them.
2. Machinery embeds its designers' assumptions and its funders' pressures. A rule about "hacked materials," a category called "misinformation," a label called "do not amplify": each reflects a judgment someone made once and the system applies a million times.
3. Government contact enters the machinery as one more input. When federal agencies send flagged items to a company that has a dedicated process for receiving them, the company's staff treat the flags as signals to be triaged, and the items move through the same queues.
4. No one in the machinery has to intend censorship for censorship to occur. The staffer who applies a label thinks he is following policy. The official who flags an account thinks she is reporting a risk. The executive who approves the process thinks it is responsible. Each acts in good faith. The aggregate is a system that suppresses speech, to an extent invisible to the speaker and to the public.
5. Therefore the risk is institutional habit, not individual villainy, and the remedy must be structural. You cannot fix a habit by firing one man. You have to change the incentives, the transparency and the legal boundaries around it.

If those premises hold, the right way to read the Twitter Files is as a case study in how well-meaning institutions drift into suppressing speech without anyone deciding to.

The best objection: this was ordinary trust-and-safety work

The strongest counter-argument deserves a full statement, because it is not stupid.

It goes like this. Every large platform needs rules about spam, fraud, harassment, impersonation, foreign manipulation and illegal content. Enforcing them requires tools, including tools that reduce the reach of abusive accounts without permanently banning them. That is not sinister. It is the job. Companies that do not do it become unusable swamps.

Moreover, the argument continues, government agencies have a legitimate role in alerting companies to threats. Law enforcement may tell a company that a network of accounts is run by a foreign intelligence service, and a company would be negligent to ignore it. Meetings between companies and agencies are routine, and the companies made their own decisions, often declining requests. The documents show staff debating, disagreeing and sometimes pushing back. Indeed the company's own people were not unanimous about the laptop story.

And finally: Twitter was a private company applying its own rules. Whatever one thinks of those rules, a company is allowed to have them. The company's new owner chose which materials to release, so the picture is curated.

I accept a good deal of that. I accept that visibility filtering is a normal tool and that a platform without it would be worse. I accept that some government flags concerned genuine threats. I accept that the release was selective. I accept that private companies may set their own rules, a point I defend in Moody v. NetChoice and the Unfinished Question of Who Owns the Feed.

But notice what the objection concedes. It says the machinery was routine. That is my point. Routine is the problem.

Why routine machinery is harder to fix than a plot

Let me answer the objection in four points.

One: a plot has an end date, and a process does not. If the Twitter Files had exposed a scheme hatched by three executives, the remedy would be to remove three executives. Because the documents show a standing apparatus, with its categories, its recurring meetings and its tooling, removing individuals leaves the apparatus in place. New people inherit the dashboard.

Two: routine machinery generates no alarms. Nobody inside a process feels that they are doing something wrong, because every step is defensible. The staffer who applies a "do not amplify" label has a policy citation. The official who flags a post has a mission statement. The usual internal brake, a person who says "this feels wrong," never engages, because nothing feels wrong. A plot at least requires someone to lie to his own conscience.

Three: opaque tools shift the burden onto the speaker. A person whose account is quietly demoted does not know it. He cannot appeal what he cannot see. The old principle that a citizen is entitled to notice and a chance to respond does not apply to a filter nobody tells you about. Transparency is the missing safeguard, and critics of the company argued that its public statements about its practices did not fully describe the tools it used, a charge the company disputed.

Four: the incentive to over-remove is built in. A platform that is criticized when harmful content stays up and rarely criticized when legitimate content comes down will drift toward removal. This is the same perverse incentive that I describe in The 1995 Prodigy Lawsuit That Accidentally Created Section 230 from the legal side. Add periodic pressure from officials and from advocacy groups, and the drift accelerates. Nobody has to want censorship. The incentives supply it.

The government contacts: where a private process meets public power

The portion of the documents that matters most constitutionally is the regular contact with federal agencies. A private company's moderation, however flawed, is protected by its own First Amendment rights and by the state-action doctrine. When the government is directing or pressing that moderation, the analysis changes.

The line, as I explain in NRA v. Vullo and the Unanimous Rule Against Government Jawboning, is between persuasion and coercion. A law enforcement agency telling a company about a foreign operation is on one side of the line. An official repeatedly pressing a company to remove lawful domestic speech, with the implicit weight of regulatory power behind the request, is on the other. The Twitter documents, at least as published, did not by themselves establish that the second category was widespread, and I will not claim they did. But they showed that a channel existed, that it was regular, and that it handled domestic content as well as foreign threats.

The Supreme Court's decision in Murthy v. Missouri, covered in Murthy v. Missouri Lost on Standing, Yet the Record It Built Still Stands, turned on whether plaintiffs could prove that particular government requests caused particular removals. That difficulty is precisely what a routine process creates. When the contact is regular, informal and unlogged, causation disappears into the machinery, and the speaker cannot prove what he cannot see.

The Founding anchor

James Madison, in Federalist No. 51, wrote one of the most quoted sentences in American political thought: "If men were angels, no government would be necessary." He continued that if angels governed men, neither external nor internal controls on government would be necessary, and that in framing a government administered by men over men, the great difficulty is first to enable the government to control the governed, and in the next place to oblige it to control itself.

Read the Twitter Files in that light. The Founders did not build the separation of powers because they expected villains in office. They built it because they expected ordinary people, with ordinary motives and ordinary blind spots, to accumulate power and use it carelessly. The design was structural: divide authority, require transparency, give each part of the government an interest in checking the others. A system built on the assumption that good people will behave well is a system waiting for a lapse.

The same logic applies to the relationship between government and private speech platforms. We do not need to believe anyone is a bad actor to see the need for rules. We need rules because the incentives are what they are. Disclosure of government requests, notice to affected users, a path of appeal and a clear legal line against official pressure are the structural equivalents of the separation of powers for the age of the feed.

What a structural remedy would look like

If the problem is institutional habit, here is what I would propose, in order.

1. Log and publish government requests. Any request by a government official to a platform concerning a specific lawful post or account should be recorded and, subject to narrow security exceptions, made public after a delay. Sunlight is a better disinfectant than any internal policy.
2. Notice and appeal for users. Platforms should tell users when their reach has been limited, why, and how to contest it. A visibility filter that cannot be seen cannot be challenged.
3. Truth in description. Companies should describe their moderation tools accurately in public terms of service, so that a promise not to "shadow ban" means what it says.
4. A bright line for officials. Agencies should adopt written rules forbidding any official from linking regulatory treatment to a company's decisions about lawful speech, and Congress should hold oversight hearings to confirm compliance.
5. Competition. Antitrust enforcement, interoperability and easy migration reduce the power of any single company's machinery. A user who can leave is a user whose complaint is heard.

None of these requires a conspiracy theory or a government referee of truth. All of them address the thing the documents actually showed.

What we should not conclude

I want to be as clear about the limits as about the lessons. The Twitter Files do not prove an election was manipulated, and I make no such claim. They do not show that every filtered account was filtered for political reasons, and some of the filters applied across the spectrum. They do not show that the company's staff were uniformly hostile to conservatives; the documents, as published, show people debating hard cases. And they do not mean that content moderation is illegitimate. A platform may remove harassment, spam and fraud.

What they do show, taken fairly, is that powerful institutions can build quiet systems for shaping what the public sees, that those systems can include a channel from government, and that the people who run them can be sincere. Sincerity is not a safeguard. For the wider case against government deciding what counts as false, see Misinformation Is Not a Legal Category, and Government Should Stop Pretending; for a federal attempt to institutionalize that role, see The Disinformation Governance Board Lasted Four Months, and That Was Too Long. For the legal myths that cloud the debate, see Publisher or Platform, Eight Myths About Big Tech and the Law.

The case in brief

1. The Twitter Files, released from December 2022 to selected journalists, documented content decisions, visibility filtering and regular contacts with federal agencies, without establishing a single master plan.
2. Moderation at scale is machinery, and machinery embeds assumptions and incentives that no individual has to intend.
3. The best objection, that this was ordinary trust-and-safety work, is largely right on its facts and proves the point: routine machinery has no end date, raises no alarms and hides its effects from the speaker.
4. The constitutional concern lies in the government channel, where the line between persuasion and coercion must be policed by transparency and clear rules.
5. The remedy is structural: logging government requests, notice and appeal for users, truthful descriptions of tools, a bright line for officials and competition.

What you can do: ask your representatives whether they will support a law requiring federal agencies to publicly log their requests to social media companies about lawful speech.