In 1917 the Committee on Public Information put 75,000 volunteer speakers in front of audiences, and every one had to be in the room. In 2024 a Russian bot farm made 968 fake Americans without anyone. The record shows generative models cut the cost of producing persuaders, not of being heard.
Sometime in 1918, in a movie theater in one of 5,200 American communities, the film stops for the intermission and a slide goes up. It gives a local man's name and says he "will speak four minutes on a subject of national importance. He speaks under the authority of The Committee on Public Information. George Creel, Chairman, Washington, D.C." Then the man stands up in front of his neighbors and talks.
He was one of about 75,000. Woodrow Wilson created the Committee on Public Information with Executive Order 2594 on April 13, 1917, and the Four Minute Men started almost at once. The idea came from Donald Ryerson of Chicago, "who made the first Four Minute speech in a Chicago theater" and brought the plan to Creel in Washington two days after the committee was formed. Topics were "assigned by the department at Washington," and local chairmen received a brief with a model talk for each campaign. The bulletins ran from "Universal Service by Selective Draft" and the First Liberty Loan through "Food Conservation" to, with no apparent irony, "Unmasking German Propaganda." The organization was demobilized on December 24, 1918.
The committee kept count. By its own tally, 75,000 speakers in 5,200 communities made 755,190 speeches, and "a fair estimate of audiences" put the total reached at 314,454,514. Allowing for towns that never reported, it estimated "a million speeches heard by four hundred million individuals" over the organization's 18 months. The audience numbers were estimates, and the committee said so. The speakers were not. Every one of them had to be in the room.
Wilson thanked the Four Minute Men in a letter dated November 29, 1918. He marveled that 75,000 speakers had done so much "at a cost to the Government of little more than $100,000 for the 18-month period," which he put at "less than $1 yearly on an individual basis."
That was the government's bill, and the committee knew it was not the cost. In the same report it priced what it had not paid for. A four-minute speech, it decided, was worth at least $4, a figure it called "ridiculously small" but chose "because of its numerical appropriateness." It valued each theater's contribution at half that, a "rental value" per speech. The summary reads like an engineer's cost model:
| item | valuation |
|---|---|
| official expenditures (headquarters) | $101,555.10 |
| contributed expenditures | $2,564,970.00 |
| one million speeches at $4 each | $4,000,000.00 |
| "rent" of theaters, etc. | $2,000,000.00 |
| speeches of traveling speakers (331) | $8,275.00 |
| publicity contributed by press | $750,000.00 |
| grand total | $9,424,800.10 |
The government's cash was about 1% of the total, which the report describes as "nearly 95 times as much" as the money spent. Most of the rest was people and rooms: volunteers who prepared a talk from a Washington brief, theater managers who gave up an intermission, audiences who had come to see a film. The Four Minute Men were cheap in dollars and expensive in presence.
That presence did three things nobody designed it to do.
It capped volume. The committee could reach as many rooms as it had volunteers willing to stand up in them, and every volunteer had to be recruited, briefed and scheduled. When it wanted more reach, it had to find more rooms: a church department carried the talks into "churches, synagogues, and Sunday schools."
It attached a name. The slide told the audience who was speaking and on whose authority before he said a word, and he was a local volunteer, someone they could find the next day.
It left records. The committee tallied speeches campaign by campaign, and the chairman's report, printed by the Government Printing Office in 1920, still lets you check the arithmetic.
Friction guaranteed that evidence existed. It did not guarantee anyone would see it. Britain's counterpart, the War Propaganda Bureau at Wellington House, went the other way. When Charles Masterman met well-known authors and poets on September 2, 1914, to recruit them, everyone present agreed to keep the meeting and the bureau itself confidential, as the Imperial War Museum's Lives of the First World War project records. The paper existed. The public was not meant to read it.
On July 9, 2024, the FBI and the Cyber National Mission Force, with Dutch and Canadian partners, published an advisory about a piece of software called Meliorator. Affiliates of RT, the Russian state-sponsored broadcaster, had used it "to create fictitious online personas, representing a number of nationalities, to post content on X." The advisory describes it as "an AI-enabled bot farm generation and management software," with an administrator panel called Brigadir and a back end called Taras. The panel's vocabulary is revealing. It had tabs for "souls," which the advisory defines as "false identities that would create the basis for the bots," and "thoughts," "the automated scenarios or actions that could be implemented on behalf of the bots." Taras held the .json files that controlled the personas.
The same day, as NBC News reported, Justice Department prosecutors filed a list in federal court of 968 fake accounts created between June 2022 and March 2024. The accounts posed as Americans. One posed as a bitcoin investor in Minneapolis. The department said the operatives built them with the help of artificial intelligence, which made it easier to produce the images and the text.
Set that next to the 1917 ledger. The Four Minute Men needed 75,000 volunteers to put a name in front of an audience. A Four Minute Man was a neighbor with his name on a slide; a Meliorator persona was a "soul" in an admin panel. The software produced 968 names without a single volunteer, room or intermission, and the biography on each account did the slide's job, falsely.
Here the story usually turns to alarm. The record says something more useful.
OpenAI's May 2024 threat report covered five covert operations that had used its models. Two were Russian, Bad Grammar and Doppelganger; one Chinese, Spamouflage; one Iranian, the International Union of Virtual Media; and one was a commercial company in Israel, STOIC, whose operation OpenAI nicknamed Zero Zeno "to reflect the low levels of engagement that its various campaigns attracted." OpenAI rated each on the Breakout Scale, which scores an operation's impact from 1 to 6 "depending on whether they are amplified by real people, quoted uncritically by the media or influencers, or have an impact on policy or politics." The report's verdict was that "none of the five operations included in our case studies scored higher than a 2, indicating activity on multiple platforms, but no breakout to authentic audiences." Bad Grammar stayed at 1. In October, OpenAI said every election-related operation it had disrupted that year was in Category Two, "meaning that their ability to reach real people across the internet remained limited."
The operators knew distribution was the hard part, and they tried to fake that too. Some used OpenAI's models to generate "replies to their own posts to create false online engagement." Zero Zeno posted texts on Instagram and X and had a further set of its own accounts reply to them. The report's comment is dry: this "is distinct from attracting authentic engagement, which none of the networks described here managed to do." Nor had the operators left the room entirely. AI, the report says, "does not change the operators themselves," and Bad Grammar "posted content that included refusal messages from our model, exposing their content as AI-generated."
The May report's explanation of all this is the hinge of this essay. "To reach any kind of audience, such content requires a distribution system, and it has to compete against all the content produced by real people across social media. In every operation we found, the content they generated failed to build interest among authentic audiences."
Two caveats belong beside that. These are the operations OpenAI found, so they say nothing about the ones nobody caught. And OpenAI warned that these operations could still climb the scale, "especially if they are unwittingly amplified by media, influencers or politicians."
Put the two records side by side and the argument sharpens. The Four Minute Men never had a distribution problem. The theater was the distribution: a room full of people who had paid for their seats and were already sitting there when the speaker stood up. Production and distribution came bundled in one body standing in one room, and what made 1917 propaganda expensive was that you could not buy one without the other.
Generative models unbundled them. They pushed the cost of producing a persuader, with a name, a photo and a posting history, close to zero, though not all the way: OpenAI found that "all of these operations used AI to some degree, but none used it exclusively." They did little for the cost of being heard. In 2024 that was the step where the operations stalled, and it is the one piece of the old friction still standing.
Spam is persuasion whose production cost fell to almost nothing, and in 2008 researchers from UC San Diego and the International Computer Science Institute measured what that does. They infiltrated the Storm botnet and followed its pharmacy spam to the sale: "After 26 days, and almost 350 million e-mail messages, only 28 sales resulted." The conversion rate was, in their words, "well under 0.00001%." Extrapolating on what they called "an admittedly dangerous assumption," they estimated roughly $3.5 million in revenue a year. At zero marginal cost, one sale in millions was still a business, so the volume kept coming, and the fight moved to the receiving end. That is where the Spamalytics authors did their measuring: how much of the mail was delivered, and how much the filters and blacklists stopped.
Cryptographers had already named the other lever. In a paper first presented at Crypto '92, Cynthia Dwork and Moni Naor proposed fighting junk mail by making each sender pay in computation: "require a user to compute a moderately hard, but not intractable, function in order to gain access to the resource, thus preventing frivolous use." That is the 1917 friction rebuilt on purpose, a cost an honest sender barely notices and a bulk sender cannot ignore. The AI companies run their own version. OpenAI's May report says "our models are designed to impose friction on threat actors," and describes Doppelganger trying to generate images of European politicians, "only to be refused by the model."
If you run anything where accounts or content can be created, whether a forum, a review system, a marketplace, a social feature or an API, you are running the 1918 theater. Your cost structure is part of your security model, and the record suggests three habits.
Charge for reach, not for writing. You cannot make writing expensive again, and honest users need cheap generation too. You can make reach expensive for identities that have not earned it. Cap how many people a new account can reach, slow mass replies and mentions, and let reach grow with a history of real interaction. That rebuilds the 1917 bundle on purpose: to be heard, you have to have been present.
Keep the slide. Every Four Minute Man was introduced by name and authority before he spoke. Anything that reaches a large audience on your platform should carry, in a form your systems can check, who is behind it: the account's age and history, any verified affiliation, and a label when the content is machine-generated.
Keep the ledger, and count breakout rather than volume. The committee tallied its speeches. OpenAI's October report says threat actors mostly used its models in "a specific, intermediate phase of activity," after they had acquired accounts and before they published, and that watching that phase lets AI companies complement the "downstream" distribution platforms. Log account creation and reach well enough to reconstruct a network after the fact. Then measure what the Breakout Scale measures, whether real people picked the content up. Nine hundred accounts talking to each other are a cleanup job. One post quoted by a real influencer is an incident.
In 1918 a speaker had to stand up in front of the people he wanted to persuade, under his own name. Nobody passed a law requiring that. It was simply what persuasion cost, and for anyone building the places where people now listen, it has become a design choice.
Every Four Minute Man was introduced by name, and on whose authority, before he said a word. If you run agents that post or publish, Chain of Consciousness keeps a tamper-evident record of what each agent did and on whose behalf, so content that reaches an audience can carry a checkable account of who is behind it.
pip install chain-of-consciousness npm install chain-of-consciousness
Or start without installing anything: Hosted Chain of Consciousness.