On a Douyin relationship show in 2025, a caller told the host she was pregnant and worried about her husband. He was spending more time with his best friend and calling him “babe” in his texts. Viewers began weighing in. The host sounded surprised. She had sent the caller the script the day before. The caller was a reporter for Sixth Tone. Seven minutes on the phone paid about forty-eight cents.
In 2017, Netflix’s engineers used Good Will Hunting to explain how the service could get two people to try the same movie. A Netflix subscriber who watched romances might see Matt Damon and Minnie Driver together on the tile. Someone who watched comedies might see Robin Williams. The film stayed the same. The invitation changed according to what each viewer had already watched.
I had never gone looking for an earwax camera. Then a stranger in a shopping clip showed me how one worked. A clip like that might come from a customer who likes the camera, someone who earns a commission every time it sells, or a host with a list of products to get through before the end of a shift. It takes longer to figure out why the person is showing you the gadget than it does to find out what the gadget does.
A feed learns which clips keep you watching. The person making a clip can buy the appearance that strangers found it and liked it before you got there. The rest of the audience brings its own curiosity and judgment to a scene that someone else has already arranged. That mixture of paid performances, automated numbers, and real attention is part of what people are trying to describe when they call the internet dead.
RCA’s television demonstrations at the New York World’s Fair in 1939 carried radio’s basic arrangement into a new medium. A station transmitted to many receivers. You could turn the dial, telephone the station, or write to complain. A household in Nielsen’s audience sample could leave a record of what it had watched. Advertisers and programmers could later study ratings and make decisions from them. Once a show was coming through the set, there was little that a family in one living room could do to change what came next.
The broadcasters found a way to put another audience in the room with the family. In the 1950s the television sound engineer Charley Douglass developed the Laff Box. Douglass selected and timed recorded laughs. A punch line repeated over several takes acquired the sound of a crowd hearing it fresh. Even when a viewer knew what a laugh track was, the sound carried information about where a reaction belonged. Later experiments found that recorded laughter could change how amused listeners said they were, depending in part on the group they thought the laughter came from.
Theatre managers and performers paid for something like this in nineteenth-century Paris. At the opera, professional applauders called claqueurs started ovations and calls for encores. Louis-Désiré Véron, who directed the Paris Opera in the 1830s, tells a story about ordering the group to keep quiet while the dancer Pauline Duvernay performed. At the end of her pirouette she smiled, expecting applause, and the house gave her silence. Véron liked to tell the story from the manager’s side of the curtain. It is still striking that he could withdraw an organized response and leave a whole room unsure when to start its own. Suetonius had made the specialized sounds of Nero’s trained applauders one of the details in his account of the emperor’s vanity.
Warner Cable made a different kind of audience possible in Columbus, Ohio. On December 1, 1977, it launched QUBE, a thirty-channel service with a wired console and five response buttons. At home you could vote for acts on Talent Search. After the blizzard of 1978, Mayor Tom Moody came on Columbus Alive! and got feedback from people who had been waiting for their streets to be cleared. The host Dick Wolfsie once asked whether viewers found him too silly and recalled getting a majority verdict against him within six seconds. The same box could register that you wanted to see a movie or buy something and bill your household.
The households that had the box and were willing to press a button made up a public QUBE could count. A show put the result back on the screen while the people who produced it were still watching. That immediacy was new. The limits of the sample were already familiar to people who knew how a survey worked. By 1981, reporters were asking what would happen to the information once a cable company had collected it. A billing record and the answer to a poll had been given for different purposes. To a company interested in that household’s habits, both were potentially valuable.
Dallas Smythe, a Canadian political economist, published his account of the “audience commodity” the same year QUBE launched. Commercial broadcasters, he argued, assembled audiences they could sell to advertisers. The advertiser got access to attention. Two-way cable let the seller know something more about what that attention was worth. QUBE’s local programming proved expensive to maintain as Warner expanded. The idea that the box could watch back survived the shows that were built around it.
The smart television made responding easier by dispensing with the button. The Federal Trade Commission said Vizio began collecting second-by-second viewing information in 2014 on millions of sets without obtaining adequate consent. In 2017, Vizio agreed to a $2.2 million settlement with the FTC and New Jersey. Roku describes automatic content recognition on its televisions when enabled. The set recognizes programming from an antenna or from equipment connected to it. A viewer who would have had to tell a broadcaster what they watched could now disclose it in the course of watching.
Netflix used what it knew about the audience to change the menu itself. Even after it had picked a film to recommend, it could still choose what part of that film to show you. A poster that once had to be printed and sent to theaters had become a variable.
Facebook gathered its users’ activity in News Feed in 2006 and gradually made the order of the stories a matter of prediction. Twitter introduced a ranking change in 2016 that placed selected posts ahead of the latest ones. X’s For You feed pulls in posts from accounts you have never followed. TikTok starts recommending videos before you have followed anybody. According to TikTok, the system uses signals including what users watch through, skip, like, share, and seek out. An independent audit by Maximilian Boeker and Aleksandra Urman found that following, liking, and viewing choices on test accounts affected subsequent TikTok recommendations.
It used to be easy to tell when you were sending something back to the station. You had to make the call or press the button. On TikTok, even a pause can register. Someone who stops because a video is offensive or baffling has also supplied evidence that it held their attention. The next set of choices arrives through a feed that has learned from that response. Watching the next clip gives the service another signal, drawn from a choice it had already rearranged.
It is tempting to explain the two minutes lost to a feed by saying “dopamine,” as if the company needed to know precisely what was going on in your head. For its purposes, it may be enough to learn when you waited for the payoff. A close-up of something coming out of somebody’s ear gives a curious viewer a reason to wait. A clip that promises to expose an enemy or embarrass a politician offers another. Somebody deciding what to make tomorrow can learn from how long people sat through today’s version. The next batch of videos comes from people who have learned something from the system that was learning from its viewers.
There is a governing power in that, and it is closer to the experience of opening the app than any argument about whether software can control the mind. A feed can give a stranger an audience. It can also decide which subjects reach you with a thousand favorable reactions already visible around them. Those are ways of ordering access, exposure, and rewards. The power of a government rests on a different set of institutions, including the law. For a merchant, a political campaign, or a person making a living on a platform, the rules of placement can still end up deciding what is possible on any given day.
Two feed experiments give a measure of that power. In research published in 2023, Andrew Guess and colleagues assigned some U.S. Facebook and Instagram users reverse-chronological feeds during the 2020 election. What users encountered changed. They spent less time on the services. The researchers found no detectable change in the political attitudes they measured, including issue or affective polarization. In a different experiment, published in 2026, Germain Gauthier and colleagues found that U.S. users who switched from a chronological X feed to its ranked feed in 2023 moved in a conservative direction on a set of policy and current-events questions. They followed more conservative accounts. Party identification did not measurably change, and switching the ranked feed off for users already on it did not produce an observable reversal. One reason the return trip could be different was plain in their following lists. Switching back left the new follows in place.
An earwax camera only has to catch somebody’s attention once to give the feed something to work with. TikTok Shop lets sellers open affiliate collaborations to eligible creators or invite particular creators to make sales content. An affiliate earns a commission on purchases attributed to a clip. One merchant may end up with a small crowd of people demonstrating the same gadget in their own styles. To a seller, those variations are a way to find out who can move the product and which pitch works. A person who encounters several of them on different days sees what looks like an object catching on in several corners of the world at once.
China’s livestream-commerce business supplies a view of the work behind the performance. On Douyin, TikTok’s Chinese counterpart, a stream can come from a company with sets, hosts, support staff, and merchandise waiting just off camera. Rest of World followed an operation whose workers in China trained presenters selling through livestreams in the United States. The apparent ease with which someone can talk about a product for an hour has an employer, a shift, and a person who knows what the camera is supposed to be showing.
After the call about the marriage, the reporter was added to a WeChat group with more than a hundred people who took jobs like it. The host and her assistant posted roles, storylines, and time slots there. The same investigation spoke to a Hangzhou operator who described agencies buying annual packages of scripted callers, comments, and product ratings. Some callers were paid to break up the silence on a stream. Others gave a host an unexpected problem to solve or a reason to sell something. A client who bought the caller and the comments could supply both the drama and the people who seemed to have tuned in for it.
Some of this work is as public as the product on the table. A professional shopping host can honestly describe the thing on the table, and an affiliate can disclose a real commission. Commercial television was full of people hired to sound comfortable in front of a camera. The deception enters when a paid relationship is concealed or a caller supplied by the show appears to be the public calling in. In the United States, the principle goes back to broadcast sponsorship and the payola provisions Congress tightened in 1960. A station could air a record someone had paid it to play if it identified the sponsorship. The FTC applied the same basic distinction to material ties between advertisers and online endorsers. The listener had an interest in knowing what brought the recommendation to air.
The dispute over who controls American TikTok came to a corporate arrangement in January 2026. TikTok USDS Joint Venture LLC was assigned responsibility for securing, testing, and retraining the recommendation system with U.S. user data and for U.S. moderation decisions. Oracle, Silver Lake, and the Abu Dhabi investor MGX each took 15 percent. ByteDance retained 19.9 percent. According to the announcement, TikTok’s global organization, through its U.S. entities, retained commercial functions including e-commerce, advertising, and marketing. The distinction is worth making because people talk about control of TikTok as if it were a single lever. The company with responsibility for the American recommendation system and the operations making money from a sale can be organized differently and still meet in the same clip.
There is a market for the response as well as the performance. In 2017 Oobah Butler used reviews from friends to get a restaurant that did not exist, The Shed at Dulwich, to the top of TripAdvisor’s London rankings. The requests for tables came from real people. What they took for the prior experience of diners had been supplied by the man who wanted them to call. The people ringing up sincerely wanted dinner. They had no reason to suspect the reputation had been manufactured.
Sherry He, Brett Hollenbeck, and Davide Proserpio studied the everyday version of the same business. Amazon sellers recruited fake reviewers in private Facebook groups. Sellers offered to reimburse purchases in exchange for reviews. The researchers followed the products’ ratings and sales ranks. Review campaigns improved both in the short run, and ratings deteriorated after the recruiting stopped. In one analysis average ratings were already below the starting level four weeks later. An operator in this business can sell something a buyer will soon have a reason to buy again. A company with the means to keep renewing its reputation can appear to have a stable relationship with the public even while it keeps borrowing pieces of it.
The same arithmetic applies to an account’s followers. In 2019 the FTC alleged that Devumi had fulfilled more than 58,000 orders for phony Twitter followers. New York investigators said some of the accounts appropriated actual people’s identities. In a later case the FTC alleged that the supplement company Bountiful had attached reviews for existing Amazon products to the listings for newer ones. The ratings had all the substance their customers put into them. The listing gave them a new object.
The MusicLab experiments followed what happened after the first people arrived. Matthew Salganik, Peter Dodds, and Duncan Watts gave different groups the same collection of songs by unfamiliar bands. For some listeners, the page showed how many people in their group had downloaded each song. Groups who could see that count produced bigger winners, and the biggest winner varied more from group to group. In a later experiment Salganik and Watts inverted the displayed rankings after establishing them. Some of the false order came true as people picked songs now made to look popular. Some strong songs recovered. The inversion also reduced total downloads. Selling the appearance of enthusiasm can change where real enthusiasm goes, and it can leave the people making their choice worse served by the list.
On a social-news site, Lev Muchnik, Sinan Aral, and Sean Taylor found that a single planted positive vote made the next vote more likely to be positive. Users tended to correct an initial negative vote. The people who followed the positive vote were making up their own minds in the ordinary way. They had been given one piece of information about the people who had apparently made the decision before them. Every real vote after it gave the next visitor a better reason to stop asking where the first one had come from.
Once I started looking at how a product could appear to have many independent advocates, a certain kind of political account on X looked different to me. An account posts videos. The same group appears underneath them, commenting or reposting. One of those accounts uploads the clip again. It reaches another cluster of people who may have missed the first one. Anybody who has been around a campaign or a fandom has seen ordinary versions of this. A handful of accounts repeating a clip in different places is a cheap way to make it look as though strangers found it independently.
In research around the 2024 U.S. election, Luca Luceri and his colleagues identified thirty-four coordinated accounts that repeatedly circulated material sympathetic to Trump and sent readers toward some of the same sites and a YouTube channel. They traced overlaps between accounts and profiles. The research did not establish who ran or financed them. A 2020 investigation had supplied a better view of the labor behind a different operation. Facebook linked accounts that posed as ordinary Americans to Rally Forge, an Arizona marketing firm working for Turning Point USA and another group. The pro-Trump campaign employed teenagers, according to The Washington Post. They made arguments, wrote comments, and came back the next day to do it again. On the screen that labor looked like the ordinary work of keeping up a profile and talking politics with strangers.
That year Mike Bloomberg’s Democratic campaign hired people to promote him in their own social circles. It set out to recruit more than five hundred organizers in California at $2,500 a month. Campaign language appeared repeatedly across personal accounts, and Twitter took action against some of the activity. The campaign maintained that staff could speak in their own voices and did not need to label their posts as sponsored. A person hearing from a friend might know the author of the post quite well and have no idea there was a payroll behind it. In Azerbaijan that same year, Meta found a network tied to the ruling party’s youth organization in which much of the reaction to the comments appeared to come from within the network. Seeing a conversation respond to itself is a different experience from seeing campaign literature arrive in your mailbox. There are other people apparently right there for you to agree with.
X gave users a new clue about who was speaking when it started displaying estimated account locations in 2025. BBC Verify reported that X showed an account presenting itself as an Ivanka Trump news account as based in Nigeria. Wired found examples of Trump sharing content from accounts X located overseas. An account can have an international audience, and the location label is an estimate open to error. The relevant question is how the account presents itself. A U.S. political identity in a profile can make a clip look like a dispatch from inside American political life. The location label gave people something new to check. It left open who was paying for what, or whether anyone was.
X and its predecessor Twitter also offer something a political operation could once get only from an editor or broadcaster. The official trends list presents a topic as something large numbers of strangers have just begun to discuss. Tuğrulcan Elmas and colleagues found that coordinated posting in 2019 could push subjects onto Twitter’s official list of trends even when much of the original activity was subsequently deleted. In the period they studied, they attributed at least one fifth of the top-ten global trends to the pattern they identified. The largest category in a separate analysis of manipulated Turkish trends was illicit advertising. Political and policy demands appeared there too. An official trend can put a subject before someone who has no connection to the people who first promoted it. If that person joins in, the next indication of interest may be genuine.
The economics of manufacturing support look different from the seller’s side. It is hard to know whether the customer who buys the ear camera would have made the purchase without the video. At least the shop can record a transaction. A political operator may be rewarded for a different thing, including how much attention a clip receives or whether a more prominent account takes it up. If four apparent strangers vouch for it and a fifth decides to repeat it sincerely, a client can point to the fifth person as evidence that the message is spreading. The fifth person’s enthusiasm can be perfectly real. It is the calculation that brought the first four into view that a reader can so easily miss.
An editor gives the operation another route to an audience. Josephine Lukito and her colleagues traced posts from Russia’s Internet Research Agency into U.S. news stories, where some appeared as examples of normal public reaction. That was one place where an invented personality did the work of a real person. Work on what those posts did to voters is harder to interpret. Gregory Eady and colleagues estimated that one percent of the users in their Twitter study accounted for seventy percent of exposure to the operation’s posts. They found no meaningful relationship between that exposure and changes in the political attitudes or voting behavior they measured. What moved through a newsroom, what went before a voter, and what changed that voter’s choices have to be established separately.
In 2024 Reuters reported on a U.S. military campaign in which accounts posing as Filipinos had attacked Chinese health supplies and, later, the Sinovac vaccine. The apparent independent testimony of people in the country was something the operation thought worth manufacturing. The reporting established the operation’s existence and origin. How many people encountered it and what they did about vaccination are still unknown. A political advertiser would recognize both the reason for using a local voice and the difficulty of showing which decision that voice changed.
A parent told New York investigators that a comment in the Federal Communications Commission’s 2017 net-neutrality docket used their ten-year-old son’s name, his real email, and the family’s address. A woman found a comment under the name of her husband, who had died of cancer the year before. The commission took in more than 22 million comments. The New York attorney general estimated that nearly 18 million were fabricated. Contractors working on a broadband-industry campaign used real people’s names without consent. On the other side, one student filed millions of comments under invented identities. The agency had a process for hearing what a rule meant to a person who had taken the time to speak. Somebody reading the headline count could instead come away with an impression of how many people on either side had bothered. The broadband group’s planning records show how closely it tracked the volume on both sides.
When Ukrainian drones hit Yandex data centers in October 2026, the online reaction demonstrated why those distinctions are easy to lose. Charts showing an abrupt decline in traffic from Yandex were taken as evidence that Russian political bots, including some supposedly sustaining the MAGA audience, had disappeared. The strikes and disruptions were real. Cloudflare found that the fall from Yandex’s search network mostly involved recognized crawlers that fetch pages and images for its search index. The company also saw some malicious automated traffic shift hosts, and no unexpected change in the social-media activity it could observe. The chart counted traffic to servers. Its users had read it as a count of people arguing on the internet. There was a real loss of traffic and a real appetite for seeing that loss reflected in the feed. The evidence did not connect the two.
That helps explain how the original Dead Internet Theory got such a hold. In January 2021, a poster using the name IlluminatiPirate laid out a version on Agora Road’s Macintosh Cafe, invoking older imageboard discussions and treating the resemblance among online voices as evidence of deliberate replacement. Kaitlyn Tiffany took the idea into a wider public in The Atlantic that August. Generative media has since given people more occasions to suspect there was no one on the other end. Imperva has reported automated traffic exceeding half the requests in its samples. A crawler making thousands of requests and a person reading one page both contribute to the tally. Counting those requests tells us something about demand on servers. It offers little help in deciding who sits behind a particular account or whether five accounts have independent reasons to talk. Adrian Rauchfleisch and Jonas Kaiser found that even Botometer, a tool designed to estimate account automation, could give a live person a high bot score and an automated account a low one. Results varied over time and by language.
Kempelen’s Mechanical Turk, exhibited in 1770, worked by hiding a chess player inside what looked like an automated cabinet. Amazon borrowed its name for a marketplace that performed work through a computer interface by giving the tasks to people. The old trick remains relevant on a feed where a paid worker can be indistinguishable from a bot to somebody looking for patterns. The Chinese caller reading a script was supplying something different again. Her performance had been arranged in advance, and the interest of the people taking her story seriously was their own.
I was trying to remember a story about a communist prison in which a group had to bear the consequences if one person refused to go along. The version I remembered, with a fixed number of prisoners automatically punished if one objected, is difficult to substantiate. There are documented histories of collective pressure. In parts of Stalin’s Gulag, brigade quotas and rations put the prisoners’ welfare at stake in one another’s work. China’s reform-through-labor system organized labor and political education in groups that could press individuals to criticize themselves and each other. Under that arrangement, a prisoner’s peers had to bear the cost of dissent with them.
.jpg){kind=link}
Maoist criticism and self-criticism also took place outside prison. In some land-reform campaigns villagers were mobilized to speak publicly about grievances against landlords. During the Cultural Revolution, a struggle session could subject a teacher, official, or other target to public accusation, humiliation, and physical assault. People in attendance supplied visible assent through denunciation or silence, under conditions in which being seen to object could be dangerous. Robert Jay Lifton, who interviewed people who had gone through Chinese thought reform, called one part of the process “milieu control.” The people setting the terms of the meeting could also have a hand in what kinds of conversation were possible outside it. Even the terms available for understanding one’s own doubts could come from the people running the meetings.
A prison or a political persecution campaign could impose hunger, confinement, and physical violence. Online, the penalties this comparison helps illuminate are being shunned, mobbed, or made to seem alone. Group penalties gave prisoners an incentive to do a guard’s work. In a feed, the apparent opinion of a crowd supplies an invitation to enforce a norm. People join a pile-on because they think the group expects it, stay out because it looks unwinnable, or finally speak when someone else has gone first. Most of those decisions are made without knowing how many people actually hold the opinion that started it. In Myanmar, UN investigators traced a network of seemingly independent Facebook pages spreading anti-Rohingya hate to the same military that was carrying out its 2017 campaign of mass violence.
Solomon Asch was able to make one part of this visible with a set of lines on cards. Participants in his experiments saw which lines matched, then heard a room of people give the same wrong answer. A dissenting ally sharply reduced the tendency to give in. Deborah Prentice and Dale Miller later found that Princeton students systematically thought their peers were more comfortable with campus drinking than they themselves were. No operator had to manufacture that impression. Elisabeth Noelle-Neumann’s spiral of silence proposes one way it can sustain itself. People who fear isolation become less willing to voice a view they think is unpopular, making it seem rarer to everyone else. The important thing on a screen is that other people’s silence already has an interpretation ready for it. The comments you see tell you almost nothing about how many people read them and decided to keep quiet.
Timur Kuran’s work on preference falsification helps explain why this can be unstable. The gap between a private opinion and the one it is safe to express sometimes widens over years. The gap between a neighbor’s private opinion and what the first person thinks it is may widen too. When enough people finally speak, the appearance of a consensus can dissolve faster than the underlying views could possibly have changed. Kuran applied the idea to the surprise collapse of Eastern European communist regimes in 1989. A feed full of real people facing much smaller costs might still contain some of the same uncertainty. It is hard to tell which way the public will move when its opinions are being inferred from the people volunteering to be seen.
William J. Brady and colleagues tested a related point on Bluesky around the 2024 U.S. election. Their experimental engagement-ranked feed exposed readers to more moral outrage and antagonism. It made participants perceive greater partisan hostility and read some discussion norms less accurately, with no detectable change in their own engagement behavior. An alternative ranking that reduced the influence of extreme users improved some of those judgments. Making a different selection of actual people easy to see was enough to make the group look different.
The older literature had names for parts of this. Walter Lippmann’s Public Opinion called the picture pieced together from reports of events we never see a “pseudo-environment.” His point was that we act within the picture and our actions have consequences outside it. The Soviet researcher Vladimir Lefebvre later took up another version in his work on reflexive control, developed in the 1960s and subsequently applied in military thinking. A person choosing a course of action reasons from what they believe the others involved know and intend. If someone can arrange the evidence from which they infer those intentions, the choice can be made voluntarily and still fall within terms supplied by someone else. Commercial recommendation grew out of its own history of advertising, information retrieval, and audience measurement. The resemblance draws attention to what somebody sees before they make a choice that still feels like their own.
It is possible to make a group appear unanimous by supplying agreement. It is also possible to put so much else in front of people that they miss the conversation that might have unsettled it. Gary King, Jennifer Pan, and Margaret Roberts studied a Chinese government posting operation in which much of the material celebrated the country or its achievements rather than taking on particular critics. Roberts used “flooding” for this kind of approach to keeping attention elsewhere. Burying a subject makes it harder for the people interested in it to tell how many others share that interest. The feed can arrive at a similar result through competition among publishers with no shared plan. Everybody who gets paid to keep somebody watching has a reason to supply another thing to watch.
Once the guards set the penalty, the people in the room had to anticipate it. In the political feed, what the surrounding people seem to believe can start to determine who feels comfortable answering. Each person deciding whether to speak reads the relative absence of everyone else as a signal. The political clip has then done a different kind of work from the earwax camera. Somebody may keep watching the camera out of curiosity. Somebody encountering what seems to be a unanimous political audience may decide that the safe thing is to stop being seen.
The viewing habits of a person and the apparent judgments of the people around them now meet in the same piece of software. One choice is where to send a video. Another is what a viewer will know about the people who got there first. Distributing the video and manufacturing its reception can be separate lines of work, and a company can make money at either one.
When polished AI videos and songs started turning up in my feed, I wondered who was paying to make them. Some of the videos went on for three, four, five minutes. I could scroll past in a few seconds. Somebody had still spent the time or money. Viewed as a single thing made for a single person, one of those videos looks expensive. A working studio solves a different problem by dividing a film into short shots, reusing a setting or character reference, generating less demanding shots with a cheaper model, and cutting around failed attempts. Voice, music, graphics, and editing supply continuity between pieces of footage that never had any on their own. Runway lists models that charge by the second and deliver only a few seconds per generation. Alibaba’s open-weight Wan2.1 offers a different cost structure, with hardware, time, and labor still to pay for. A subscriber judging the price from a few failed shots can be looking at a very different budget from a shop reusing the same characters and techniques across dozens of videos. Some people also put hours into making something strange because they like making it. A polished clip on its own gives very little away about who paid for it.
That answers part of the question I had about the money. The cost of one ambitious video is a different question from the cost of keeping a channel stocked. A series carries techniques and earlier work forward. One good-looking upload may earn its maker an audience that will still be there for the next one. The larger investment might look foolish on the proceeds of a single clip and sensible if the producer is building an account that can draw people back. The odd video that drifts past your phone is one item in an inventory you may never see together.
The incentive to keep the shelves full was already visible on YouTube in 2017. Parents and journalists encountered sprawling runs of children’s videos pieced together from familiar characters, search terms, and bits of stories that might have made sense somewhere else. James Bridle described a production environment in which the ingredients could keep turning up in new combinations because a recommendation might put the next one in front of a child. YouTube responded with tighter rules for what could appear or carry advertising. The central advantage of the arrangement had already become evident. A producer can make a business out of supplying the kinds of things a system is disposed to serve.
Renée DiResta and Josh Goldstein traced a small part of this business on Facebook. They studied 125 pages that had each posted at least fifty AI-generated images. Some used the pictures to collect an audience. Others directed people toward sites or offers and in some cases displayed signs of scams or page takeovers. Facebook recommended the pages to people who had never decided to follow them. Producing an image had become cheaper. Having a way for one of those images to come before a stranger was the other half of the business. The producer could try many pictures before any one of them attracted enough attention to pay.
Liz Pelly’s reporting on Spotify gives an account of an industry doing this with human composers and recording artists. Through the company’s reported Perfect Fit Content program, certain suppliers had a route to mood playlists for work credited to many different names. Spotify had publicly denied creating fake artists or using them to escape royalties. The arrangements Pelly describes show how much of a seemingly independent supply of music a limited set of intermediaries can assemble. In 2025 The Velvet Sundown went a step further in public. It came with band photographs, tracks, and enough listeners to generate stories about whether the band existed. The account eventually described the project as synthetic music guided by people. Listeners could find the work because they liked it or because they wanted to find out what was going on. The discovery that part of the story had been manufactured could give the music an additional route to a genuine audience.
Others had figured out how to make money without waiting for that last part. Michael Smith used a catalog of AI-generated songs and streams from automated accounts to collect more than $8 million in royalties, according to the federal account at sentencing. He pleaded guilty to conspiracy to commit wire fraud in March 2026 and was sentenced in October to eighteen months in prison and two years of supervised release, with roughly $8.09 million in forfeiture. The stream was a tally used to decide who got paid. Smith needed the service to treat the tally as evidence of listening. Whether anybody found the music worth putting on made little difference to that transaction.
There is a short passage from selling a fabricated response to selling a fabricated part of the world. In Event of Moon Disaster, released online in 2020, used a genuine contingency speech written for Richard Nixon in 1969 and made it look and sound as if he had delivered it. The Apollo 11 astronauts came home. Nixon never gave that address. The MIT project was forthright about what it had created, and the archives hold the real memorandum. An isolated excerpt could travel without that explanation. It would still have a script from the White House, a recognizable face, and the sound of an American president. The usual signs that something belongs in the historical record could help put a fabricated event into it.
The lawyers Bobby Chesney and Danielle Citron call the opposite opportunity the “liar’s dividend.” A person who really was recorded can now appeal to what viewers know computers can do. There are two ways for the past to get harder to see. A fabricated record can acquire the appearance of a source. An authentic record can acquire the appearance of something a person could have made this morning. In a feed there may be little distance between where a viewer finds the recording and where they find a crowd ready to tell them what to believe about it.
The film behind Netflix’s two pictures was the same. A service able to make a coherent program for each viewer and adjust it as they watched would have more to work with. It might make a comedy feel more familiar, give a favorite character more to do, or produce a fictional memory with the visual texture of the programs someone grew up watching. At present that is a possibility, with costs and technical difficulties that matter. The part about supplying an audience response has a much longer commercial history. There is no particular reason the business of making the program and the business of making it look as if other people love it should stay separate.
Lippmann’s reader could compare his paper with the same paper on his neighbor’s doorstep and see that they had been given the same picture. With a personalized feed, even that small act of checking has changed. You can show somebody the clip you saw. It is harder to put them in the position you were in when it appeared, after the things you had just watched, among the replies that happened to be put first. An exchange between two people may start from a disagreement about what everybody knows, when each has been learning what “everybody” thinks from a different view of the crowd.
On the old television, everyone watching the same show heard the clapping in the same places. In the feed, the people clapping get there before you do, and what you see depends on which of them the feed puts first. The FTC’s 2024 rule prohibits buying fake views or followers to materially misrepresent influence for commercial purposes if the buyer knew or should have known the numbers were fake, leaving purely political amplification outside the rule. The next time I see a cheap camera in somebody’s hand, I can look to see whether they’re being paid to sell it. The people who appear to have found it before I did are harder to account for.