The crowd you're arguing with
isn't real.

The census

Nobody knows how many bots there are. The counts we have are staggering.

Start with the web itself. In 2024, automated traffic reached 51 percent of everything moving online — bots overtook people for the first time since measurement began. "Bad bots" — scrapers, fraud tools, fake-account software — alone made up 37 percent of all traffic, a share that has risen six years straight. The report's explanation for the surge: AI has made bots cheap to build and hard to spot.^[1]

On Facebook, the removal numbers tell their own story. The platform took down about 3.5 billion fake accounts in 2025 — and that is a routine year, not a purge. In several years Facebook has banned more fake accounts than it had monthly active users; the ratio peaked at 2.7 to 1 in 2019.^[2] Meta's last public estimate put fake accounts at roughly 5 percent of monthly active users — and it has since stopped disclosing Facebook's monthly active users at all, which makes the comparison impossible to update.^[2][5]

And removals only count the bots that got caught. On Twitter — now X — the company's filings long insisted fewer than 5 percent of accounts were fake, while independent researchers estimated 9 to 15 percent.^[6] Detection is an arms race, and the referee works for one of the teams.

The mechanism

How a few fake voices move real opinion

You don't learn what other people think from surveys. You learn it from the room — and for most of us, the room is now a feed. That is the vulnerability bots exploit. They don't need to persuade you of anything. They only need to change your sense of what everyone else believes, and normal human social instincts do the rest.

Put those two findings together and the threat comes into focus. It has never been necessary to fake half a platform. A few percent, placed early and aimed at the right people, is enough to bend what an entire network believes the consensus is.

Manufactured grassroots

Fake crowds, real consequences

This isn't theoretical. Manufactured consensus has been deployed — against elections, against regulators, against you — and documented each time only after the damage was done.

None of this is a rogue-actor story anymore. It's an industry. Oxford's Internet Institute found organized social-media manipulation campaigns in all 81 countries it surveyed in 2020 — up from 70 the year before — with bot accounts deployed in 57 of them, and governments in 48 countries hiring private "strategic communications" firms that have collected some $60 million for the work. The report's lead author put it flatly: misinformation "is now produced on an industrial scale."^[11] Meta itself reports removing more than 200 covert influence networks, originating in 68 countries, between 2017 and 2022.^[12] Two hundred that were caught.

2024–2026

The bots got a brain

Everything above describes the cheap, dumb generation of bots — copy-paste accounts that a careful reader could spot. Large language models removed that ceiling. A bot can now hold a conversation, argue in your dialect, and joke its way out of being accused of botting. And the price of a thousand convincing voices has collapsed.

Facebook's algorithm does the distribution. Stanford and Georgetown researchers tracking AI-generated spam pages on Facebook — the "Shrimp Jesus" wave — found hundreds of pages whose synthetic images racked up hundreds of millions of interactions, pushed by Facebook's own recommendation engine into the feeds of people who never followed them. Many users showed no sign of knowing the content was fake. The pages then funneled their audiences to ad farms and scams.^[13]

Influence is now a subscription service. In 2025, Anthropic disclosed a commercial "influence-as-a-service" operation that used its AI model, Claude, to run more than one hundred fake political personas on Facebook and X. The model itself decided when each persona should comment, like, or share, and wrote the replies in the target audience's language. The operation served multiple clients across four campaigns and engaged tens of thousands of real accounts. When users accused the personas of being bots, they answered with sarcasm — and kept going.^[14]

Whole conversations can be synthetic. OpenAI reported disrupting ten covert operations in a single three-month stretch of 2025. One China-linked network, "Sneer Review," generated comments across TikTok, X, Reddit, and Facebook — and then generated the replies to its own posts, staging entire threads of fake people agreeing with each other for a human audience of onlookers.^[15]

The real problem

Why "open discussion" can't happen on an outrage engine

The bots are only the third act of the problem. Walk through what a Facebook conversation actually is, and the idea of the feed as a "public square" comes apart piece by piece.

First, the referee is biased toward anger. The feed is not ordered by merit or by time; it is ordered by predicted engagement, and moral outrage is the most engaging thing humans produce. Across 27 studies and 4.8 million posts, each added moral-emotional word in a post increased its spread by about 13 percent — the original study measured 20.^[16] Facebook didn't just observe this; it weighted the angry reaction five times a like. The mechanics are on The Machine.

Second, the "public" was never the public. The most active quarter of American adults on Twitter produced 97 percent of the tweets.^[17] During the 2020 election, 2,107 people — a third of one percent of a 660,000-voter panel — spread 80 percent of the fake news, by hand, no automation required.^[18] What reads as "everyone is saying" is a tiny, unrepresentative, self-selected sliver — pre-sorted by an algorithm that favors its angriest members.

Third, the crowd is spiked. Onto that distorted sample, add everything documented above: bots that amplify junk in its first seconds, personas that argue back in fluent idiom, fake threads agreeing with themselves, and an influence industry operating in 81 countries. Every signal you might use to gauge what people actually think — likes, shares, comment sections, trending topics — is for sale.

And the house takes a cut either way. Engagement is revenue no matter who generates it. Internal documents reported by Reuters in late 2025 showed Meta projecting that scam ads, schemes, and banned goods would account for roughly 10 percent of its 2024 revenue — about $16 billion — while its platforms showed users some 15 million high-risk scam ads a day. Meta's internal policy: ban an advertiser only when its systems are at least 95 percent certain of fraud. Below that threshold, the penalty for probably being a scammer was a higher ad rate.^[19] A company that charges fraud a premium instead of showing it the door is not going to wage genuine war on fake engagement — fake engagement inflates the very numbers it sells.