In May 2024 I argued that social platforms had two ways out of the bot problem: verify every user against a national ID, or charge a small fee to join. The fee was the option I liked. Bot farms run on volume, volume runs on free accounts, so put a price on an account and the economics collapse.
The argument was clean. It was also testable, and by the time I wrote it, X had already run the test. It did not go the way I said it would.
Why the fee argument is so persuasive
Start with the sum, because the sum is genuinely convincing and it is why this idea will not die.
What a fee does to a bot farm, on paper
Accounts one budget buys per year, at each price point.
Bars are capped at 10,000 accounts so the paid tiers stay readable; the numbers are exact. $8/mo and $20/mo mirror X Premium’s public pricing. This is the theory only — what happened in practice is further down the post.
At zero, a spam operation is limited only by how fast it can register. At eight dollars a month, a hundred thousand dollar budget buys about a thousand accounts a year. That is a real constraint. If bots were purely a volume game, this would work.
X ran it, in two countries
In October 2023 X launched a test called Not A Bot in New Zealand and the Philippines. New accounts paid one dollar a year to post, like or reply. This is the exact policy I would go on to propose seven months later.
Two things are worth noting. X's own framing was that the fee makes abuse more expensive, not that it removes bots. And no before-and-after bot rate was ever published. The single cleanest natural experiment on this question produced no public result.
The correlation went the wrong way
The strongest evidence arrived in February 2025. Researchers examined 2.5 million election-related images posted to X between July and September 2024 and classified how many were AI-generated.
Then the finding that undoes my 2024 argument. The accounts spreading AI-generated images were more likely to be X Premium subscribers, and more likely to show automated behaviour. The people paying were disproportionately the ones doing the thing the payment was supposed to stop.
That is correlation, not proof the subscription caused it. But my claim required the opposite correlation to hold, and it does not. A fee does not separate humans from bots. It separates people who will pay from people who will not, and a spam operation with a profit motive is exactly the kind of actor that will pay.
Independent work presented at USENIX in 2023 reached a compatible conclusion from another direction: paid verification clustered around particular political and crypto-promotional communities, and did not establish that an account belonged to a real, accountable person.
How bad is it, actually
I claimed in 2024 that the feed was mostly bots and invoked Dead Internet Theory. When you go looking for measurements rather than vibes, the picture is smaller and stranger than that.
- X: about 12% of election images AI-generated in one large sample, but only 1.4% of text. No platform-wide bot percentage is published anywhere.
- Meta: roughly 4% of monthly active accounts across its apps are fake, by its own May 2026 estimate to a parliamentary committee. That is accounts, not posts, and not Instagram specifically.
- Reddit: machine-generated text was generally a small minority across 51 subreddits from 2022 to 2024, reaching around 9% in some communities in some months.
None of those is a majority. The internet is not mostly machines. So why does it feel like it is?
It is concentration, not volume
The same X study found that roughly 10% of the accounts spreading AI images accounted for about 80% of that spread. The synthetic content is not evenly smeared across the platform. It is produced by a small, hyperactive set of accounts and then amplified into exactly the places you look: replies under big posts, trending tags, the top of the feed.
That reframes the whole problem, and it explains why my fix was aimed at the wrong target. A per-account fee taxes the long tail of cheap accounts. The damage is being done by a concentrated head that is already willing to spend money, because spreading this content is its business.
The other option still costs what it cost
My second option was national-ID verification, and I want to keep the part of the 2024 post that still holds. Tying accounts to government identity would work better than a fee, and the price is anonymity. That price falls hardest on the people who need anonymity most: journalists, dissidents, whistleblowers, anyone posting from a country where the wrong opinion is a criminal matter.
I said then that this trade is bad. I still think so. What I would add now is that we have real evidence the cheap alternative does not work, which makes the pressure to reach for the expensive one much stronger than it was two years ago.
What I got wrong
The instinct was fine. Bots are an economics problem, and raising costs is a reasonable lever. The mistake was reasoning to a conclusion from a model of how spammers behave, and never checking whether anyone had measured it.
They had. X shipped the experiment a year before I proposed it, and the evidence since points the other way. That is the whole lesson: the arithmetic being persuasive is not evidence, and a clean argument is the easiest thing in the world to be confidently wrong with.