The grassroots that weren't there

In 2017, the FCC asked the public what it thought about net neutrality. The public answered — more than 22 million times. On paper, it was one of the largest outpourings of civic input in the agency's history.

Most of it was fake.

When the New York Attorney General finished digging, four years later, the number landed like a brick: nearly 18 million of those comments — about 80 percent — were fabricated.[1]

Some of it came from industry. An organized campaign generated more than 8.5 million fake comments backing repeal — most of them impersonating real Americans, stolen names and addresses stapled to opinions those people never held.[1] Three of the firms behind that machinery later paid more than $4.4 million to settle.[1]

And here's the part that should keep everyone honest: the other side cheated too. One 19-year-old college student, working alone, pushed 9.3 million comments the other way — for net neutrality — through automated software and fake names.[1] Pick your villain. The tooling doesn't care about your politics.

Astroturf isn't a left problem or a right problem. It's a process problem — and the process was already losing.

The friction was the feature

Nobody designed this on purpose, but public comment used to be protected by friction. To flood a docket with fake support you needed money, or scripts, or a room full of people copying and pasting. And agencies could often catch it — the same paragraph showing up 3,000 times is not subtle.

That friction was doing quiet work. It meant that, most of the time, a comment roughly corresponded to a person who bothered to write it.

Then the cost went to zero

Generative AI didn't invent the fake comment. It deleted the price.

A 2017-style campaign cost millions and left fingerprints — identical text, clumsy duplication. A 2026-style campaign costs about the price of an API call and produces a million unique messages, each one a little different, each reading like a real, slightly annoyed constituent.

And we're not guessing about whether that works. In a field experiment, researchers at Cornell sent more than 32,000 messages — some written by people, some by GPT-3 — to over 7,000 state legislators. Then they watched who wrote back.

Human letters got a reply 17.3 percent of the time. The AI letters? 15.4 percent. A two-point gap. On a couple of issues the machine actually did better. The offices reading their own constituent mail could not reliably tell a person from a program.[2]

Why "count the comments" breaks

Most institutions still treat public input as a tally. More comments on one side than the other, and that side has "the people" behind it. That instinct made sense back when faking the people was expensive.

It doesn't make sense now. When volume is free, counting volume just rewards whoever wrote the better script. A process that weighs input by quantity is, in effect, running an auction — and the currency is compute, not citizenship.

Washington has noticed, a little. In 2024 the House passed the Comment Integrity and Management Act, built on one quiet idea: make sure each comment comes from a real person.[3] Human verification, not vibes.

The only fix is upstream

You can't moderate your way out of this after the fact. Once a million plausible comments are in the pile, no detector sorts them cleanly — and every false positive silences a real person who actually wrote in.

So the fix has to happen before the input counts at all. Not "is this comment well-written," but "is there a verified human behind it" — and then weighting what they say by that, instead of by raw volume. It's less satisfying than a big number. It's also the only version that survives contact with cheap AI.

That's the uncomfortable design problem under every civic platform built after 2022, ours included: the point was never to collect the most voices. It was to make sure the voices belong to someone.

The 22 million comments weren't a triumph of participation. They were a preview. The question isn't whether the bots show up — they already have. It's whether we keep counting them.