The cheapest and most effective governance is good community design — it just can’t scale, and it can’t last.
Community Watch: the community solves its own troubles
Trusted citizens are brought in to “sweep up” spam comments, and every action is public and auditable (who did it, when, why, whether anyone appealed, and the review outcome) — and in the first-version stage, the rules barred AI from removing anything automatically. Over a one-week trial, twenty-odd people swept up 339 comments in total (one “cleanup superhero” alone handled nearly 90% of them), 70% of the swept accounts came from four bot accounts, and there was not a single appeal. These public removal records are themselves the most reliable material for training the model, forming a “crowd labels → feed the model” loop.
339
Spam comments swept in one weekPorn ads 219 + abusive ads 20
Nearly 90%
Handled single-handedly by one “cleanup superhero”One person swept 304 of them
0
Appeals from the swept accounts70% of which came from four bot accounts
Public records aren’t about harsher punishment. They thread removal, appeal, and review into a traceable, out-in-the-open loop of community self-governance.
The watch team and the machine make each other stronger: a self-reinforcing cleanup loop
Crowd power isn’t a one-off cleanup — it’s a loop that reinforces itself: every public removal the neighborhood watch team makes is itself the most reliable training data for the model. Members hand their flags to the patrol bot; once it has learned, it takes over the clearest repeat spam, freeing people up to handle the harder-to-judge content. The more you sweep, the more accurate the model gets, and the less effort people spend — passed hand to hand like a relay, the whole system gets stronger the more it sweeps.
Round one: the crowd hands its public records to the machine
①
The crowd cleans up.The neighborhood watch team removes spam comments one by one — every action is public and auditable. This is where the loop begins.
②
Labeled data accumulates.Every public removal record is itself the most reliable training data for the model; these records are gathered on a public bulletin board and enter moderator review.
③
Train / calibrate the model.The collected spam patterns are fed into a new model — the machine leveled up: v1 → v2 — and it sweeps more accurately.
④
The bot cleans up too.A stronger patrol bot takes over more of the clear-cut spam in one pass, v2 → v3, freeing people from repetitive labor.
⑤
People are freed up, back to ①.Members go after the harder-to-judge content; the new patterns they surface are handed off to the machine again. The more you sweep, the stronger the model.
Comments use a very small model (multilingual-e5-small, about 100 million parameters), with a false-positive rate of just 0.33%; articles are then calibrated once more by an LLM A large AI model like ChatGPT or Claude that can read, write, and understand language. , pushing false positives down more than 10-fold. To avoid over-sweeping, the bot also has a brake: the same spam must appear three times over and match a spam-ring (Ring) pattern before it acts.
The coast-guard bot: a machine on the team, but with the brakes tied on
It patrols automatically, reusing the comment model to detect spam. Its actions come in two tiers (a score ≥0.55 files a report, and only ≥0.80 removes — the latter off by default), plus one more “overturn-rate” brake: should the watch team’s recent “overturn rate” exceed 10%, the bot simply goes on strike. In effect, it turns human review into an automatic, real-time safety valve.
≥ 0.55
Threshold to file a report≥ 0.80 to remove, and off by default
> 10%
The moment the overturn rate crosses the line, the bot stops workingHuman review = an automated safety valve
A history of failures: why handing verdicts to ordinary people is hard
Community litigation (2019) foundered on bloc-voting cliques and ballot-stuffing by sock-puppet accounts; the architect jury (2021) folded in under a year — first because “the expertise a verdict requires got overlooked,” and second because the architects, too, took sides and formed factions: “What ate up the most of the platform’s time wasn’t the cases themselves, but the squabbling and mediation inside the jury.” Even so, the team held on to its value: “Letting the community decide the community’s own affairs is still precious — you just have to think carefully about what system will let it last.”
“What ate up the most of the platform’s time wasn’t the cases themselves, but the squabbling and mediation inside the jury.”
Further reading: How users see governance — voices U1–U4 Hack the Platform: small remixable experiments