The censor nobody elected wasn’t Claude: it was Meta’s automation, with no human safety net

WHAT HAPPENED, AND WHAT DIDN’T. On October 17, the No Kings coalition held its ‘Vote Early’ day of action. Organizers say it included nearly 3,000 events across all 50 states. At 1:30 p.m. New York time, its main Instagram account disappeared. According to the coalition, it became visible again around 3:00 p.m., but administrators did not regain access until 5:33 p.m. Meta has offered two explanations. The first, from spokesperson Andy Stone, blamed unusually high rates of posting and commenting. The next day he revised it: engineers had detected ‘a Claude tool reading profiles, posts, and hashtags’ at about four times the normal rate, in bursts. That pattern, Meta said, ‘resembled automated bot traffic.’ The company called it an error and apologized. Note what Meta has not said: which tool it was, who was using it, or why. Anthropic has made no public statement. Hence our thesis: the headline ‘AI censored a protest’ misdescribes what happened. Claude did not decide which speech could circulate. A Meta anti-bot system did, and nobody reviewed it in time.

HOW A MACHINE MODERATES. Two pieces need separating. The first is integrity systems, which watch behavior: request speed, odd logins, bot-like patterns. They act without reading what you say. That is what fired here. The second is content moderation, which judges what gets posted. Today it works in layers. Classifiers (models that assign a label such as ‘spam’ or ‘harassment,’ with a confidence score) filter the bulk. Increasingly, a large language model (LLM, the technology behind ChatGPT or Claude) receives the company’s policy rewritten as instructions and rules on the borderline cases. The most sensitive cases should go to a person. Meta is speeding up the shift. According to the Financial Times, LLMs already handle about half of content and ad reviews, with a goal of more than 90% for some categories. The company argues its models make 13% fewer errors than human reviewers and catch 10% more violations. Those figures come from its own tests, without external audit. According to the same paper, Meta employees warn the rollout is too fast and that the models still remove or demote harmless content.

DIFFERENT ERRORS: FEWER, BUT IN BULK. Our reading is that a machine does not fail the way a person does. A human moderator errs at random: a bad day, an unfamiliar context, a case they misread. A model is more consistent, because it applies the same rule the same way every time. But when the rule or the model is wrong, the error repeats thousands of times at once, with the same bias. Academic research is starting to measure this. A University of Queensland study (April 2026) found that LLMs judge hate speech differently depending on the political persona they are given, even though overall accuracy barely changes. A paper at the NDSS 2026 security conference found that political content varies most by country and language. The No Kings case adds a third, more mundane failure mode. A system tuned to catch bots punishes anyone who browses very fast. More and more people browse with AI agents, meaning assistants that read and act on the web on their behalf. So an organizer using an agent to track mentions mid-campaign looks a lot like a bot. That collision is not a one-off. It will grow.

‘THE AI DID IT’ CANNOT BE AN ALIBI. The chain of responsibility here is clear, even if the narrative muddies it. Meta set the threshold, Meta suspended the account, and Meta decided this account would have no human protection. Its own Oversight Board, which the company funds, said on October 20 that the error ‘should have been caught’ by its mistake-prevention systems. It also asked why No Kings accounts were not in ‘cross-check,’ the program that gives human review to important civic accounts. Naming Claude in the statement shifts attention to a third party that made no decision. We saw this with layoffs: ‘the algorithm decided’ sounds like technical fate, but someone chose to deploy it that way. We are not claiming Meta acted with political intent. There is no evidence of that, and the account was restored automatically within an hour. But the coalition, 14 state attorneys general led by Connecticut’s William Tong, and Senator Chris Van Hollen are all asking the same things: whether any human was involved, which other election accounts have been restricted, and how legitimate political activity is told apart from automated traffic. The attorneys general set an October 29 deadline. These are the right questions.

WHAT SHOULD CHANGE. In the short term, the risk is real and has a date. Days before an election, a platform that automates 90% of its decisions can silence vote organizers within minutes. And it can do so without anyone intending it. We propose three concrete demands. First: mandatory human protection for civic and electoral accounts during campaigns. Second: transparency on the automated system’s error rates, audited by third parties rather than only self-reported. Europe’s Digital Services Act already requires platforms to explain each decision and offer an appeal; the United States has nothing equivalent. Third: clear rules for legitimate AI-agent traffic, which will soon be normal. In the long term, we are optimistic. AI moderation can be better than today’s: faster, more consistent and, above all, able to explain each decision. A language model can give reasons for its verdict, which an opaque classifier cannot. Well governed, it would free thousands of moderators from the most traumatic job on the internet. But that only happens if responsibility is not automated along with the decision. The censor nobody elected is not the AI. It is the company that decides where not to put a person.