The comment you never see

Every large platform decides, thousands of times a second, what you will not see. This is usually described as moderation, which makes it sound like a librarian removing graffiti. The scale makes it something else.

A human reviewer looking at a reported post has, on average, a few seconds. In that time they must judge context they do not have: whether a word is an insult or a greeting between friends, whether an image is violence or a news report about violence, whether a sentence is a threat or a quotation of a threat. Reviewers in one documented case were expected to reach a decision in under thirty seconds, across languages they did not speak natively.

Automated systems take most of the load, and they are good at the clear cases and poor at the rest. They are particularly poor at irony, at minority languages with little training data, and at anything where meaning depends on who is speaking. A sentence that is abuse from a stranger may be affection between two friends, and the system cannot tell which it is looking at.

The result is two errors running at once, in opposite directions. Material that should come down stays up, and material that should stay up comes down, and the second kind is invisible: the person whose post vanished rarely learns why, and nobody else knows it existed.

There is no version of this with no errors. The honest questions are which errors a platform chooses to make, who they fall on, and whether anyone can appeal — and those are questions about policy, not about technology.