Content moderator — tasks, one by one

Working the queue

One very large platform publishes the split because the law makes it, and the number is not close: in the six months to 30 June 2026 it removed around 104 million pieces of content in the European Union and states that automated systems actioned 94.1% of it without human review. Read the denominator carefully — that is a share of enforcement actions, not a share of the working day. What it does establish is that the default path through this task no longer passes a person, and that the human queue is now the exception route rather than the main one.

A share of actions is not a share of jobs, and this platform does not publish moderator headcount alongside it. It also says nothing about the hard half: an automated system that clears 94% of a queue may be clearing the easy 94%, leaving a residue that is slower and heavier per item than the old average. Nothing here measures that, and anyone claiming the remaining work is proportionally smaller is going beyond the document.

The call the rule does not decide

This is the residue the automated path leaves behind, and it is defined by needing context the item does not carry: who is speaking, to whom, about what, in which week. Platforms keep escalation tiers and policy specialists for exactly this, and the published rules themselves are written with exceptions that require somebody to recognise the exception.

No platform publishes how many items reach this route, and that share is exactly what would decide whether this is a shrinking specialism or a growing one. A judgement that a task is human-led is a judgement about what the work requires, not a prediction that the headcount holds.

Reviewing an appeal

Appeal is the route by which an automated decision gets a person, and European law makes the route compulsory for the largest platforms rather than optional. That places a permanent human step downstream of an automated one, which is a different shape from a task simply being automated: the volume of this work is set by the error rate of the system above it.

Compulsory is not the same as staffed, and a right to appeal says nothing about how long the queue is or who is in it. This page has no measurement of appeal volumes or of how many are upheld.

Turning a value into an enforceable line

The rule is the thing everything else executes, and writing one is a drafting problem with a measurement attached: a policy that two trained people read differently cannot be enforced consistently by either people or software. Automating the enforcement raises the value of this work rather than lowering it, because an inconsistent rule now produces inconsistent results at machine scale.

Nothing on this page measures how many people do this at any platform, and it is plausibly a small number even where the queue is large. A task being hard to automate does not make it a common job.

Correcting the system that replaced the queue

Work that exists because of the automation, not despite it: a system actioning the overwhelming majority of a queue needs a supply of corrected examples and someone who can trace a wrong outcome back to either the rule or the model. This is the same shape as the reviewer roles that appear wherever a generative system reaches production, and it is the part of this occupation that did not exist a decade ago.

New work is not necessarily equivalent work: this page holds no evidence about how many such roles exist, what they pay, or whether they are offered to the people whose queue was automated. Naming a task that appeared is not a claim that it absorbs the people displaced.

Looking at the worst of it

Automated detection removes volume from this task and does not remove the task: a classifier flags, and the cases where a person must confirm are disproportionately the severe ones. So the direction of travel here is that the share of a moderator's remaining day spent on the worst material goes up rather than down, which is the opposite of what a headline automation rate suggests.

That direction is an inference from the shape of the split, not a measurement: no platform publishes what proportion of human-reviewed items are severe, and none is obliged to.