Content moderator — how we know

The only curve on this site whose later points can be checked against a number the employer is legally obliged to publish. It starts high because classifier-based takedown of spam, nudity and known-bad hashes was already routine before language models — the 2022 level is not an AI story, it is a decade of matching systems. The steep stretch is 2023 into 2024, when the same platforms began applying general-purpose models to the part that had needed reading rather than matching: context, intent, and the caption that changes what the image means. It flattens after 2025 for a reason the published reports show rather than a reason we assume — the automated share of enforcement actions is already above nine in ten at the largest platforms, so there is very little headroom left in the measure, and what remains is the contested residue that has resisted every increment so far. It does not reach the top because appeal against an automated decision is compulsory in the European Union, which puts a permanent person downstream of the machine.

Assessment date

2026-09-23

Basis of the task judgements

2 evidence-backed · 4 platform inference · 0 not enough evidence

Verified events

2

How we assess an occupation →