LO
Get told when a verified record lands on this occupation →
Content moderator
Decides whether a piece of content stays up, against a written rule — the one occupation on this site where the employer is legally obliged to publish how much of that deciding is done without a person.
Find your platform's most recent statutory transparency report and read what it publishes about automated versus human review. If the automated share is published and high, the queue you are in is the part being measured away — and the report will usually also name which teams the exceptions route to.
This is not a probability of losing your job. It combines how much of the role's task load is exposed to automation with how far adoption has actually gone — useful for comparing occupations on one consistent basis, and for nothing else.
Written for people who apply a platform's written rules to user content at scale: the review queue, appeals, and the policy and quality work around them. It does not cover the platform's lawyers, its law-enforcement liaison, or the engineers who build the classifiers — those are separate jobs with separate task mixes. The evidence here leans heavily on very large platforms operating in the European Union, because that is where the law requires the numbers to be published; a forum with two moderators and no statutory duty is a different situation, and nothing on this page is measured there. One consequence of that shape is worth stating up front: the published rates describe enforcement actions, not working hours, so a high automation rate does not by itself say how many people are still employed.
What is actually changing#
The unit of analysis is the task, not the job title. A role is not replaced — its task mix shifts.
Is this your job? Say so and this page narrows to your share of it.
A job title is a bundle of tasks bought together, and no two people hold the same bundle. Nothing is sent anywhere — it stays in this browser.
Read all 6 tasks in full — direction, reasoning and limits →
Recent changes#
The company publishing its own figures because the Digital Services Act requires a report every six months — which is what makes this measurable at all, and why the numbers exist for platforms in the European Union and almost nowhere else. The period is 1 January to 30 June 2026 and the post is dated 31 August 2026. What it states: around 104 million pieces of content removed for violating the company policies, counted across video, live streams, ads, product listings and comments; and that automated systems actioned 94.1 percent of violating content without human review. Three limits, and the first decides how the number may be used. The denominator is enforcement actions, not working hours and not staff: a platform can action almost everything automatically and still employ many people, because the cases a person still sees are the ones the system could not settle, and those are slower per item. The post publishes no moderator headcount beside the rate, so nothing here supports a claim about employment in either direction. And it is one platform in one jurisdiction; the report also notes that the period includes the launch of its shop in eight further member states, so the content mix inside that 104 million is not constant across the half-year either.
An employer has put it into production. Can move the baseline — weighted by scale and how similar the setting is.
What this means for you#
The entry job people describe when they say content moderation — clearing a queue of clear violations — is the part with a published automation rate above ninety per cent at the largest platforms. Going in through that door means going in through the narrowest one. The parts that are not narrowing are the ones that need judgement about context, the ones that need someone who can write a rule other people can apply, and the ones that exist to correct the automated system. All three are reachable from moderation work, and none of them is reachable without having done it.
Your leverage is that you know where the written rule and the real case diverge, and that is exactly what neither a classifier nor a policy document contains. Two concrete moves: be the person who can produce the corrected examples a model is trained on, and be the person who can say whether a bad outcome came from the rule or the model. Both are visible inside the company in a way that queue throughput is not.
Your options#
Four directions, each with its real constraints and one thing you can test this week. Continuing as you are is a legitimate choice — it just has to be a chosen one.
Move to the side that writes the rule
Policy is where the judgement lives once enforcement is automated, and it is staffed disproportionately by people who worked the queue, because writing an enforceable line requires knowing the cases that break it. It is also the work that does not shrink when the automation rate rises.
There are far fewer policy seats than queue seats, and many sit in a different country from the queue. The move usually requires writing in the platform's working language to a standard the queue never tested.
Find your platform's most recent statutory transparency report and read what it publishes about automated versus human review. If the automated share is published and high, the queue you are in is the part being measured away — and the report will usually also name which teams the exceptions route to.
Become the person who corrects the classifier
A system that actions the overwhelming majority of a queue needs a steady supply of corrected cases and someone who can tell a rule problem from a model problem. That role exists because of the automation, and the people best placed to do it are the ones who can say what the right answer was.
It usually sits with a data or engineering team and is graded on their ladder, which can mean a slower first year. It also concentrates exposure: the cases sent for correction are the contested ones.
Take ten decisions you personally overturned in the last month and write, for each, whether the rule was wrong or the model was. If you can do that reliably, you already have the skill the role is hiring for; if you cannot, that gap is the thing to close.
Stay, and own a language or a region
Automated detection is uneven across languages, and the unevenness is structural rather than temporary: the training data exists where the users are most numerous. A moderator who is the reliable judgement in a language the system handles badly is holding the part of the queue that cannot be cleared automatically.
It is a bet on a gap closing slowly, and the platform decides when to invest in that language. Being the only person who can judge a language is also a single point of failure, which is uncomfortable in both directions.
Check whether your platform's transparency report breaks moderation down by language, and whether yours appears. A language that is not reported separately is usually one where the automated tooling is weakest — which is where the human judgement is worth most.
Common questions#
At the largest platforms operating in the European Union the number is published, because the law requires it. One of them reports that in the six months to 30 June 2026 automated systems actioned 94.1% of the roughly 104 million pieces of content it removed, without human review. Read what that number counts: enforcement actions, not working hours and not jobs. It tells you the default path through the queue no longer passes a person. It does not tell you how many people are still employed, because the same report does not publish moderator headcount beside it, and the cases a person still sees are the ones the system could not settle.
We do not answer that with a number of years, and on this occupation the reason is unusually concrete: the thing to watch is published twice a year by your own employer. Take your platform's statutory transparency reports and put two consecutive ones side by side. Three signals are in them and none requires a forecast — whether the automated share is still rising or has flattened, whether the report breaks moderation down by language and whether yours is in it, and whether appeal volumes are rising while the automated share rises, because that combination means the system is clearing more and getting more of it wrong. A trend you can read yourself, in a document your employer is obliged to publish, is worth more than anyone's date.
Because that share is of actions, and actions are not evenly hard. The items a classifier settles are the ones it is confident about, which are disproportionately the clear ones; what is left is contested, context-dependent and slower per item. Two structural pieces also keep a person in the loop rather than beside it: appeals against automated decisions, which European law makes compulsory for the largest platforms, and the corrected examples the system itself needs in order to keep working. None of that guarantees a headcount, and this page does not claim one — it explains why a high automation rate and continued hiring are not a contradiction.
Writing the rule, not applying it. A policy has to produce the same verdict in two people's hands before it can produce a consistent one in a model's, and turning a value into a line that survives satire, reclaimed language, news value and a war video takes someone who has seen the cases that break it. Automating enforcement raises the cost of getting that wrong rather than lowering it, because an inconsistent rule now produces inconsistent results at machine scale. The honest caveat is that hard to automate is not the same as widely employed: policy seats are far fewer than queue seats.
What these judgements rest on#
1 of 6 task judgements on this page are backed by a verified event and 5 are platform inference, each labelled where it appears. Behind them sit 2 technology dimensions, a reconstructed trajectory since language models reached the public, and 1 verified events.