The interview below is a composite. The speaker is not one person, and no one at this magazine interviewed them. Every answer is assembled from things content moderators have actually said or sworn to: published reporting, court filings, witness statements and settlement documents from the United States, Ireland, Spain and Kenya. We have kept the interview form because the people who do this work are bound by non-disclosure agreements that make a real, named, first-hand interview a legal risk for them — and that fact is part of the story.
You work for Facebook, but Facebook doesn’t employ you. Explain that.
My badge said something else. Accenture, Cognizant, Teleperformance, CPL, Sama — depending on the site, the country, the year. The queue was Facebook’s, or TikTok’s, or YouTube’s. The rules were theirs, the software was theirs, the accuracy targets were theirs. The paycheque, the HR department, and — this becomes important — the liability were somebody else’s. There are tens of thousands of us around the world, and most of us are arranged this way: the platform writes the policy, and a contractor absorbs the cost of enforcing it. When people imagine moderation they imagine a filter, something technical. It’s a labour arrangement. The filter is a person, and the person is deliberately kept at arm’s length.
What does a shift actually look like?
A screen, a queue, a timer. A piece of content arrives — a video, a photo, a comment thread — and you decide whether it stays or goes. You have seconds per item. There’s a figure that gets quoted a lot, that moderators make decisions in seconds, thousands of times a day; it comes from a 2017 piece in The Atlantic, not from a stopwatch study, but nobody who has done the job disputes the shape of it. What is measured precisely is accuracy. In the US class action, moderators described being held to a 98 percent success rate on their decisions. Ninety-eight percent, on material that ranges from nudity to beheadings to child abuse, at queue speed, where the difference between “terrorist propaganda” and “documentary evidence of a war crime” is a judgment call lawyers would argue about for a week. You make that call before the timer runs out.
And the content itself?
You stop describing it, is the honest answer. In the Kenya case — nearly 190 moderators suing Meta and Sama in Nairobi’s employment court — 144 of us were examined by Dr Ian Kanyanya, the head of mental health services at Kenyatta National Hospital. All 144 were found to have PTSD, generalised anxiety disorder and major depressive disorder. In 81 percent of cases the PTSD was severe or extremely severe. These were people moderating African-language content between 2019 and 2023, paid roughly eight times less than their US counterparts for the same queue of the worst things humans do to each other. The claim alleges intentional infliction of mental harm, human trafficking, modern slavery and unlawful redundancy. Meta and Sama have contested aspects of it. The diagnoses are in the court record.
Has anyone been made to pay for this?
Sometimes, quietly. In May 2020 Facebook agreed to pay $52 million to settle a class action brought in San Mateo County by moderators hired through third-party contractors — more than 10,000 current and former workers, each receiving at least $1,000 and up to $50,000 for documented severe psychological harm. The settlement still needed a judge’s approval when it was announced, and it settled the money without settling the structure. In Ireland, a moderator named Chris Gray filed a High Court action in 2019 against contractor CPL and Facebook, alleging PTSD; reporting at the time noted that CPL’s health insurance excluded mental-health cover. You can injure your mind doing the job; insuring that mind was apparently a separate product. And in January 2024 a Barcelona court ruled that a Teleperformance subsidiary was responsible for the mental disorders of a 26-year-old Brazilian moderator who started in 2018 and spent five years in psychiatric treatment. The court treated the illness as a workplace accident — reportedly the first time a Spanish court held a moderation company liable for a worker’s psychological harm. A workplace accident. As if the mind had been caught in a machine. Which, in a sense, it had.
The platforms would say the system is improving. AI now removes most of it.
TikTok told the Bureau of Investigative Journalism that 80 percent of violative content is now removed by automated systems. Notice what that means for the remaining 20 percent. Automation takes the easy cases — the spam, the duplicates, the clearly flagged. What’s left for humans is the ambiguous, the novel, the worst. The better the machines get, the more concentrated the residue becomes. We are the residue.
Why does the outsourcing matter so much? Isn’t it just how global companies operate?
Because subcontracting isn’t a side detail of this industry — it’s the design. Meta’s net profit in 2024 was $62 billion. TikTok’s parent made $33 billion. Teleperformance, one of the big contractors, made $589 million. The money pools at the top; the psychological cost and the legal exposure pool at the bottom, in a different corporate entity, often in a different country with a different labour regime. Outsourced Meta moderators in Ghana earn a base wage of £65 a month. Fairwork’s Mark Graham calls it a race to the bottom: contracts move between countries, and wages and conditions follow the contracts downward. If moderation were done in-house, in Menlo Park, at Menlo Park salaries, with Menlo Park occupational health, the true cost of the product would be on the platform’s own books. The queue doesn’t change. Only whose ledger the damage appears on.
The NDAs. You can’t talk about any of this?
The agreement says discussing the work is a breach of contract. That’s why you’re reading a composite instead of a name. But consider what the silence buys. The public argument about moderation is conducted almost entirely at the level of policy — what the rules should be, whether the platforms enforce them fairly. Almost nobody asks what the enforcement costs, because the people paying that cost are contractually prevented from saying. The NDA isn’t just protecting trade secrets. It’s protecting the arrangement itself: the distance between the company that profits from the decision and the person who has to make it in eight seconds, at 98 percent accuracy, for £65 a month.
Last question. What would you want a user to understand?
That the clean feed is a receipt. Someone examined everything that isn’t in it, item by item, and the platforms have built an entire corporate architecture — the contractors, the offshore sites, the quotas, the NDAs — to make sure you never see that person and they never have to employ them. Every so often a court forces the door open: $52 million here, a workplace-accident ruling there, 144 diagnoses in Nairobi. Then the door closes, the queue keeps moving, and the timer starts again.