Domain Atlas / Content moderation & editorial AI
Sama Nairobi: the review workforce as the governed subsystem
Explore this deployment in the PAN Lab ↗
In the PAN Lab, the readouts of this case's model organization carry a shaded evidence band whose width follows the least-established class among the modeling inputs the readings rest on.
The least-established input behind this case's model organization's readings is an assumption, not a measurement. Evidence base: 1 assumed · 9 published baseline.
The governed subsystem at the Samasource Kenya EPZ Limited (Sama) Nairobi delivery centre was the outsourced human review layer itself rather than any scoring model, and the arrangement's defining feature is a split principal. Automated detection and user reports upstream of the vendor fed a queue; items that stage could not dispose of were routed to reviewers employed by Sama, who applied the client's written policy through the client's review tool and returned an action. TIME reported approximately 200 reviewers covering roughly eleven African languages for a sub-Saharan Africa queue in February 2022; the Employment and Labour Relations Court's ruling of 2 June 2023 records approximately 260 moderators affected by the January 2023 redundancy. The party that set the policy, supplied the tool, composed the queue and set both review targets — Meta Platforms, Inc. and Meta Platforms Ireland Limited — is not the party that employed, insured or medically supported the reviewers, and its controlling legal position throughout the litigation is that it is not an employer at all. Everything the pipeline exists to filter out passed through the reviewers' eyes by construction; the exposure classes named in the court record and the filed medical assessments include gruesome killings, self-harm and suicide, sexual violence, explicit sexual content, child physical and sexual abuse, mutilated bodies, and conflict footage from the Ethiopia-Tigray war. No enforcement error rate for this deployment is published in any source read for this record.[4]
What happened
Start with the shape of the arrangement, because everything in this case follows from it. One company employed the reviewers. A different company set the policy they applied, supplied the tool they applied it in, composed the queue they worked, and set both of the targets they were measured against. And that second company's controlling legal position, maintained through four years of litigation, is that it is not an employer at all. That position is not a technicality here. It is what the case is about, and it is why the claims were brought as constitutional petitions in a labour court rather than as ordinary employment claims against the supplier alone.
The work itself is quickly described. Automated detection and user reports, upstream of the vendor and owned by the client, feed a queue; what that stage cannot dispose of goes to a person. Samasource Kenya EPZ Limited, trading as Sama — EPZ for Export Processing Zone, the tax-privileged export sector the company was registered in — staffed that human layer at its Nairobi delivery centre for Facebook and Instagram across sub-Saharan Africa, with reviewers recruited from across the continent, many on work permits tied to the job. TIME reported approximately 200 of them covering roughly eleven African languages in February 2022. The Employment and Labour Relations Court's ruling of 2 June 2023 records approximately 260 affected by the January 2023 redundancy. Everything the pipeline exists to filter out passes through those reviewers' eyes by construction, and the court record and the filed medical assessments name the classes: gruesome killings, self-harm and suicide, sexual violence, explicit sexual content, child physical and sexual abuse, mutilated and dismembered bodies, and conflict footage from the Ethiopia-Tigray war.
The operational terms come from one source and are attributed here everywhere they are used. TIME's February 2022 investigation, working from payslips and worker accounts, reports an average handling time target of fifty seconds per ticket, a quality or accuracy score requirement of at least eighty-four per cent audited against the client's policy, shifts of up to nine hours with monitored screen time, and an implied quota of roughly 580 items per reviewer per day. It reports take-home pay of about $1.46 an hour for Kenyan staff and about $2.20 an hour, roughly $440 a month, for non-Kenyan staff; a moderator interviewed in December 2023 gave about $429 a month with a $200 supplement for non-Kenyan staff paid three times a year, consistent in order of magnitude. Sama's position is that moderators earned about triple the Kenyan minimum wage. None of these operational figures is confirmed in any of the six Kenyan judgments read for this file. They are the load-bearing quantitative inputs to the model this case file stands beside, and that is exactly why they carry their attribution rather than shedding it.
Now consider what two targets do to one person. The clock says leave; the score says be right; and the same manager reads both meters. Something has to give, and the record names what did. TIME reports guidance to watch only the first fifteen seconds of a video before actioning it, where the title and the surrounding comments appeared innocuous. That is not carelessness. It is a sampling rule inside the review step, and it is the pressure-release the design requires — the only slack a reviewer can take alone.
The other slack is the one they could not take alone, and it is the mechanism this case turns on. Sama provided wellness counsellors on site and one hour of wellness break a week. Its vice-president has stated that the company provides on-site licensed mental health professionals that employees can access at any time. But a former counsellor told TIME that managers, rather than counsellors, held the final say over whether a break was granted, and frequently refused on productivity grounds. Put that beside the metering and the asymmetry is complete: the signal that would relieve the load is filtered by the party whose objective is the load, and the party that set the target sits on the other side of a contract it says is not an employment relationship. The petitioners separately pleaded that the counsellors were not qualified psychiatrists or psychologists, and that the insurance provided was inadequate. Those are pleadings.
Then look at what this deployment measured. Throughput was measured per ticket, continuously, with monitored screen time, and fed directly into performance management of the individual. Psychological load was measured once. It was measured retrospectively, from outside, by the claimants' own advocates: on 4 December 2024, medical reports for 144 of the 185 claimants who volunteered for assessment were filed at the Nairobi Employment and Labour Relations Court. The head of mental health services at Kenyatta National Hospital classed 81 per cent of those assessed as suffering severe post-traumatic stress disorder, with generalised anxiety disorder and major depressive disorder also diagnosed, and at least 40 were reported as misusing alcohol or drugs. Those are filed medical opinions in support of a pending claim. They are not adjudicated findings, and the causal attribution of the diagnoses to the work is a pleaded allegation. Meta declined to comment on the reports because of the ongoing litigation, saying it takes moderator support seriously, that its contracts with third-party firms set expectations on counselling, training and fair pay, and that moderators can customise the content-review tool so that graphic content appears blurred or in black and white. Samasource did not respond to the same request. Both of those are the parties' own positions and neither has been tested by any independent source read for this file.
Collective voice arrived, and it arrived late twice. In July 2019 a group of more than a hundred Sama moderators organised as the Alliance and petitioned for a doubling of wages. The organiser of that drive was suspended and dismissed on 20 August 2019, on grounds recorded as bullying, harassment and coercion said to have placed the relationship with the client at risk. He alleges the dismissal was union-busting; that allegation has not been adjudicated. Nearly four years later, on 1 May 2023 — five weeks after the terminations took effect — more than 150 current and outsourced workers moderating for three different platform clients met in Nairobi and voted to register a content moderators union, the first such body on the continent, reported under two names across the coverage. No source read for this file confirms that registration with the Kenyan labour office was completed, so this file says voted to register and stops there. The voice that might have governed this capacity was constituted five weeks after the capacity was dissolved.
The exit itself has a chronology and it should be read to the day. On 10 January 2023 Sama announced it was leaving content moderation to concentrate on computer-vision data annotation, and would not renew the Meta contract, which ran to the end of March. Redundancy notices followed on 10 January, with a last working day of 28 February; a revised notice of 18 January moved that to 31 March; termination letters issued on 8 February 2023. Approximately 260 reviewers were affected. The work moved to Majorel. The dismissed reviewers brought Constitutional Petition E052 of 2023, alleging the redundancy was retaliation for the earlier petition and for complaints about pay and conditions, and alleging the successor supplier had been instructed not to hire former Sama moderators. Both are allegations.
What the courts have actually done has to be stated in both directions, because the secondary record almost universally states it in one. On 6 February 2023 Justice JK Gakeri disallowed the Meta entities' strike-out application as inopportune at that stage and directed that the petitioner be given the opportunity to comply with the rule governing service outside the jurisdiction, expressly leaving weighty outstanding issues to be determined. That is a refusal to strike out at an interlocutory stage, not a holding that Kenyan law governs the client's conduct. On 20 March 2023 the court restrained implementation of the redundancy; the paid-leave order was extended in March and April, and the immigration status of foreign moderators was preserved. On 2 June 2023 Justice B Ongaya granted thirteen interim orders, including a requirement that the respondents provide proper medical, psychiatric and psychological care for the petitioners and other Facebook content moderators in place of wellness counselling, a restraint on refusing to recruit qualified moderators on the ground of prior engagement through Sama, a restraint on retaliation for filing the petition, and directions to named state bodies to review occupational-safety and employment law for virtual and digital work. On 7 December 2023 Justice MN Nduma dismissed both contempt applications, holding that placing employees on paid leave was not an action that constituted willful or deliberate disobedience and that the electronic evidence was not sufficient to prove that the respondents had replaced the petitioners; contempt requires a near-criminal standard of proof, which was not met.
Then came 20 September 2024, and two judgments of the Court of Appeal issued the same day in opposite directions. In the appeals against the interim ruling, consolidated as Civil Appeal E595 of 2023 with E602 and E615, the court held that the trial judge had impermissibly and dangerously delved into contested issues of fact and law at an interlocutory stage, that orders extending expired contracts and compelling medical and psychological care have the effect of final orders, and that no special circumstances justified mandatory injunctions. It ordered that the ruling of 2 June 2023 is set aside in its entirety together with all consequential orders arising therefrom, and substituted an order dismissing the moderators' application. The court-ordered care therefore no longer stands. On the same day, in Civil Appeal E232 and E445 of 2023 consolidated, the court dismissed the Meta entities' jurisdiction appeals with costs, holding that whether the appellants are engaged in virtual business in Kenya and whether the pleaded violations occurred in Kenya are contested questions of fact best resolved in a full hearing as opposed to an interlocutory application. Claimant-side and press coverage carried the second result widely and the first hardly at all. Both are in the same day's cause list.
Since then the case has moved slowly and the deployment has ended. Court-encouraged mediation before a former Chief Justice began in August 2023 and collapsed in October, with claimant counsel saying the respondents were buying time. Leave to serve the client outside the jurisdiction was granted on 23 January 2024. On 26 May 2025 Justice MN Nduma dismissed the Meta entities' application to stay the consolidated trial pending certification of a Supreme Court appeal, holding that the trial court lacked jurisdiction to grant such a stay while certification remained before the Court of Appeal. Rulings expected in the consolidated petitions on 12 February 2026 were not delivered; the court issued a ruling on notice and adjourned without fixing a date, and the petitioners protested publicly. As of August 2026 no merits determination has been reported. And on 16 April 2026 Sama issued 1,108 redundancy notices at the Nairobi delivery centre after Meta terminated its remaining contract — by then annotation work rather than moderation, three years and one line of business away from the redundancy this case is about, and not to be merged with it.
One number should be stated and not resolved. The claimant count differs across the record and none of the figures is wrong in its own context: 43 in the June 2023 caption, 183 in the December 2023 caption, 184 in May 2023 reporting and the salary order of 11 May 2023, 185 in the advocacy and press reporting and in the 144-of-185 medical-evidence figure, and 186 or 187 in the two Court of Appeal captions. This file uses whichever number its document uses and asserts none of them bare.
Finally, the comparative baseline, because this is not the first time. The same outsourced structure — a vendor site, a quota, an accuracy audit, in-house wellness provision, and psychological injury — was documented at the United States sites in 2019, studied in the peer-reviewed literature on moderator well-being in 2021, and settled in the United States in 2020 for $52 million to moderators with post-traumatic stress disorder, without any admission or finding of liability. That settlement is a different case in a different legal system and no damages figure travels from it to here. Nor does the $1.6 billion figure circulating in coverage that mixes this employment litigation with a separate Kenyan High Court petition about the amplification of hateful content during the Ethiopia conflict. That is a different case, and no damages number attaches to this deployment at all.
The sociotechnical reading
Most cases in this atlas ask what a system did to the people it was pointed at. This one asks what it did to the people it was made of.
The inversion is worth stating precisely, because it changes what a governance question even means. In every other content-moderation deployment here, human review is the correction channel: the thing you add when the classifier is wrong, the redundancy that catches the false positive. Cross-check delays it; hash-sharing consortia route around it; a triage queue rations it. In this deployment the human review layer is not the remedy. It is the exposed surface, and it is exposed by construction rather than by failure. Nothing malfunctioned to put child abuse imagery in front of a person in Nairobi. That is the design working. The whole pipeline exists to keep that material away from everyone else, and the mechanism by which it does so is a person looking at it, at a reported fifty seconds a ticket, for up to nine hours.
Once you see it that way, the governance question stops being about accuracy and becomes about who owns which dial. Four dials matter here: what enters the queue, how fast it must leave, how right the leaving must be, and whether a person may stop. The first three are set by a company that employs nobody in the room. The fourth is held by the company that does, and specifically by the function whose objective is the first three. There is no configuration of that arrangement in which the relief valve opens easily, because the party approving the break pays for it in throughput and the party setting the throughput pays nothing for the refusal.
That is the split principal, and it is why the legal claim had to be constitutional rather than contractual. An ordinary employment claim against the supplier reaches the party that approved or refused a break and cannot reach the party that made the break expensive. The claimants' answer was to sue both and to argue jurisdiction, and the Kenyan Court of Appeal's answer on 20 September 2024 was that whether the foreign client is engaged in virtual business in Kenya, and whether the pleaded violations occurred there, are contested questions of fact for a full hearing. That is a real and unusual result. It is also, four years in, the only thing anybody has actually won, and on the same day the same court removed everything else.
The second structural fact is what this deployment chose to instrument. Throughput has a meter that runs continuously, per item, with monitored screen time, and feeds a performance process. Exposure has no meter at all. It has, instead, a single retrospective measurement made by the wrong party at the wrong time for the wrong purpose: 144 assessments, filed as evidence, four years after the work, by the claimants' own advocates, in a case that has not been decided. A system that measures what it wants to maximise and leaves the thing it damages to litigation is not a system with a monitoring gap. It is a system whose monitoring is complete for the purpose it was built for. The gap only looks like a gap if you believe the reviewer's condition was ever one of the quantities being managed.
The third is a timing observation and it is the one this file would most like a reader to keep. Two channels in this record could have changed the arrangement while it ran, and both arrived after they could. The organising drive of July 2019 was the workforce trying to acquire process discretion — a say over the dials it did not hold — and it ended, five weeks later, with its organiser dismissed on recorded disciplinary grounds he says were pretextual and no court has ruled on. The union vote of 1 May 2023 came five weeks after the terminations took effect: the first content moderators union on the continent, constituted over a workforce that no longer existed. And the courts, the one channel that has bound anyone, took twenty-eight months to grant interim protection and then had it removed fifteen months later, with no merits ruling four years and three months after filing. Governance that arrives after the capacity it governs has been dissolved is not weak governance. It is a different thing from governance, and this case is the clearest documented instance of it in the atlas.
The fourth is about portability, and it is the quietest. When the arrangement became costly the capacity moved. The supplier left the line of business in January 2023 and approximately 260 jobs went with it; the work was with a successor within months; in April 2026 the client ended the remaining contract at the same site and 1,108 more notices followed. Every part of this deployment turned out to be portable except two: the injury, which stayed with the people, and the standing to sue over it, which stayed with the people too and is worth less each year it is not decided. A petitioner told reporters after the February 2026 adjournment that the delay was not merely a procedural setback but a profound failure of the judicial promise, and another described what the prolonged case had done to her ability to find work. Those are quotes about the remedy, not about the harm.
Two boundaries hold across this file and neither is decoration. Served people are not modelled: the platform's users, the people who posted the content and the people depicted in it are outside the network the Lab board draws, and no outcome for any of them is computed from anything on it — this deployment's operator network is the labour arrangement itself, which is precisely the inversion. And the four registers stay apart: what a published Kenyan judgment establishes, what one investigative publication attributes to worker accounts and documents, what remains a pleaded allegation in a case with no merits ruling, and what a party says about itself are four different weights of evidence, and no sentence here lets one borrow the authority of another. The people in this record are identified individuals with published psychiatric particulars. They are named in the public record and they are not named here.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.