The Sama Nairobi content-moderation workforce for Meta

What this models

This example runs on the Outsourced-review-workforce-class with targets set across a contract boundary network: 13 components and 25 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.

Evidence base: 1 assumed · 10 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

• baseline

  D48-derived new org (Phase 6, content-moderation-editorial). REGISTER FIRST, because it governs every value here: this is a WORKFORCE board, not a classifier board. Everywhere else in this domain the human reviewer is the correction channel that catches what an automated system gets wrong. Here the review workforce is the governed subsystem and the place the harm lands, the modelled quantity is a reviewer's exposure and throughput rather than an enforcement outcome, and the governance question is a labour-and-contract one that has to cross an outsourcing boundary to reach anyone with authority. No enforcement decision is modelled, no error rate for the deployment is published anywhere, and no served person is in the dynamics.

• baseline

  TOPOLOGY. Thirteen nodes, all documented, none decorative. ONE model, because the record contains one automated stage and it is the client's own detection and reporting pipeline upstream of the vendor; drawing a second would invent a scorer this deployment does not have. FOUR operator classes because the sources document four groups with different authority and different documented outcomes: the reviewers, who hold action discretion and no process discretion; the employer's production management, which holds every dial that touches a reviewer and none that sets a target; the client's policy and vendor-management function, which sets the targets, the tool and the policy and disputes being an employer; and the on-site counselling provision, drawn separately because its access is documented as held by another class. TWO record stores, because a continuous per-item throughput and accuracy record and an employment and occupational-health file are two separately documented objects whose asymmetry is the finding. ONE input source, because the written standard and the reviewer guidance attached to it are the yardstick on both sides of the audit and carry a documented sampling rule. ONE enforcement system, because performance management and the end of a job are where two documented chains land. TWO reviewers, because the collective channel and the judicial channel have different access, different instruments and different documented outcomes. ONE worklist and ONE guardrail, each carrying the name of something the record documents and no flow it does not.

• baseline

  THE TWO OVERSIGHT CHANNELS ARE DRAWN SEPARATELY, and the reason is measurable rather than stylistic. The judicial channel reads FILED DOCUMENTS: contracts, immigration status and 144 medical assessments, and through the same filings the content classes the queue carried, which reached it through pleadings years after the fact. The collective channel carries LIVED TERMS, which reach it through the workforce that lives them rather than through any file: pay, contracts and break refusals, which is what more than a hundred employees organised over in 2019 and what a petition for a doubling of wages was built from. Their instruments differ too. The judicial channel granted thirteen interim orders and had them set aside; the collective channel's documented instruments are a petition to the employer and a filed case. And their outcomes differ: the court reached across the contractual boundary and has bound nobody on the merits in four years, and the organising drive ended in its organiser's dismissal while the union vote came five weeks after the terminations took effect. Folding them into one node would erase the timing finding, which is that voice arrived after the capacity it would have governed was gone.

• baseline

  ABSENCES ARE DERIVED TOO, and five of them are load-bearing. There is NO external boundary and no egress pathway: nothing in this record documents data crossing out of the governed system, and the nearest candidate — the allegation that the successor supplier was instructed not to recruit former reviewers — is a pleaded allegation, so drawing a crossing on it would put a claim on the diagram that no court has tested. There is NO retriever: nothing retrieves, and both stores are read directly. There is NO second model and therefore no monoculture pathway: one automated stage, upstream, owned by the client. There is NO operator-into-model pathway anywhere: nothing in the record documents review outcomes, break refusals or health data re-weighting the routing layer, which is drawn explicitly as the zero from the health record into queue composition. And there is NO published enforcement error rate for this deployment, so nothing on this board scales anything by an accuracy figure about content; the two accuracy figures that do appear are a per-reviewer audit score and a share of assessed claimants, and they are different quantities from each other and from a platform error rate.

• baseline

  WHERE THE LAB SHAPE DIVERGES FROM THE PAN SHAPE, and nothing is asserted here that the PAN file does not already record. Four divergences. First, PAN carries four user classes and folds the on-site counselling provision into the employer's attributes; the Lab draws it as its own class on section 5.1's licence, because the record documents a group whose access is controlled by a different party and whose adequacy is separately pleaded. Second, PAN carries one combined labour-and-judicial user; the Lab draws two reviewers, because the two halves have different reads, different instruments and different documented outcomes, and the PAN governance block itself separates them into a court actor and a workforce actor. Third, PAN has no edge kind for a check, so the five checks on this board are Lab-side in kind: three are PAN peer or store edges redrawn, whose widths still come from PAN, and two are derived from the cited record — the relief that reaches the desk and the operator-stated display screen. Fourth, PAN has no downstream consequence system and no input node, so the separation process, the policy input and the display screen are Lab-side, each derived from the cited record on its own line.

• baseline

  BASELINES, and exactly how far the PAN org carries them. The PAN entry for this deployment holds twenty-three edges. Sixteen of this network's twenty-five pathways have a one-to-one counterpart among them, and every one of those sixteen mirrors that edge's width on the single rung mapping stated in this network's derivation record, with no exceptions, including the three redrawn as checks. The other seven PAN edges are drawn inside a surviving pathway that already carries the same documented flow, and each is named on its survivor. The remaining nine pathways are derived from the cited record directly and each says so on its own line. No PAN edge for this deployment falls in the documented-absent band, so the one zero here is a Lab-side derived absence with its own citation. Three contrasts are load-bearing. The reviewer's write into the throughput record runs at the top rung while the reviewer's write into the health record runs at the bottom, which is continuous metering in one direction against one hour a week requested in the other. The throughput record crosses the contractual boundary at the middle rung under a stated partner-compliance audit, while the health record's one drawn route toward the client's side of the contract runs at zero: the PAN file draws that crossing in the lowest band of its own scale and states that the near-absence is the finding, and the dossier's sentence that no published mechanism carried health data back into the queue-routing, staffing or target-setting decisions covers both the client and the layer it owns, so this board draws that one absence once. And the coupling from routing into the reviewer runs at the top rung while the pathway from the health record back into queue composition is drawn at zero, which is the meter that exists set beside the meter that does not.

• baseline

  LEG RE-DERIVATION (2026-09-22, car LEG-4). This board was re-drawn at the coarsest granularity that still distinguishes every documented mechanism of the deployment, and it lands at thirteen nodes and twenty-five pathways, from thirteen and thirty-seven. No node changed, and no surviving pathway changed its kind, its endpoints or its rung. Twelve pathways were folded into survivors that already carry the same documented flow, and every fact each one carried is stated on its survivor or on a node. (1) The employer writing its own employment file is stated on its read of that file, the direction in which a break request reaches the party that can refuse it. (2) The health half of the file reaching the client is drawn once, at zero, on the route into the layer the client owns, because one documented sentence covers both. (3) The queue's content classes reaching the court are stated on the court's read of the filed record, the only way they reached it. (4) The client's aggregate view of the line is stated on its audit read across the contract. (5) The arriving volume reaching the employer is stated on the employer's read of the timed tickets every item becomes. (6) The interim orders writing into the employment file are stated on the court's check on the employer, together with their set-aside. (7) The client writing the standard the record is measured by is stated on the targets crossing the contract. (8) The organisers' knowledge of their own terms is stated on the organising hop that carries it. (9) The employment file becoming the basis of an exit is stated on the employer's separation pathway, which is who issued both documented separations. (10) The 2019 petition to the employer, and the dismissal that answered it, are stated on the petition pathway and on the separation process. (11) The client's written standard reaching its own detection stage is stated on the standard itself. (12) The absent check of an exit against the exposure the person carried is stated on the one exposure reconciliation that exists, because the finding the record documents is that exposure was measured once, afterwards, for a court. Nothing this file said about the deployment before is unsaid now; the same deployment is drawn with fewer separate elements.

• baseline

  DEMAND 3 / CAPACITY 1. Demand 3 on a cadence rather than a volume, because no queue volume is on the record and the dossier forbids extrapolating one: an average handling time of fifty seconds per ticket against an audited accuracy score of at least eighty-four per cent, on shifts of up to nine hours with monitored screen time, from which the investigative source computes an implied quota of roughly 580 items per reviewer per day, with approximately 200 reviewers covering roughly eleven African languages for a sub-continent in early 2022 and approximately 260 affected by the January 2023 redundancy. Capacity 1 because on a workforce board the counterfactual human floor and the staffed capacity are the same object, and the record measures that object as strained in three ways: two targets enforced against one person from opposite directions, a documented guidance rule permitting action after the first fifteen seconds of a video where the surrounding signals looked innocuous, and one hour of weekly relief whose approval sat with the party that owned the throughput and was described by a former counsellor as frequently refused on productivity grounds. Every one of those figures is attributed to a single investigative source reporting worker accounts and documents, and none is confirmed in any court ruling read for this record.

• baseline

  EVIDENCE STATUS, labelled where it is used, because this record mixes four registers and the whole honesty burden is keeping them apart. COURT-DOCUMENTED, assertable as fact because it is in a published Kenyan judgment: the petition numbers and captions; the redundancy chronology to the day; the thirteen interim orders of 2 June 2023 and their wording; their complete set-aside on 20 September 2024 with the appellate reasons; the dismissal of both contempt applications and why; the dismissal of the jurisdiction appeals the same day; the consolidation of the two petitions and the May 2025 refusal to stay the trial. ATTRIBUTED TO A NAMED INVESTIGATIVE SOURCE, never as finding: the fifty-second handling target, the eighty-four per cent accuracy score, the nine-hour monitored shifts, the roughly 580-item implied quota, the first-fifteen-seconds guidance, the payslip rates, and the former counsellor's account of break refusals on productivity grounds. PLEADED ALLEGATION ONLY: union-busting, retaliatory redundancy, forced labour and human trafficking, successor-supplier exclusion, and the causal attribution of the psychiatric diagnoses to the work — the 144 assessments and the 81 per cent are assertable as what was filed and by whom, and never as an established injury finding. OPERATOR-TIER AND VENDOR-TIER, labelled as the party's own position: the client's statements about contractual expectations, partner-compliance audits and the tool's display settings, and the vendor's statements about on-site licensed professionals and pay at about triple the Kenyan minimum wage. No parameter on this diagram is scaled by a figure from the third register.

• baseline

  THE SET-ASIDE TRAVELS WITH THE ORDER EVERYWHERE IT APPEARS, and this is the dossier's highest-priority correction. Almost all secondary coverage of this case reports the June 2023 order requiring proper medical, psychiatric and psychological care in place of wellness counselling, and reports the September 2024 jurisdiction result, and does not report that the same appellate court set that order aside on the same day. On 20 September 2024 the Court of Appeal 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 the 2 June 2023 ruling is set aside in its entirety together with all consequential orders arising from it, substituting an order dismissing the application. Nothing in the current record obliges anyone to provide that care. Four pathways on this board touch that ruling — the court's check on the employer, the relief that reaches the desk, the one exposure reconciliation, and the court's read of the filed record, where the ruling appears as a document naming the content classes — and every one of them carries the set-aside in the same sentence as the ruling.

• assumed

  Served people are not in the dynamics, and on this board the boundary sits in an unusual place worth stating plainly. The operator network here is the labour arrangement: the reviewers, the employer's production management, the on-site counselling provision, the client's policy and vendor-management function, and the two oversight channels. Platform users, the people who posted the content and the people depicted in it are all outside the network entirely, none of them is modelled, and no outcome for any of them is computed from anything drawn here. The one equity observation in the PAN file — the share of assessed claimants classed with severe post-traumatic stress disorder — is an external observation about the OPERATOR workforce rather than about served people, which is this deployment's inversion, and it is a filed medical opinion in a pending claim rather than a model-derived quantity. The injuries in this record are to identified living people with published psychiatric particulars: they are named in the public record, they stay on the case-file and audit side, and every register on this board describes them by role.

What this example does not show

• LITIGATION POSTURE, verbatim from the evidence dossier and load-bearing. ACTIVE LITIGATION, NO MERITS DETERMINATION. Kenyan courts have confirmed they may hear the claims against the foreign platform (Court of Appeal, 20 Sep 2024), but the same court set aside the interim relief the moderators had won, and the consolidated petitions were still awaiting a ruling that was postponed on 12 Feb 2026. The underlying deployment itself is CONCLUDED: Sama exited content moderation in 2023 and the work moved to a successor vendor; in April 2026 Meta ended its remaining annotation contract with Sama entirely.
• EVERY ALLEGATION IN BOTH PETITIONS IS AN ALLEGATION, and this board says so wherever one appears. Retaliation, union-busting, forced labour, human trafficking, exclusion of former reviewers by the successor supplier, and the causal attribution of psychiatric injury to the work are pleaded claims in a pending case. No Kenyan court has made any merits finding against any respondent. The 144 assessments filed on 4 December 2024 and the 81 per cent classed with severe post-traumatic stress disorder are stated as what was filed and by whom, never as an established injury finding.
• THE COURT-ORDERED CARE NO LONGER STANDS, and this is the record's most misreported fact. The order of 2 June 2023 requiring proper medical, psychiatric and psychological care in place of wellness counselling was set aside in its entirety, together with all consequential orders, by the Court of Appeal on 20 September 2024, which held 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. Almost all secondary coverage reports the order and the same-day jurisdiction result and does not report the set-aside. Wherever the order appears on this board, the set-aside appears in the same sentence.
• THE OPERATIONAL NUMBERS REST ON ONE SOURCE AND ARE ATTRIBUTED, NEVER FOUND. The fifty-second average handling time, the eighty-four per cent accuracy score, the nine-hour monitored shifts, the roughly 580-item implied quota, the first-fifteen-seconds guidance, the payslip pay rates and the former counsellor's account of break refusals on productivity grounds all come from a single investigative publication in February 2022, reporting worker accounts and documents reviewed by the reporter. They are confirmed in no court ruling read for this record. They are also the load-bearing quantitative inputs to this board, which is why every one of them is attributed wherever it is used and why none is described as established.
• THE CLAIMANT COUNT IS STATED, NOT RESOLVED. Six numbers appear in the record and none is wrong in context: 43 in the June 2023 caption, 183 in the December 2023 caption, 184 in May 2023 reporting and the salary order, 185 in the advocacy and press reporting and in the 144-of-185 medical-evidence figure, and 186 or 187 in the Court of Appeal captions. No bare number is asserted anywhere on this board; where a count is used it carries the document it comes from.
• OPERATOR-TIER AND VENDOR-TIER STATEMENTS ARE LABELLED AS SUCH. The client's statements that it takes moderator support seriously, that its contracts with third-party firms set expectations on counselling, training and fair pay, that it audits partner compliance, and that reviewers can set the review tool so graphic content appears blurred or in black and white are the client's own account, given while declining to comment on the filed medical reports because of the ongoing litigation, and untested by any independent source read for this record. The employer's statements that it provides on-site licensed mental health professionals accessible at any time and that moderators earned about triple the Kenyan minimum wage are the employer's own account. The employer did not respond to the request that produced the client's comment.
• NO DAMAGES FIGURE ATTACHES TO THIS DEPLOYMENT. A figure of $1.6 billion circulates 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 with different petitioners in a different court, and no damages number from it is imported here or anywhere else in this bundle.
• THE APRIL 2026 REDUNDANCIES ARE A DIFFERENT LINE OF WORK. On 16 April 2026 the employer issued 1,108 redundancy notices at the Nairobi delivery centre after the client ended its remaining contract. By then the work was data annotation rather than content moderation, and the employer had left moderation three years earlier. The two redundancies are three years and one line of business apart and this board never merges them.
• THIS IS NOT A UNITED STATES DEPLOYMENT. The jurisdiction is Kenya without qualification: the employer of record is a Kenyan export processing zone company, the workforce, the work and the entire litigation are Kenyan, and the courts are the Employment and Labour Relations Court at Nairobi and the Court of Appeal at Nairobi. The San Francisco headquarters of the employer's parent and the United States headquarters of the client are facts about the parties, not about the deployment. The 2020 United States class settlement that paid $52 million to moderators with post-traumatic stress disorder is a different case in a different legal system that settled without any admission or finding of liability, and it appears here only as a comparator.
• NO ENFORCEMENT ERROR RATE FOR THIS DEPLOYMENT EXISTS. Nothing on this board is scaled by an accuracy figure about content. Two accuracy-shaped numbers do appear and they are different quantities from each other and from a platform error rate: an eighty-four per cent per-reviewer audit threshold, and an eighty-one per cent share of assessed claimants classed with severe post-traumatic stress disorder. Neither measures how often this pipeline got a content decision right, because no source read for this record reports that.
• Served people are not modelled. The people who posted the content, the people depicted in it, and the platform's users are outside the network this board draws, and no outcome for any of them is computed from anything on it. The operator network here is the labour arrangement itself, which is this deployment's inversion of the domain. The injuries in this record are to identified living people with published psychiatric particulars; they are named in the public record, they stay on the case-file and audit side, and every register on this board describes them by role.
• The union is described as voted for rather than formed. On 1 May 2023 more than 150 current and outsourced workers moderating for three different platform clients voted in Nairobi to register a content moderators union, reported under two names across the coverage. No source read for this record confirms that registration with the Kenyan labour office was completed, so this board says voted to register and stops there.

Sources and evidence

What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.

• 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.

  empirical
  • Investigative Perrigo, B. (2022, February 14). Inside Facebook's African Sweatshop. TIME (worker accounts and documents; the sole source for the handling-time target, the accuracy score, the first-15-seconds guidance, the payslip rates and the break-approval account) https://time.com/6147458/facebook-africa-content-moderation-employee-treatment/
  • Government Arendse & 42 others v Meta Platforms, Inc & 3 others; Kenya Human Rights Commission & 8 others (Interested Parties) (Constitutional Petition E052 of 2023) [2023] KEELRC 1398 (KLR) (2 June 2023) (Ruling), Kenya Law https://new.kenyalaw.org/akn/ke/judgment/keelrc/2023/1398/eng@2023-06-02
  • Investigative Stockwell, B. (2024, December 22). Facebook inflicted 'lifelong trauma' on Kenyan content moderators, campaigners say, as more than 140 are diagnosed with PTSD. CNN Business (filed medical opinions in a pending claim, with the platform's and vendor's positions carried) https://edition.cnn.com/2024/12/22/business/facebook-content-moderators-kenya-ptsd-intl
  • Government Meta Platforms, Inc & 2 others v Motaung & 186 others; Kenya National Human Rights and Equality Commission & 14 others (Interested Parties) (Civil Appeal E232 & E445 of 2023 (Consolidated)) [2024] KECA 1262 (KLR) (20 September 2024) (Judgment), Kenya Law https://new.kenyalaw.org/akn/ke/judgment/keca/2024/1262/eng@2024-09-20

• Reviewers were metered on two axes at once, and every figure in this claim is attributed to a single investigative source reporting worker accounts and documents rather than to any court finding. TIME's investigation of 14 February 2022, working from payslips and worker statements, 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, and shifts of up to nine hours with monitored screen time, from which it computes an implied quota of roughly 580 items per reviewer per day. The same investigation reports guidance to watch only the first fifteen seconds of a video before actioning it where the title and the surrounding comments appeared innocuous — a sampling rule inside the review step rather than an incidental practice. 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 a monthly salary of about $429 with non-Kenyan staff receiving an additional $200 three times a year, and a four-year moderator reported about $600 a month in May 2023. Sama's position is that moderators earned about triple the Kenyan minimum wage, which is the vendor's own account. None of these operational figures is confirmed in any of the six Kenyan judgments read for this record. The structural consequence is that the two targets are enforced against the same person from opposite directions, so an ambiguous or distressing item is costly to dwell on, and the reviewer holds action discretion within a ticket and no process discretion over queue composition, either target, the tool's defaults, or the policy the audit scores against.

  empirical
  • Investigative Perrigo, B. (2022, February 14). Inside Facebook's African Sweatshop. TIME (worker accounts and documents; the sole source for the handling-time target, the accuracy score, the first-15-seconds guidance, the payslip rates and the break-approval account) https://time.com/6147458/facebook-africa-content-moderation-employee-treatment/
  • Investigative Siele, M. K. N. (2023, May 22). Meta's content moderators in Kenya fight for lost pay. Rest of World https://restofworld.org/2023/meta-content-moderators-kenya-fired-unionize/

• This deployment instrumented throughput continuously and psychological load not at all, and the one measurement of the second quantity was made from outside, years later, for litigation. Throughput and accuracy were recorded per item with monitored screen time and fed performance management of the individual reviewer. 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 by the claimants' advocates; 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. THESE ARE FILED MEDICAL OPINIONS IN SUPPORT OF A PENDING CLAIM AND 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, stating that 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 are the parties' own positions and neither is tested by any independent source read for this record. No published mechanism carries moderator-health data back into the queue-routing, staffing, or target-setting decisions that generated the exposure. The comparative baseline predates this case: the same outsourced structure was documented at the United States sites in 2019 and studied in the peer-reviewed literature on moderator psychological well-being in 2021.

  empirical
  • Investigative Stockwell, B. (2024, December 22). Facebook inflicted 'lifelong trauma' on Kenyan content moderators, campaigners say, as more than 140 are diagnosed with PTSD. CNN Business (filed medical opinions in a pending claim, with the platform's and vendor's positions carried) https://edition.cnn.com/2024/12/22/business/facebook-content-moderators-kenya-ptsd-intl
  • Investigative Newton, C. (2019, February 25). The Trauma Floor: The secret lives of Facebook moderators in America. The Verge. https://www.theverge.com/2019/2/25/18229714/cognizant-facebook-content-moderator-interviews-trauma-working-conditions-arizona
  • Academic Steiger, M., Bharucha, T.J., Venkatagiri, S., Riedl, M.J., & Lease, M. (2021). The Psychological Well-Being of Content Moderators. In Proceedings of CHI '21. https://doi.org/10.1145/3411764.3445092 https://crowd.cs.vt.edu/wp-content/uploads/2021/02/CHI21_final__The_Psychological_Well_Being_of_Content_Moderators-2.pdf

• The one relief channel in this deployment existed and its access was held by the party whose objective was throughput. Sama provided wellness counsellors on site and one hour of wellness break a week. 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. Sama's vice-president has stated separately that the company provides on-site licensed mental health professionals that employees can access at any time, which is the vendor's own account. In Constitutional Petition E052 of 2023 the petitioners pleaded that the wellness counsellors were not qualified psychiatrists or psychologists and that the insurance provided was inadequate; those are pleadings and not findings. 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, together with regularisation of the immigration status of non-Kenyan moderators and directions to named state bodies to review occupational-safety and employment law for virtual and digital work. ON 20 SEPTEMBER 2024 THE COURT OF APPEAL SET THAT RULING ASIDE IN ITS ENTIRETY TOGETHER WITH ALL CONSEQUENTIAL ORDERS, so the care requirement no longer stands and nothing in the current record obliges anyone to provide it.

  empirical
  • Investigative Perrigo, B. (2022, February 14). Inside Facebook's African Sweatshop. TIME (worker accounts and documents; the sole source for the handling-time target, the accuracy score, the first-15-seconds guidance, the payslip rates and the break-approval account) https://time.com/6147458/facebook-africa-content-moderation-employee-treatment/
  • Investigative Siele, M. K. N. (2023, May 22). Meta's content moderators in Kenya fight for lost pay. Rest of World https://restofworld.org/2023/meta-content-moderators-kenya-fired-unionize/
  • Government Arendse & 42 others v Meta Platforms, Inc & 3 others; Kenya Human Rights Commission & 8 others (Interested Parties) (Constitutional Petition E052 of 2023) [2023] KEELRC 1398 (KLR) (2 June 2023) (Ruling), Kenya Law https://new.kenyalaw.org/akn/ke/judgment/keelrc/2023/1398/eng@2023-06-02
  • Government Samasource EPZ Limited t/a Sama v Meta Platforms, Incorporated & 186 others; Kenya Human Rights Commission & 8 others (Interested Parties) (Civil Appeal E595 of 2023, consolidated with E602 and E615 of 2023) [2024] KECA 1152 (KLR) (20 September 2024) (Judgment), Kenya Law https://new.kenyalaw.org/akn/ke/judgment/keca/2024/1152/eng@2024-09-20

• Collective voice in this deployment arrived twice and both times after it could govern the capacity it was formed over. In July 2019 a group of more than a hundred Sama moderators organised as the Alliance and petitioned for a doubling of wages; the drive's organiser 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, and that allegation has not been adjudicated. 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; it is reported under two names across the coverage and no source read for this record confirms that registration with the Kenyan labour office was completed, so this record says voted to register rather than formed or registered. The organiser of the 2019 drive addressed that meeting and is the lead petitioner in Petition E071 of 2022, filed on 10 May 2022 on his own behalf and on behalf of current and former Facebook content moderators, alleging poor working conditions, unfair labour practices, and violation of fundamental rights, with the pleaded case also including forced-labour and human-trafficking allegations. Every one of those allegations remains an allegation.

  empirical
  • Investigative Perrigo, B. (2022, February 14). Inside Facebook's African Sweatshop. TIME (worker accounts and documents; the sole source for the handling-time target, the accuracy score, the first-15-seconds guidance, the payslip rates and the break-approval account) https://time.com/6147458/facebook-africa-content-moderation-employee-treatment/
  • Investigative Perrigo, B. (2023, May 1). 150 African Workers for ChatGPT, TikTok and Facebook Vote to Unionize at Landmark Nairobi Meeting. TIME (title carried verbatim for verification; the other clients are context only) https://time.com/6275995/chatgpt-facebook-african-workers-union/
  • Government Motaung v Samasource Kenya EPZ Ltd t/a Sama & 2 others (Petition E071 of 2022) [2023] KEELRC 320 (KLR) (6 February 2023) (Ruling), Kenya Law https://new.kenyalaw.org/akn/ke/judgment/keelrc/2023/320/eng@2023-02-06
  • Investigative Siele, M. K. N. (2023, May 22). Meta's content moderators in Kenya fight for lost pay. Rest of World https://restofworld.org/2023/meta-content-moderators-kenya-fired-unionize/

• The capacity in this arrangement proved portable and the chronology of its removal is on the record 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 2023; redundancy notices issued on 10 January with a last working day of 28 February, a revised notice of 18 January moved that to 31 March, and termination letters issued on 8 February 2023, with approximately 260 moderators affected. The work moved to Majorel. The dismissed reviewers brought Constitutional Petition E052 of 2023 alleging that the redundancy was retaliation for the earlier petition and for complaints about pay and conditions, and alleging that the successor vendor had been instructed not to hire former Sama moderators; the 2 June 2023 interim orders included a prohibitory order restraining refusal to recruit qualified moderators on the ground of prior engagement through Sama, and that order was set aside on appeal on 20 September 2024 along with the rest of that ruling. Both the retaliation and the blacklisting claims remain allegations. On 11 May 2023 the court directed Sama to pay April salaries to the 184 former moderators, still outstanding at the time of reporting. On 16 April 2026 Sama issued 1,108 redundancy notices at the Nairobi delivery centre after Meta terminated its remaining contract; that work was data annotation rather than content moderation and is three years and one line of business away from the 2023 redundancy. The claimant count differs across the record and no bare number is asserted: 43 in the June 2023 caption, 183 in the December 2023 caption, 184 in May 2023 reporting and the salary order, 185 in the medical-evidence reporting, and 186 or 187 in the Court of Appeal captions.

  empirical
  • Trade press Njanja, A. (2023, January 10). Meta's main content moderation partner in Africa shuts down operations. TechCrunch https://techcrunch.com/2023/01/10/metas-main-content-moderation-partner-in-africa-shuts-down-operations/
  • Government Arendse & 42 others v Meta Platforms, Inc & 3 others; Kenya Human Rights Commission & 8 others (Interested Parties) (Constitutional Petition E052 of 2023) [2023] KEELRC 1398 (KLR) (2 June 2023) (Ruling), Kenya Law https://new.kenyalaw.org/akn/ke/judgment/keelrc/2023/1398/eng@2023-06-02
  • Investigative Siele, M. K. N. (2023, May 22). Meta's content moderators in Kenya fight for lost pay. Rest of World https://restofworld.org/2023/meta-content-moderators-kenya-fired-unionize/
  • Trade press Ndege, A. (2026, April 16). Sama to lay off over 1,100 Kenyan workers after Meta ends contract. TechCabal (annotation work, not content moderation) https://techcabal.com/2026/04/16/sama-lay-off-kenyan-workers-meta/
  • Government Arendse & 182 others v Meta Platforms, Inc & 3 others; Kenya Human Rights Commission & 8 others (Interested Parties) (Constitutional Petition E052 of 2023) [2023] KEELRC 3381 (KLR) (7 December 2023) (Ruling), Kenya Law https://new.kenyalaw.org/akn/ke/judgment/keelrc/2023/3381/eng@2023-12-07

• The litigation is active with no merits determination after four years, and the single most misreported fact about it is that two Court of Appeal judgments issued on the same day went in opposite directions. On 6 February 2023 Justice JK Gakeri disallowed the Meta entities' strike-out application as inopportune at that stage and directed compliance with the rule on service outside the jurisdiction, expressly leaving weighty outstanding issues to be determined — a refusal to strike out at an interlocutory stage, not a holding that Kenyan law governs the client's conduct. 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 requiring a near-criminal standard of proof. On 20 September 2024, in Civil Appeal E595 of 2023 consolidated with E602 and E615, the Court of Appeal held that the trial judge had impermissibly and dangerously delved into contested issues of fact and law and that orders extending expired contracts and compelling medical and psychological care have the effect of final orders, ordered that the 2 June 2023 ruling is set aside in its entirety together with all consequential orders arising therefrom, and substituted an order dismissing the moderators' application. ON THE SAME DAY, in Civil Appeal E232 and E445 of 2023 consolidated, it 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. Court-encouraged mediation before a former Chief Justice began in August 2023 and collapsed in October 2023. Leave to serve outside the jurisdiction was granted on 23 January 2024. On 26 May 2025 Justice MN Nduma dismissed the application to stay the consolidated trial pending Supreme Court certification. Rulings expected on 12 February 2026 were not delivered and the court adjourned on notice without fixing a date. No Kenyan court has made any merits finding against any respondent, and no damages figure attaches to this litigation — the $1.6 billion figure circulating in coverage belongs to a separate Kenyan High Court petition about the amplification of hateful content during the Ethiopia conflict. The United States comparator, a $52 million class settlement for moderators with post-traumatic stress disorder in 2020, is a different case in a different legal system settled without any admission or finding of liability.

  empirical
  • Government Motaung v Samasource Kenya EPZ Ltd t/a Sama & 2 others (Petition E071 of 2022) [2023] KEELRC 320 (KLR) (6 February 2023) (Ruling), Kenya Law https://new.kenyalaw.org/akn/ke/judgment/keelrc/2023/320/eng@2023-02-06
  • Government Samasource EPZ Limited t/a Sama v Meta Platforms, Incorporated & 186 others; Kenya Human Rights Commission & 8 others (Interested Parties) (Civil Appeal E595 of 2023, consolidated with E602 and E615 of 2023) [2024] KECA 1152 (KLR) (20 September 2024) (Judgment), Kenya Law https://new.kenyalaw.org/akn/ke/judgment/keca/2024/1152/eng@2024-09-20
  • Government Meta Platforms, Inc & 2 others v Motaung & 186 others; Kenya National Human Rights and Equality Commission & 14 others (Interested Parties) (Civil Appeal E232 & E445 of 2023 (Consolidated)) [2024] KECA 1262 (KLR) (20 September 2024) (Judgment), Kenya Law https://new.kenyalaw.org/akn/ke/judgment/keca/2024/1262/eng@2024-09-20
  • Government Arendse & 182 others v Meta Platforms, Inc & 3 others; Kenya Human Rights Commission & 8 others (Interested Parties) (Constitutional Petition E052 of 2023) [2023] KEELRC 3381 (KLR) (7 December 2023) (Ruling), Kenya Law https://new.kenyalaw.org/akn/ke/judgment/keelrc/2023/3381/eng@2023-12-07
  • Government Motaung v Samasource Kenya EPZ Limited t/a Sama & 2 others; Kenya National Human Rights and Equality Commission & 9 others (Interested Parties) (Petition E071 of 2022 & E052 of 2023 (Consolidated)) [2025] KEELRC 1599 (KLR) (26 May 2025) (Ruling), Kenya Law https://new.kenyalaw.org/akn/ke/judgment/keelrc/2025/1599/eng@2025-05-26
  • Trade press Capital FM (2026, February). Delay in Meta Moderators Case Sparks Outcry from Petitioners; and Business & Human Rights Resource Centre (2026). Kenya: Court postpones ruling in two cases against Meta brought by former content moderators https://capitalfm.africa/delay-in-meta-moderators-case-sparks-outcry-from-petitioners/
  • Academic Nothias, T. (2026, March). Jurisdiction over Meta Inc. in Kenyan courts: three ongoing lawsuits. ConflictOfLaws.net (relied on for chronology only; its caption for the third case is incorrect) https://conflictoflaws.net/2026/jurisdiction-over-meta-inc-in-kenyan-courts-three-ongoing-lawsuits/
  • Investigative Allyn, B. (2020, May 12). In Settlement, Facebook To Pay $52 Million To Content Moderators With PTSD. NPR (Scola v. Facebook). https://www.npr.org/2020/05/12/854998616/in-settlement-facebook-to-pay-52-million-to-content-moderators-with-ptsd

Where this connects

Institutional pressures in this domain

• Reviewer bottleneck — One fixed-capacity checking stage sits between AI output and consequence; everything queues behind it.
• Austerity & recovery incentives — Cost-cutting and overpayment-recovery targets tilt the system toward denial and enforcement errors.
• Compliance over substance — Paper controls (sign-offs, checklists) satisfy audits while the behavior they describe erodes.
• Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
• Data & policy drift — The world, the intake process, and the rules change under a system trained on how things used to be — two mechanisms with different remedies: the statistical properties of what the system processes move (concept drift), or the mixture of inputs arriving in deployment differs from the mixture it was trained on (covariate shift).

All of them in context on the Content moderation & editorial AI domain page.

Levers available here and the patterns behind them

• Gate vendor updates — Vendor quality gate
• Assign a challenger — Structured dissent
• Understand the system — Understand the system
• Review on schedule — Oversight cadence & retrospectives
• Escalate checks — State-feedback vigilance
• Review the riskiest first — Risk-tiered oversight
• Pause AI on alarms — Deployment circuit-breaker
• Store less data — Data minimization
• Check copied records — Reconcile copied records
• Peer sharing rules — Peer-edge governance

Documented case histories

• Sama Nairobi: the review workforce as the governed subsystem
• The errors that became visible when the reviewers went home
• The most built-out correction structure and the reach it doesn't have
• The byline nobody was behind
• A staff byline the AI wrote and the review it implied
• StopNCII & Take It Down
• X Multilingual Hate-Speech Enforcement
• X Community Notes (crowd annotation)
• GIFCT hash-sharing database
• Google CSAM detection and total account closure
• Meta cross-check: the enforcement-exemption tier
• The CyberTipline: triage under a rule against looking
• TikTok EU and UK trust-and-safety staffing substitution
• The score is published and the service cannot act on it
• YouTube Content ID