They didn’t go quiet. You just can’t sue from Madagascar. “Acquisition, then silence” was the wrong mechanism: the $14 billion stake did not buy silence, and ghost labor does not vanish when you buy the vendor. What the record actually shows is standing, then settlement; no standing, then a paragraph. Disclosure here is a by-product of litigation, not governance. The door never closed. The baseline was never built. The work stays cheap because nobody is counting.
A good slogan is a small act of violence against detail. “Acquisition, then silence” has the rhythm of something that must be true: three words, a comma, a door closing. It comes from a sociologist, in an interview, about the people who label the data that AI learns from, and it names a pattern. A firm gets uncomfortable scrutiny over how those workers are treated; the response is to absorb the company that employs them, so that nobody outside has to be told anything again. It is an elegant theory of corporate behaviour. This essay did the unglamorous thing that slogans rarely get, and checked it. The result is awkward for the slogan, more interesting than the slogan, and, to be clear from the start, aimed at one compressed line from an interview and not at the research behind it, which is good.
One note before the check. The author works at a global consulting firm, the same species as the firm in the story below, and did not check whether that employer has any annotation or offshore data supply chain. Section VII says more.
I. The Part of the Job Nobody Counts
Before the slogan there is an argument, and the argument is good.
Antonio Casilli, a sociologist who has spent years on the labour behind digital platforms, published a chapter in 2025 in the SAGE Handbook of Digital Labour on what he calls the inconspicuous production of artificial intelligence. The idea is simple and slightly maddening once seen. Every job, including skilled ones, contains a hidden sliver of work that nobody has written down: the checking, the fixing, the labelling, the small human corrections that make a system look like it runs itself. His complaint about the best-known studies of AI and jobs is methodological. They break jobs into tasks, count the tasks a machine can do, and so measure only the conspicuous, codifiable part. The remainder is silently assumed to be work that “cannot be automated, thus is not work”.
The workers he names are scattered across the map: click farmers in Venezuela, moderators for a large social platform in Kenya, data annotators in India, ride-hailing drivers in Germany. And he adds a twist that deserves more attention than it gets. The more a system automates the visible part of a task, the more it leans on the hidden part, because the hidden human labour is what keeps supplying the data that keeps the automation working. Automation does not retire the ghosts. It puts them on a treadmill.
The philosophical point is the durable one. A measurement that cannot see something does not report it absent; it reports it free. An economy of ghosts is cheap to run and hard to audit, and the two facts are related.
II. A Week in Madagascar
The conditions are documented in published fieldwork; the client’s reaction rests on one account.
Casilli does not only theorise. With Clément Le Ludec and Maxime Cornet he co-authored a peer-reviewed study, published in Big Data & Society in 2023, of how work is outsourced between France and Madagascar for AI. In two of the co-authors’ own account in The Conversation, workers there earned between 96 and 126 euros a month, with local supervisors paid eight to ten times as much, at the far end of a long chain of subcontracting. A follow-up on Madagascar and Venezuela appeared in 2026. The study names no firms, which matters later.
Against that background, an interview published in February 2026 by Philonomist carries the scene that gives the slogan its force. In 2022 Casilli spent a week in Madagascar at a company doing annotation work. He found workers paid about 90 euros a month, a figure close to the range in the published fieldwork. He then showed the evidence to senior managers at the consulting firm that was the client, and met, he says, with silence, a deafening one. The firm, he says, later bought WNS, one of its own data-annotation providers, in order to bring the work in-house, keep control of it and shield itself from external oversight.
The interview text does not name the client in the Madagascar sentence itself. The conversation around it points to Capgemini, and the pronouns do the rest; that identification is firm enough to work with and not firm enough to treat as settled. This essay worked from the published text only, Casilli was not asked to clarify, and because the fieldwork anonymises its firms, the published research cannot confirm the client either. The purchase itself is not in doubt. Capgemini announced the WNS deal on 7 July 2025 and closed it on 17 October 2025 for 3.3 billion dollars in cash. Its own stated reason is different and perfectly respectable: it wanted to build a leader in what it calls agentic-AI-powered intelligent operations. The two accounts are not exclusive. A firm can buy a supplier for growth and also be relieved by what that does to the view from outside.
Here is the first gap. WNS is mainly a business-process company, with delivery centres its own pages list in thirteen countries, and Madagascar is not among them. It does sell data annotation through a product data operations unit. The interview does not say that the workers Casilli met in Madagascar were WNS’s, and nothing in the public record found for this essay says so either. The firm bought a supplier of annotation work. Whether it bought the supplier that employed the people in his story is unestablished.
III. The Slogan, Checked
In its literal terms the slogan fails twice; in its narrower reading it is unresolved.
The slogan’s second example is Scale AI, the data-labelling company. Casilli says that since Meta acquired it in June 2025 for 14 billion dollars, nothing more has been heard about what happens there; and that it is always the same: “acquisition, then silence”. Start with the first word. Meta’s own quarterly filing describes what it took as a non-voting minority of Scale’s equity, which it accounts for as an investment in a company over which it has no significant influence. Scale’s announcement said the investment valued the company at over 29 billion dollars, that Meta would hold a minority, that its founder Alexandr Wang would go to Meta while staying on Scale’s board, and that Scale remained independent. The press put the sum at about 14.3 billion dollars for 49 per cent.
In fairness, “acquired” is a common press shorthand for a deal like this, and not a Casilli invention. And a 49 per cent non-voting stake plus the founder’s move might amount to something close to capture in practice, a point returned to below. But the literal mechanism the slogan needs, absorbing a supplier and so deleting its disclosures, requires control, and a non-voting minority does not give control. As a description of what happened, “acquisition” is wrong.
Now the second word. Before the deal, Scale was already being sued by contractors in California over pay and classification, and a second suit alleged psychological harm from exposure to traumatic content. TechCrunch reported in May 2025 that the US Labor Department had dropped its investigation into Scale, a month before the deal, and said the reasons were unclear. After the deal, the record does not go quiet. Bloomberg and others reported layoffs at Scale in July 2025. An AFP report in October 2025, relayed on Casilli’s own site, quoted a Kenyan labeller calling the work modern slavery, and Scale was reported to have responded on wages and mental-health support. The official settlement website confirms a 12.5 million dollar fund for California contributors, preliminarily approved by the San Francisco court, with a final approval hearing set for 30 October. Casilli published his own account of leaked documents in September 2025. Silence turned out to be a remarkably noisy thing.
IV. What Survives
The narrower claim is smaller than the slogan and, unfortunately, truer.
Strip the slogan down and something does stand, and on the narrower reading Casilli may mean, that the company itself said little about conditions, the record is closer to him than the paragraph above allows. In the sources checked, very little of that noise came from Scale. It came from plaintiffs and their lawyers, from a wire service, from an academic with a blog, and, until it was dropped, from a regulator. Scale’s own contributions, in those sources, were a layoffs memo, responses to journalists and the formal notices a settlement requires. Nothing found shows Scale published labour-conditions material before the deal either, so a fall in its candour cannot be shown, only a low level of it. The one institutional witness, the Labor Department, left the room before the deal, and the reporting could not say why.
The pattern the record supports is therefore not “acquisition, then silence”. It is that the people behind the model become visible mostly when somebody with a grievance and a way to file it makes them so. That second condition is unevenly distributed. The class action and the settlement concern contributors resident in California. The reporting that reaches the people furthest from the buyer’s lawyers, in Kenya or Madagascar, arrives as journalism and scholarship, which is slower and rarer. Put crudely, the workers who can sue get settlements, and the workers who cannot get a paragraph in an interview, or, in Madagascar’s case, a peer-reviewed paper that does not name the firms. Disclosure in this industry looks less like a reporting practice and more like a by-product of litigation, and litigation has jurisdictions.
V. Silence Is a Measurement
Absence of news means something only against a baseline, and baselines can be built.
To say that silence followed an event, one needs to know what there was before. That baseline need not be an external index. It can be the firm’s own earlier habit of disclosure, or the habits of comparable vendors that were not bought. A company that filed reports and then stopped has gone quieter in a way anyone can measure. For the people at the centre of Casilli’s anecdote, no such baseline was found: no regular, independent instrument tracking conditions for annotators, and no evidence that their employer published any account of them before or after. Where nothing was being said, nothing can be seen to stop.
A real test would be dull. Was the supplier independent before, in the sense that someone outside could read its disclosures? Did it publish anything about its workers’ conditions? Did independent reporting about those workers measurably fall afterwards, compared with similar vendors that were not acquired? On the evidence found, Scale fails the first and the third questions, which is to say the pattern is not there. For WNS, no case has been identified, because nobody has shown that the Madagascar workers were its workers, and so there is nothing yet to score. That is not a vindication of the slogan. It is a description of how thin the instrumentation is, which is, if one is honest, Casilli’s deeper point wearing different clothes. A slogan that cannot be tested is not yet a finding.
VI. The Counter-Case, in Fairness
Buying a supplier has more motives than hiding it, and swapping one has a different outcome.
Large acquisitions have many reasons: revenue, capability, control of quality, speed, the plain wish to own what one depends on. Capgemini says growth in agentic-AI operations. Meta’s investment bought a stake in a data supplier and a founder’s services at the same time. Nothing in the record found here lets anyone rank those motives, and the stated purpose in the interview is Casilli’s reading of someone else’s intentions, however confidently put. There is also a fair argument in the other direction. Work brought in-house comes under the buyer’s own employment law, its own compliance machinery and its own reputational exposure, and can end up better governed than work spread across a chain of contractors nobody has to answer for. Vertical integration might do some workers good. This essay has no evidence either way, which is exactly the problem.
A better-known case points somewhere different. As reported, moderators who worked in Nairobi for Meta’s contractor Sama sued Meta, Sama and a later vendor, Majorel, over dismissals and alleged blacklisting after Sama ended that work in 2023, and Kenyan courts ruled that Meta could be sued there. That is a vendor swap, not an acquisition, and the visibility came from litigation, which fits section IV better than it fits the slogan. It is also a reminder that companies have more than one way to put distance between themselves and a workforce, and that “acquisition” is only one of them.
And a slogan from an interview is not a peer-reviewed claim. Compression is what interviews are for, and it is a little unsporting to cross-examine a headline. The academic argument in the first section, and the fieldwork in the second, are separate things and untouched by anything above.
VII. The Interest, Stated
An essay about someone else’s supply chain should say where its author sits.
The interest belongs here, in the open. The author sells nothing into the annotation or data-labour territory, which makes this one of the few essays in this newsletter with no product at stake in its conclusion. But the author works at a global consulting firm, the same species as the firm in the anecdote, and did not check whether that employer has any annotation or offshore data supply chain. Treat that as a gap, not a clean bill of health. A writer who scrutinises one consulting firm’s supply chain and says nothing about his own is exactly the kind of omission the essay is about. Nobody named here, Casilli, Capgemini, WNS, Meta or Scale, was approached for comment; the essay works from published material, and a right of reply would be the first thing to add.
VIII. The Case Against This Essay
A check of a slogan is cheaper than the slogan, and deserves the same suspicion.
First, the check leans on secondary reporting for much of the post-deal record. Meta’s own filing and Scale’s own announcement establish the structure; the 14.3 billion dollar and 49 per cent figures, the layoffs and the October report come from press coverage, and the settlement terms come from its official website, not from the court order itself, which was not seen. A fuller check would read the docket. If any of those items is wrong, the sentence about noise has to be redone.
Second, absence of evidence cuts both ways. The search found no link between WNS or Capgemini and Madagascar, and it cannot rule one out. A private contract, a subcontractor or a shifted vendor would leave no public trace, and the fieldwork anonymises its firms by design. The anecdote may be exactly as described.
Third, “silence” may have meant something narrower than the slogan’s literal reading: that the companies themselves said nothing about conditions. On that reading the record is closer to Casilli than the third section allows, and part of this essay is a quarrel with his phrasing.
Fourth, the minority stake may not be as innocent as the structure suggests. Money and a founder’s allegiance can shape a supplier’s behaviour without a single voting share. This essay did not test whether independence was real in practice, or whether the deal’s form was chosen for reasons unrelated to labour.
Fifth, the settlement and the suits concern US-resident contractors, so the conclusion that disclosure follows standing is a hypothesis built on a skewed sample, and one case does not make a pattern. It is also not new: scholars of labour have long argued that visibility follows organising and legal standing, and this essay restates that tradition with a fresh example and no new proof.
Sixth, this essay is the product of a single drafter and a handful of checks. Several of its facts were corrected during its own preparation, including the Meta stake, the post-deal record and the credit owed to Casilli’s fieldwork, which is a fair warning about how easily slogans, including this essay’s, drift from the record.
What would prove this wrong: independent evidence that reporting about a vendor’s workers measurably fell after a full acquisition, compared with similar vendors that were not acquired; a Capgemini or WNS disclosure placing Madagascar annotators inside the acquired business; or a reading of the full interview in which Casilli plainly means only company disclosure, which would turn this essay’s headline objection into a misreading.
IX. A Research Question, Not a Replacement Slogan
The more accurate line is harder to say, and is better offered as a question.
Slogans win because they are short, and facts arrive late because they are long. A more accurate line would be harder to say in an interview and harder to put on a slide, and it should be held as a question and not a slogan: does public visibility of AI’s hidden workers follow legal standing? Where workers can sue, the record shows litigation and, eventually, settlements. Where they cannot, it shows an anecdote and a reporter’s paragraph. Standing, then settlement; no standing, then silence. That is a hypothesis resting on a skewed sample, and, as the previous section says, a restatement of an old argument. It would take comparison, across vendors with and without litigation, over time, to test it.
The practical conclusion is dull and stubborn. If the people behind a model are to be visible, somebody has to build the instrument that sees them before anyone buys anything, because silence can only be measured against a baseline, and none was found. Until then the industry’s hidden labour will remain what Casilli says it is: inconspicuous by design, and cheap precisely because nobody is counting.
Dr Luke Soon is an AI Leader at PwC Singapore, covering fourteen Asia Pacific markets, and co-author of Singapore’s Model AI Governance Framework for Agentic AI. He writes at genesishumanexperience.com under the Genesis: Human Experience in the Age of Artificial Intelligence banner.
Leave a comment