The Workers Who Trained the AI Are Being Let Go With Nothing

Seven hundred and twenty workers at Meta's Sandyford campus in Dublin have been told their jobs are at risk. They did AI annotation and content moderation — the unglamorous, labour-intensive work that sits beneath every language model and content feed. Roughly 400 of them will receive zero statutory redundancy pay because they have been in post for less than two years. Meta's own 8,000 employees, cut in the same broad restructuring, received a minimum of four months' severance plus additional pay tied to tenure. The gap between those two numbers is the story.

What happened on 29 May. Covalen workers picketed Meta's European headquarters at Ballsbridge, Dublin. The Communications Workers Union (CWU) has voted for strike action and is calling for government intervention. Covalen — the outsourcing firm that employs the workers, rebranded from Majorel Ireland — notified workers and the Irish government of the potential redundancies after Meta announced it was reducing reliance on external vendors as AI systems take on annotation and moderation tasks previously done by hand.¹²

Meta's statement, as reported by WIRED: "Meta is reducing its reliance on external vendors as AI takes on more of this work."¹ Clean. Framed as a vendor relationship, not a workforce decision. The people involved are, legally, someone else's employees.

The legal structure is doing exactly what it was designed to do. Ireland's Redundancy Payments Act 1967 sets a two-year continuous service threshold for statutory redundancy. Workers below it receive nothing beyond their contractual notice period. The statutory rate for those who do qualify is two weeks' pay per year of service plus one bonus week, capped at €600 a week gross.³ That cap is not generous. But zero is less than not generous.

The two-year threshold was not designed as a loophole. It was designed for a labour market where most employment relationships were long-tenured and the short-tenure exception was genuinely exceptional. What it was not designed for is a labour market in which a specific category of worker is deliberately cycled through on contracts shorter than two years, at industrial scale, specifically to supply a function that a client company does not want to carry on its own books.

That is the annotation labour model. Short-cycle contracts, renewable, disposable when the model iteration is trained or the content-moderation volume drops. The two-year threshold does not accidentally catch these workers. It was always going to catch these workers, because the job design was always structured to catch them there.

Who holds the legal obligation matters here. Meta is not the employer of record. Covalen is. The redundancy obligation — statutory or enhanced — sits with Covalen, not with the company whose campus the workers report to, whose systems they trained, and whose vendor-reduction decision triggered the redundancy process. This is not a technicality. It is the whole architecture of the arrangement.

A CWU representative described it plainly: "The outsourcing model is being used as a shock absorber for AI transitions — that has to change."⁴ I think that framing is analytically correct, not merely rhetorical. The shock absorber metaphor is precise. When AI capability improves to the point where annotation volume can be reduced, the cost of that transition — in redundancy payments, in disrupted income, in careers that end mid-cycle — does not land on Meta's income statement. It lands on Covalen's (to whatever extent Covalen is liable), and on the workers (to the extent they are not). The 400 workers below the two-year threshold absorb the full cost themselves.

The comparison with Meta's direct employees is not incidental. In early 2025, Meta announced cuts of approximately 8,000 direct employees, roughly 10% of its global workforce. Those workers received a minimum of four months' severance plus additional pay based on tenure.⁵ The same broad restructuring that is generating this contractor reduction produced that package for people who happened to sit inside the employment relationship rather than outside it.

The 720 Covalen workers were doing work that made Meta's AI models function. Annotation — labelling data, flagging edge cases, providing the human judgements that supervised learning requires — is not peripheral to the AI product. It is constitutive of it. The people doing this work at the Sandyford campus were not ancillary to Meta's AI development. They were part of it, structured to remain invisible to the balance sheet.

Efficiency for the operator is precarity for the worker.

The counterpoints deserve engagement, not dismissal. There are three serious ones.

The first is that the precarity predates AI. Content moderation and annotation were never structured as stable long-tenure careers. The job design was always temporary, always outsourced, always high-churn. If that is true — and it largely is — then attributing the workers' situation specifically to AI displacement elides a prior design choice that was made long before current AI capabilities made annotation redundancy plausible. The workers were always going to be disposable. AI just accelerated the schedule.

This is true and it does not let anyone off the hook. The prior design choice was also wrong. The fact that the precarity was baked in earlier makes the case for structural reform stronger, not weaker. If the design was always exploitative, the current moment is not an exception to be managed — it is a crisis that reveals what the design always was.

The second counterpoint is that AI annotation work may not disappear entirely. Some labour economists argue that as base models improve, demand for higher-quality complex annotation (edge cases, fine-tuning, reinforcement learning from human feedback) may partially offset displacement of routine annotation work. This is theoretically coherent. There is no evidence it applies to the 720 Covalen workers in Dublin, and the partial-offset argument has been used to undercount displacement in previous technology transitions often enough that I hold it with some scepticism. But it is not nothing, and a piece that ignored it would be less honest than this one.

The third counterpoint is that Ireland's two-year threshold is not an outlier in European terms. Several EU member states have qualifying periods of one to two years for statutory redundancy. Employer groups argue the threshold protects businesses from liability for genuinely short-term hires. The argument has some validity in its intended context. It has much less validity when the threshold is being systematically used to structure an entire workforce category below it.

The Irish government's position is structurally uncomfortable. The CWU is calling for government intervention. What intervention would mean in practice is either lowering the qualifying period, extending redundancy liability to client companies, or some combination of both. Each of those options increases the cost of operating the outsourcing model in Ireland, at the margin.

Ireland's attraction as a European hub for US tech platforms rests substantially on a combination of corporate tax rate (12.5%) and the infrastructure of outsourced labour that allows companies to maintain large European operations without directly employing the people who do much of the work.⁴ The Irish state has a fiscal interest in the continued presence of Meta, Google, Apple, and their peers that is not abstract. It is revenue. The same government being asked to strengthen worker protections in outsourced tech labour is the government that has structured its foreign direct investment model around the conditions that make those workers vulnerable.

That tension is genuine. It does not resolve cleanly. ORA is not going to pretend it does.

What makes this case matter beyond Dublin. The annotation and content-moderation workforce at Meta's Sandyford campus is one instance of a global model. The workers doing AI data labelling in Dublin are structurally similar to workers doing the same work in Nairobi, Manila, and Bucharest — cycling through short-term contracts with outsourcers contracted to US platforms, building capabilities that will eventually reduce the demand for their own labour. The outsourcer absorbs the headline employment relationship; the platform absorbs the capability gains; the workers absorb the exit cost.

When AI systems become capable enough to reduce demand for a given annotation function, the transition cost does not appear on the platform's income statement as a redundancy liability. It disappears into the contractor relationship, and often below the statutory threshold, because the workforce was designed to sit there.

This is not specific to Meta. It is not specific to Ireland. It is the playbook for AI-adjacent labour at scale, and Dublin on 29 May is just where it became visible.

What a CWU sector official said outside Meta's Ballsbridge headquarters: "These workers trained the AI that is now taking their jobs and they deserve better than being cast aside with nothing."² That sentence is not rhetorical. It is a description of the distributional incidence. The productivity gain from the AI system these workers helped build will flow to Meta's shareholders and to the users of an improved product. The cost of the transition — for 400 workers, the entirety of that cost — flows to workers who have no statutory floor and no collective bargaining relationship with the company whose AI they were training.

That is what redistribution of risk without redistribution of reward looks like at the operational level. It does not require a political position on AI. It requires only reading the structure of who pays what and to whom.

The conversation about AI's benefits is real. The productivity gains are real. The improved systems are real. What keeps getting skipped is the question of how the costs of transition are allocated, and to whom. In Dublin, the answer is: to the 400 workers below the threshold, who trained the model, and who will receive nothing for having done so.

Footnotes

¶