Wired reported on April 28 that more than 700 workers at Covalen, a Dublin-based Meta contractor doing AI training and content moderation, have been told their jobs are at risk. Roughly 500 are data annotators โ the people who check material generated by Meta's AI models against the company's rules against dangerous and illegal content. Workers were informed via a brief video meeting on Monday with no questions allowed. The cuts are the second round at Covalen in five months; together with the November layoffs (around 400 jobs, ending in a worker strike), Covalen's Dublin headcount is on track to be almost halved. Meta announced one-in-ten company-wide job cuts last week while separately announcing it will nearly double AI spending.
The work being cut is what most AI labs call adversarial annotation and red-teaming. Workers craft elaborate prompts to try to bypass model guardrails โ provoking the model into generating CSAM, suicide content, or other prohibited output, then logging where the model fails. One worker described it to Wired: "you spend your whole day pretending to be a pedophile." Another said: "it's essentially training the AI to take over our jobs. We take actions as the perfect decision for the AI to emulate." That is the tradeoff Meta is making: humans absorb the psychological cost of red-teaming, their judgments get distilled into automated safety classifiers, the classifiers then replace them. A six-month cooldown clause in the contracts โ laid-off workers cannot apply to competing Meta vendors โ adds a labor-rights wrinkle the Communications Workers' Union is contesting.
Two implications stand out. First, the safety question. If Meta is cutting its adversarial-annotation workforce while doubling AI spending, the bet is that automated red-teaming and constitutional-AI methods are now good enough to replace humans on the hardest categories โ CSAM, self-harm, jailbreaks. That bet is industry-wide; OpenAI, Anthropic, and Google use similar mixes of automated and human red-teaming. Whether automated red-teaming actually catches the most adversarial human prompts is an open empirical question, and the failure mode is asymmetric: a missed CSAM jailbreak is a real-world harm, not just a metric. Second, the labor pattern. The people who do the worst psychological work in AI development โ looking at the worst content the internet can produce, all day โ are also the most precarious. They are contractors, not employees. They sit in low-cost geographies. They have cooldown clauses. The economics of AI safety annotation now looks like the economics of social-media content moderation a decade ago, and that did not end well for the workers either.
For builders, the practical reading is short. If you build on safety-tier APIs from any major lab, your guardrails are increasingly trained by automated systems with shrinking human oversight at the adversarial-annotation layer โ verify your own application's edge cases instead of trusting upstream safety classifiers. If you build tools that involve human-in-the-loop labeling, the precedent here is that those workers get treated as disposable; the Covalen โ November strike โ April cuts pattern is the cautionary tale, not the exception. And for anyone watching Meta specifically: the same week Meta cut adversarial annotators, it announced a near-doubling of AI capex. The signal isn't ambiguous โ Meta thinks the AI training the AI is now good enough to replace the humans training the AI. We will find out whether that is true the way we usually find out: when something goes wrong.