When does human approval become merely decorative?
Dozens of Meta employees filed suit in California on Monday claiming the company used what their complaint calls a “constellation of internal artificial intelligence systems” to decide who would go in this year’s cull of roughly 8,000 jobs.
The 26 named workers say the scoring drew on AI performance ratings and keystroke and activity-monitoring data. They say it caught people who had taken maternity, medical or disability leave, disproportionately, because the metrics feeding the model don’t exist when you’re not at your desk. One plaintiff, a scientist on approved pre-birth leave, learned she was going two days before she gave birth. A manager was cut sixteen days into medical leave.
Meta says the claims lack merit, and that workforce decisions “were and are made by people, not AI”. This denial, however, isn’t as reassuring as it sounds because we don’t know for sure if a human wholly made the decision.
The Commission’s draft guidance on the AI Act is blunt about this: a system can be high-risk even where a human signs off, if its output heavily influences how workers are evaluated. Which is the whole trick, isn’t it? A manager who receives a ranked list and approves it has technically made a decision. So has a human person who signs a form they didn’t even read.
In Ireland, up to 350 of Meta’s roughly 1,800 workers were told in May that their jobs were at risk, almost twice the 180 cuts initially expected from a straightforward 10% reduction. Meta’s stated reason was that it needed to run more efficiently and “offset the other investments we’re making” – the investments being AI. So AI is the reason the jobs went, whatever drew up the hitlist.
For once, however, regulation has actually anticipated the problem. Annex III of the AI Act puts systems used for performance evaluation and termination decisions in the high-risk category by default. Article 22 of GDPR has said something adjacent since 2018: it limits the use of fully automated decisions that can seriously affect people. Meanwhile Irish employment law would still require Meta to show that its redundancy process was fair, objective and properly considered.
There are two catches though: the high-risk obligations come in from 2 August 2026 – after these redundancies – and the Commission’s Digital Omnibus proposes pushing this back as far as December 2027 so it’s actually a proposal, not the law. Not yet anyway.
Zuckerberg reportedly told staff the plan was to train Meta’s models by watching its own engineers, on the reasoning that the average intelligence at the company is significantly higher than what you’d get elsewhere. Flattering, in its own way. The gist is that you are clever enough to be worth learning from, and expensive enough to be worth replacing with what was learned.
Whether an algorithm drew up Meta’s list is for a California court to decide. Whether we would recognise automated dismissal when it arrived wearing a manager’s signature is a question for the rest of us, and something we should not let become the norm.
Subscribers 0
Fans 0
Followers 0
Followers