When AI Decides Who Gets Fired
2026-07-19 · 5 min read · Janaina Maia
Twenty-six Meta employees are suing the company for allegedly using AI to decide who gets laid off. Their claim: Meta's internal AI tools — including a system called "Metamate," employee-monitoring dashboards, and algorithmic performance scoring — were used to select which workers to terminate. The selected employees were disproportionately on disability or family leave. Meta says humans made the decisions. The employees say the AI did.
Ars Technica reported that this is apparently the first lawsuit against a major US company challenging the use of AI in conducting layoffs. That milestone alone should get every product designer's attention. But the design implications run deeper than the legal ones.
What the lawsuit alleges
The complaint says Meta used a combination of internal AI systems to score, rank, and select employees for its May 2026 layoffs of roughly 8,000 workers. The inputs included performance ratings, "calibration scores," productivity metrics, keystroke monitoring, AI token usage, and an internal classification that labelled employees as "AI Native," "AI First," or "AI Enabled" — a ranking of how enthusiastically they adopted Meta's own AI tools.
The problem: if you score people on AI tool adoption, keystroke activity, and output volume, you are systematically penalising anyone who took medical leave, maternity leave, or disability accommodation. They could not accumulate AI token usage or keystrokes during leave. The scoring system did not adjust for this. The result, the plaintiffs allege, is that workers exercising legally protected rights were disproportionately flagged for termination.
One employee was told she was being laid off the day before her water broke. Two days before she gave birth.
This is a product design problem, not just a legal one
The lawyers will argue about whether Meta used AI or humans to make the final call. But for product teams, that distinction is almost beside the point. The real question is: what happens when you build a scoring system that cannot account for protected leave, disability, or accommodation?
Think about it as a design decision. Someone designed the inputs. Someone chose the weights. Someone decided that "AI Native" was a positive signal and that low activity was a negative signal. Someone decided not to neutralise the scoring for people on leave. Each of those is a product design decision that produced a discriminatory outcome — whether or not anyone intended it.
This is exactly what I wrote about when I said AI privacy is also a self-curation problem. Systems that infer from behavioural exhaust — keystrokes, tool usage, activity logs — will always over-index on the people who are always-on, always-present, always-producing. They will systematically undervalue the people who took time away for health, family, or recovery. The "objective" data is not objective. It reflects who had the privilege of being continuously present.
The "AI Native" classification is a red flag
The lawsuit says Meta categorised employees by AI adoption levels: "AI Native," "AI First," and "AI Enabled." This is worth pausing on, because it turns AI adoption from a personal productivity choice into a career-defining classification.
Imagine being an employee who took parental leave, came back, and discovered that your "AI Native" score had decayed to "AI Enabled" because you were away during the rollout period. Your productivity metrics dropped because you were recovering from childbirth. Your keystrokes were low because you were on approved leave. And now a system you never consented to feed has flagged you as lower-performing.
Classification systems in enterprise products are not neutral infrastructure. They are power structures. When a company categorises its own workers by how enthusiastically they adopt the company's AI tools, it creates a feedback loop: adopt our AI quickly or your career metrics suffer. That is not measuring performance. That is measuring compliance.
What product teams should ask
- Does your scoring system account for legally protected status? If you are scoring employees on output, activity, or tool usage, you must ask whether those metrics systematically disadvantage people on leave, with disabilities, or with accommodations. If you do not adjust, you are building discrimination into the math.
- Who designed the inputs, and who reviewed the weights? Every input in a scoring system is a product decision. Someone chose to include AI adoption. Someone chose to include keystroke frequency. Someone chose not to neutralise those inputs for protected leave. These are not neutral choices. They need the same design review you would give any feature that affects user outcomes.
- Can the people being scored see and challenge the system? If an AI system is influencing someone's employment, they need to know what data it used, what weights it applied, and how to correct or contest the result. This is not a radical demand. Credit scoring has had legal requirements for transparency and correction for decades. Employment scoring should be no different.
- Is "human in the loop" actually human? If a human rubber-stamps an AI-generated ranking without understanding or adjusting the inputs, the human is not really in the loop. They are a liability shield. Real human oversight means the human understands the system, can override it, and is accountable for the override.
- Are you measuring compliance or performance? If your metrics reward people for using your company's own AI tools, you are measuring adoption, not impact. These are different things. An employee who produces excellent work without AI tools should not be penalised for low AI token usage. That metric says more about the company's AI rollout than the employee's contribution.
The broader pattern this week
This lawsuit lands in a week already full of AI accountability signals. Germany ruled that AI Overviews are media content, not search results, making platforms liable for what their models produce. The Meta Oversight Board found that leading LLMs systematically refuse to criticise authoritarian governments while freely criticising democracies. New York became the first US state to ban data center construction, citing the cost to residents. Linus Torvalds told open-source AI critics to "fork it or walk away," drawing a line on how far AI scepticism should go in practical engineering.
The connecting thread: every one of these stories is about who is accountable when AI systems produce outcomes that affect real people. The German court said the platform is accountable. The Oversight Board said product teams are accountable. The New York government said the infrastructure companies are accountable. And now twenty-six employees are asking: when an AI system flags you for termination, who is accountable?
My take
I wrote recently that AI does not make work worthless because it makes work easier. This lawsuit is the flip side of that argument: AI does not make decisions fair because it makes decisions efficiently. An algorithm can discriminate just as effectively as a person — and in some ways more efficiently, because the discrimination is hidden inside a scoring system that looks objective.
The product lesson is straightforward: if you build a system that scores, ranks, or selects people, you need to design for the people the system will disadvantage. Not after a lawsuit. During the design phase. The inputs need review. The weights need review. The outputs need to be tested for disparate impact. And the people affected need a way to see, understand, and challenge the result.
Meta may win this case on the legal question of whether humans or AI made the final call. But the design question will outlast the verdict. Scoring systems that penalise protected leave are not neutral. Classification systems that reward AI adoption are not measuring performance. And "human in the loop" is not accountability if the human does not understand what the loop is doing.
If you are building AI that affects people's lives — and in enterprise product, you almost certainly are — this case is your warning shot. The question is not whether AI can make employment decisions. It is whether the product team designed the decision surface carefully enough that those decisions are fair, transparent, and contestable. If you did not, you have already built the discrimination in. You just have not been sued yet.