High Risk Is Not a Compliance Category

What the Meta lawsuit teaches every organisation running AI on people — and what it doesn’t.

Twenty-six Meta employees are suing the company in federal court, alleging that AI-assisted systems selected them for layoff — and that the selection disproportionately hit people on medical, parental or family leave. According to the complaint, the inputs included performance ratings, calibration scores, productivity metrics, activity monitoring and even AI-token consumption. Metrics that, in the plaintiffs’ words, by design cannot be accumulated by someone on protected leave.

The company denies the claims, and a court will decide what actually happened. But the structure of the allegation deserves every HR leader’s attention, whatever the verdict — because the structure is reproducible in almost any organisation that measures people with dashboards.

I know this pattern from the inside. Over the past year, most of my technical work has been inference on unstructured data — typically deriving job levels from job descriptions and taxonomy definitions. A job level is not a technicality: level is usually one of the defining dimensions of an equal-value work group, which makes it the foundation of the pay gap calculation and of every salary benchmark built on top. Nothing decorative about it. And here is what that work teaches you fast: job descriptions are mixed-quality raw material — outdated, inconsistent, inflated, vague. The code can be excellent; it will still only ever build on what you feed it. If the input doesn’t fully describe the context, the output will be skewed — and so will every decision stacked on top of it. You cannot build a castle out of slime, however good your architect is. That’s why, in everything we build, two things are non-negotiable: the quality of the data going in, and a named human validating what comes out.

The flaw isn’t where everyone is looking

The tempting reading of this case: “the AI failed.” It’s tempting because it’s comfortable — a model can be retrained, a vendor can be replaced, a procurement box can be ticked.

But if the claims are accurate, the model did exactly what it was designed to do. It counted output. The problem is what counting output means for a person who spent four months on parental leave: their absence is recorded as underperformance — not by error, but by definition.

That is not a model flaw. It is a metric-design flaw — and this distinction matters enormously, because our entire quality-assurance reflex is aimed at the model. Accuracy tests won’t catch it: the model is accurate. Standard fairness metrics won’t reliably catch it either, if protected leave isn’t represented in the test design. You can audit the algorithm to perfection and still ship the harm, because the harm was decided upstream, when someone chose what to measure.

This is why serious algorithm-audit practice treats AI as a sociotechnical system: the model, the metrics feeding it, the people acting on it, and the process around it are one unit of analysis. Audit any single layer in isolation, and the finding is meaningless.

What “high-risk” actually means

Under the EU AI Act, AI used in employment decisions — hiring, promotion, task allocation, termination — sits in the high-risk category. And here is where I see organisations make a quiet but consequential mistake: they read “high-risk” as a paperwork tier. A label that determines how many documents you owe.

That’s backwards. High-risk means the potential harm is high-impact. In money: penalties, litigation, reputational and direct business damage. In people: rights and opportunities lost — the two harm categories that algorithm-audit taxonomies name explicitly as opportunity loss and infringement on rights. The Meta complaint, if accurate, is a textbook illustration of both at once: livelihoods lost, allegedly connected to the exercise of a legally protected right.

This is where a socio-technical lens earns its keep. In algorithm-audit practice I like BABL AI’s CIDA model — Context, Input, Decision, Action — an evaluation and transparency frame that forces you to examine all four layers of an AI system, not just the model in the middle. Map the Meta allegations onto it and the anatomy becomes clear: the Context is a mass layoff touching populations with legally protected absences; the Input is a set of metrics that, by design, cannot be accumulated on leave; the Decision is selection for termination; the Action is thousands of jobs eliminated — allegedly without the individualized review the context demanded. Notice where the harm enters: at Input, before any model runs. That is exactly the layer a model-only audit never sees — and exactly why evaluating the four layers together is not methodological pedantry, but the difference between auditing the algorithm and auditing the system.

And risk has a formula: risk = magnitude × likelihood, where magnitude = severity × scale. Eight thousand people in a single layoff round is scale. A metric that structurally penalises protected leave is severity. When both are present, likelihood barely matters — the exposure is already enterprise-grade.

The deployer thought experiment

Here is the exercise I’d run with any leadership team: imagine the same system — same metrics, same dashboards — deployed as a third-party tool inside an organisation with a working AI governance frame. What would have had to happen before go-live?

The system would appear in an inventory and receive a risk classification. Employment decisions: high-risk. That single act triggers everything downstream — nothing about the system stays informal after it.

An impact assessment would ask the uncomfortable question in advance: can any input metric, by design, disadvantage someone exercising a legal right? Leave. Disability. Accommodation. This one question, asked at design time, is this entire lawsuit avoided. Asked after deployment, it’s a settlement negotiation.

Data governance would interrogate the inputs themselves. What is token consumption actually a proxy for? What does an activity count measure during parental leave? Metrics are never neutral; they are choices, and choices can be reviewed.

Human oversight would be structural, not decorative: a defined pause-and-review rule for protected situations, with a named person accountable — judgment kept inside the organisation, not outsourced to a dashboard.

And the evidence would already exist — testing records, decision logs, deviations — before anyone asks. Documentation assembled under litigation pressure convinces no one, least of all a court.

None of this blocks AI adoption. That’s the point most commentary misses: governance done this way is not the brake — it’s what makes deployment defensible, which is what lets it survive contact with legal, with works councils, and with courts.

“But the deadlines moved”

One more objection I hear weekly, and it deserves a direct answer. Yes — the recently adopted Digital Omnibus postponed the core obligations for standalone high-risk systems under the AI Act to the end of 2027. Some transparency duties still arrive in August 2026, but the heavy compliance machinery got breathing room.

Here is what did not move: discrimination law. Family and medical leave protections. Disability rights. The Meta plaintiffs are not suing under the AI Act — they’re suing under employment law that has existed for decades, applied to algorithmic conduct. The lesson of this case is precisely that waiting for the AI-regulation deadline is a category error: the exposure exists the moment your metrics touch people, and it always has.

The deadline tells you when the regulator starts asking. The harm decides when a lawyer does.

The question that is the audit

So, one question for every organisation running productivity, performance or selection AI on people:

Could any of your metrics, by design, disadvantage someone exercising a legal right?

If nobody in your organisation has asked that question yet — that’s the audit.

If nobody has asked that question in your organisation yet — that is the audit.

Independent AI system audit →Explore FairPath →