The Meta Lawsuit Nobody's Talking About: When AI Decisions Disappear

Twenty-six Meta employees are suing over layoffs they say an algorithm shaped. The hard part of their case isn't proving the AI was biased. It's proving what the AI actually did.

I've spent years watching regulated firms adopt AI tools faster than they can document them. This Meta lawsuit is the bill coming due—and it won't be the last one.

The Real Exposure Isn't Bias

Everyone is bracing for biased AI. Chief compliance officers are asking about fairness audits. Legal teams want to know about protected class analysis. Those are real concerns, but they're the second-order problem.

The bigger exposure is AI nobody can reconstruct.

A model can be perfectly fair and still sink you in court if no one can show how a specific decision was actually made. "We used AI responsibly" is not a document you can hand a judge. When the discovery request arrives—and it will—you need to produce the actual decision path for John Smith's layoff on March 15th. Not your fairness methodology. Not your vendor's white paper. The specific chain of inputs, weightings, and outputs that led to one person's termination.

I've sat in rooms where companies couldn't answer that question about decisions made six months ago. The model's been retrained. The data's been aggregated. The vendor says the specific run isn't reproducible because they've "improved" the algorithm since then.

That's when the CFO starts doing math on settlement costs.

Sarbanes-Oxley Already Taught This Lesson

After Enron collapsed in 2001, Congress didn't ban spreadsheets and email. Those tools had enabled the fraud—analysts built elaborate models that hid billions in debt, executives coordinated the coverup through digital communications. The fix wasn't eliminating the tools. It was making the decision path provable.

Sarbanes-Oxley mandated documented controls and retained evidence. The scandal was never that people used technology to make financial decisions. It was that no one could reconstruct what happened when investigators started asking questions.

Twenty years later, we're running the same playbook with different technology.

Companies are embedding AI into employment decisions—screening resumes, ranking candidates, flagging performance issues, optimizing layoff lists—without building the equivalent of SOX controls. They're treating audit trails as compliance overhead instead of the core product of an AI employment decision.

Nobody gets fired the day the algorithm ships. The company just slowly builds unreconstructable liability.

What Regulators Already Decided

The regulators have actually settled this question—firms just aren't listening yet.

The EU AI Act treats employment AI as high-risk and requires a reconstructable trail for every decision. New York City's Local Law 144 mandates independent bias audits and retained records for automated employment decision tools. Illinois, California, and Maryland have similar laws in motion.

The through-line is identical across jurisdictions: show the decision path, don't just assert it was fair.

When I talk to firms about this, they point to their vendor's fairness documentation. That's table stakes—it proves you thought about bias in the abstract. It doesn't prove what happened to employee #4,817 in the third quartile of last year's reduction-in-force.

The uncomfortable question isn't whether your model is fair. It's whether you can prove what it did.

The Audit Trail Is the Product

Here's the reframe that compliance teams resist: in an AI employment decision, the audit trail isn't overhead. It's the deliverable.

You're not buying a recommendation engine that happens to need documentation. You're buying a documentation system that happens to generate recommendations. If you can't reconstruct the decision, you don't have a defensible HR tool—you have a liability generator with a nice UI.

I've watched firms evaluate AI hiring platforms by testing accuracy and time-to-hire metrics. Almost none ask the vendor: "If we get sued eighteen months from now, can you recreate exactly how the system scored candidate #4,291 in the applicant pool on July 12th?"

The ones who do ask get very interesting answers. Lots of hedging about "model versioning" and "approximate reconstruction" and "our general methodology shows..."

That hedge is the sound of your legal budget evaporating.

What This Looks like in Practice

Last year I was advising a client who'd implemented an AI performance review tool. Elegant system—aggregated peer feedback, analyzed communication patterns, identified flight risks. Saved managers hours of work.

Then an employee filed an EEOC complaint claiming the algorithm downgraded her rating because she took parental leave. The vendor's response? "Our model doesn't use leave status as an input." True—but also not the question. The question was: can you show what inputs DID drive her specific score, and can you prove leave-correlated factors weren't proxies?

They couldn't. Not because they'd done something wrong, but because they'd never architected the system to answer that question. The model had been retrained twice since her review. The specific input weightings weren't versioned. The decision was six months old and functionally unreconstructable.

Settlement took four months and six figures. For one employee review.

The Monday Morning Question

So here's what to ask your security team, your HR technology team, or your vendor this week:

If a court ordered us to reconstruct a single AI-influenced employment decision from six months ago—inputs, model version, weightings, output, human override if any—could we produce that documentation?

Not "do we have fairness audits." Not "does the vendor certify compliance." Can you specifically recreate the decision path for one person?

If the answer is "probably" or "we'd need to check with the vendor" or "it depends on what we still have in the logs," you already know you have a problem.

The firms that survive this wave won't be the ones with the most sophisticated AI. They'll be the ones who treated documentation as seriously as the algorithm itself.

The audit trail isn't compliance overhead. In an AI employment decision, it's the product. Everything else is just expensive litigation prep.

---

What are you seeing in your organization? Are audit trails being built into AI employment tools, or bolted on as an afterthought? I'm watching this pattern repeat across industries—would value your perspective.