Hugo Vance: You know what 1964 and 2026 have in common? A law that assumes the discriminator is a person.
Lila Soto: Oh, we're starting there — okay, yeah, I'm in.
Hugo Vance: We are, because the Meta Platforms lawsuit makes it unavoidable. Does 1 through 26 — twenty-six anonymous current and former Meta employees — file in federal court in Oakland alleging that Meta's AI systems picked workers on protected medical, parental, and disability leave for termination. The AI's name, at least one of them: Metamate. And a judge — William Orrick, U.S. District Court — hears the request to halt the layoffs, says the case raises 'serious questions' about AI discrimination, and denies it. The 26 plaintiffs' layoffs were proceeding July 22nd, 2026. Five days after the ruling.
Lila Soto: Wait — he said 'serious questions' and still said no?
Hugo Vance: He did. And I think that is the confession embedded in the ruling — the court sees the harm as plausible and cannot figure out how to stop it. That is not a close procedural call. That is a doctrine that has not caught up to the technology. Title VII was written for intent. For motive. For something a jury can infer from a document or a manager's testimony. An AI productivity-scoring system that no engineer can fully explain is not that kind of defendant.
Lila Soto: So the question we're actually after is — can you even build a discrimination case when the thing that discriminated doesn't have a mind to interrogate?
Hugo Vance: That's it. And Meta laid off approximately 8,000 people — ten percent of its global workforce — beginning in May 2026. This is not a boutique case. This is the leading edge of a very large, very normalized problem.
Lila Soto: Normalized — that's the word I keep snagging on, yeah.
Hugo Vance: Normalized, yes — but I want to push back on the frame we've been using, because I think 'the law hasn't caught up' actually overshoots what the evidence shows.
Lila Soto: Oh — say more, because I think that's the thing I actually want to get at.
Hugo Vance: Disparate impact doctrine. It's already in Title VII. A facially neutral rule that disproportionately harms a protected group is unlawful — no proof of intent required. That tool exists.
Lila Soto: Right — but the part that doesn't fit is who wrote the rule. The analogy that comes to mind: imagine a company says everyone who ran a marathon last year gets promoted. That policy doesn't mention disability anywhere. But it structurally locks out people with disabilities. That's textbook disparate impact. And what Meta's AI did with keystroke monitoring and AI-token usage dashboards is almost identical — it measured activity that workers on legally protected medical or parental leave were, by definition, not performing. The system didn't target them. It just — mm — counted something they couldn't do.
Hugo Vance: The doctrine fits the harm. What it doesn't fit is the author of the harm.
Lila Soto: Exactly — and genuinely new things are happening here, I think. Princeton and University of Chicago researchers found that LLMs — ChatGPT, Claude, Gemini — can develop biases through simulated hiring experience that weren't in the original training data at all. The system stereotypes more severely than humans do, and it gets there on its own. So the question of 'who do we depose' isn't just procedurally awkward — it's that there may be no human decision anywhere in the chain that produced the discriminatory output.
Hugo Vance: Well. That is the genuinely novel part. Not that the harm is new — structural exclusion through ostensibly neutral metrics is very old — but that the bias generator is now self-modifying in ways no engineer wrote and no engineer can fully narrate.
Lila Soto: So the doctrine exists. The gap is auditability — we can't prove what the system learned after deployment, which means disparate impact becomes nearly impossible to litigate without some way to open the black box.
Hugo Vance: But here's where the arbitration clause bites — and I think this is where the hot take actually lands. The Meta case was ordered into private arbitration. Not public federal court. So Orrick's 'serious questions' language? It appears in a TRO denial. Not a binding precedent. Not case law. It vanishes into a confidential forum and the next twenty-six workers face the exact same legal vacuum.
Lila Soto: Oh — so the 'serious questions' just... disappear.
Hugo Vance: Into a settlement no one can read. Yes.
Lila Soto: Okay, and that's — mm, think about what that means for someone like Sarah. Project manager, takes three months medical leave for surgery in January, comes back in April. May rolls around, Meta's scoring algorithm runs, and her keystroke dashboard shows a flat line for those three months. She lands in the bottom quartile on activity metrics — not because she underperformed, but because the system literally measured absence the same way it measured disengagement. She files, she maybe gets a confidential settlement, and — wait, actually — no court ever rules on whether that measurement constitutes unlawful disparate impact under Title VII. The next Sarah doesn't even know that case existed.
Hugo Vance: And contrast that with Mobley v. Workday — filed in early 2023, alleging Workday's AI hiring tools discriminate by age, race, and disability under Title VII and California's FEHA. That case is in open federal court. The doctrine for algorithmic hiring is being built in public. The doctrine for algorithmic firing is being built in secret.
Lila Soto: That asymmetry is — yeah, that's genuinely alarming. Because Mobley is about the door in. Meta is about the door out. And we'll have case law for one and none for the other.
Hugo Vance: Amazon scrapped an AI recruiting algorithm — early, before any litigation forced it — because it was amplifying historical biases against women. The industry had a warning. It was not broadly internalized.
Lila Soto: And the part that comes later in this conversation makes that even more uncomfortable — because the question isn't just whether law is lagging, it's what happens when ninety percent of companies are already doing this and courts haven't set a liability standard yet.
Hugo Vance: Indeed. The conduct is normalized before the verdict arrives.
Lila Soto: And that ninety percent number — I mean, that's not a trend, that's a done deal. Manpower Group says more than ninety percent of companies are already using AI in talent acquisition. So whatever liability standard courts eventually set, it lands on ground that is already completely built over.
Hugo Vance: The retroactive problem. Yes. Conduct that is widespread and tacitly permitted today — because courts haven't stopped it — could be ruled unlawful in three years. And the workers harmed in the interim have already been let go into confidential arbitration.
Lila Soto: They paid the lag fee.
Hugo Vance: That is precisely the phrase I would use.
Lila Soto: And then — okay, this is the part that actually made my jaw drop — Netchex launched a product called Mesh on July 20th, 2026. An AI HR teammates suite, targeting deskless workforces specifically, designed to, quote, anticipate HR problems proactively. That is three days after Orrick's ruling.
Hugo Vance: The industry accelerated at the exact moment the litigation signaled exposure. That is not a coincidence — that is a market reading 'serious questions but no injunction' as a green light.
Lila Soto: Which brings us to AI transparency and auditability — the actual missing piece. Because without the ability to examine how Metamate weighted a keystroke against a protected leave period, nobody can challenge the output. Not in arbitration, not in open court. The accountability gap isn't legal — it's technical, and it's baked into the product architecture.
Hugo Vance: So the calibrated claim — the one I think actually holds — is this: the law is not broken. Disparate impact doctrine exists. The problem is the system running three years behind deployment, with no auditability requirement forcing companies to generate the evidence that would make that doctrine usable. Workers in the gap bear the cost of that distance.
Lila Soto: Which takes me back to the very first thing you said — the law assumes the discriminator is a person. Judge Orrick said 'serious questions' on July 17th. Five days later, the layoffs happened. That's not a broken system. That's a system that saw the problem clearly and had no instrument to stop it in real time.
Hugo Vance: Fine. The law isn't obsolete — it's being quietly bypassed, one arbitration clause at a time. Until courts establish what AI auditability actually requires, every worker ranked by a system they cannot see is one employment contract away from being the next Doe 1 through 26.
Lila Soto: We started with 1964 and a law built for humans. We're ending with 2026 and a judge who agreed something was wrong — and watched it happen anyway. I mean, that's the quiet rhyme, isn't it.
Hugo Vance: It is. And I find that — mm, not comforting, but clarifying. The distance between seeing a problem and having a tool to stop it has always been where people get hurt.
Lila Soto: Yeah. Good thinking-through with you today.
