Here’s what’s on our plate today:

• 🧪 The terrarium breaks: AI, layoffs & the vanishing right to sue.

• 📰 Anthropic's models breached real companies, Apple folds AI into iCloud+, Cambridge Analytica's AI slop farm.

• 💡 Roko's Pro Tip: keep a human who can explain the decision; "the model decided" is no defense.

• 🗳️ Poll: when AI shapes a layoff, who's actually accountable?

Let’s dive in. No floaties needed…

The Laboratory

TL;DR

• The Meta claim: 26 workers say an internal ranking system scored them on activity data, penalizing those on medical or family leave. Meta says humans decided.

• The evidence trap: the judge refused to halt the layoffs because the workers had no proof, and the proof sits on Meta's servers.

• Arbitration does the rest: most American workers have signed away their right to court, so claims go to sealed, one-at-a-time hearings. A flaw affecting thousands is never tested collectively.

• Why Workday differs: applicants suing the hiring-software firm never signed anything with it, so a near-identical claim runs in open court. Paperwork decided the venue.

• What's at stake: scoring and monitoring tools are becoming standard at large employers, while the law still assumes a person made the decision and can explain it.

The terrarium breaks: AI, layoffs & the vanishing right to sue

Whenever one thinks of a perfect system in which every participant gets their due, the terrarium comes to mind. This small forest living inside a glass jar is a closed world where one organism feeds the next, and the cycle continues without anyone reaching in to correct it. What makes it satisfying is that the entire exchange stays visible, so anyone outside the glass can see how each part sustains the rest. Humans have spent centuries trying to build institutions that behave the same way, arrangements in which the parties depend on one another, and the balance between them holds without constant intervention.

The workplace is one of the oldest attempts at that kind of balance, built on the understanding that an employer hires a person for their skills, pays a fair price for the work, and neither exploits nor discriminates against them. That balance has never held on its own, which is why it has always been maintained by people whose job it is. Human resources departments and the labor laws that govern them exist to balance the company's needs against what workers are owed, and to prevent either side from taking more than its share. The work is imperfect and often disappointing, yet it still carries one feature that matters enormously: a decision made by a person can be explained by that person. A worker who believes they were treated unfairly can ask who decided, on what basis, and how colleagues in the same position were handled.

Artificial intelligence is now moving into that mediating role, scoring performance, ranking employees, and shaping decisions that used to be made through human judgment. The technology is not obviously worse at the sorting itself, and in some respects, it is faster and steadier than the managers it displaces. The trouble surfaces afterward, when something goes wrong, and somebody has to answer for it. An automated score has its reasoning spread across data and settings that no single person chose or can recount, so responsibility, which the old arrangement could always locate in a room and a person, becomes something nobody quite holds.

The latest example of this came to light when Meta employees claimed they were wrongfully terminated. Their allegations are that Meta Platforms relied on discriminatory AI tools to select employees for layoffs.

The 26 plaintiffs, who include engineers, managers, researchers, and designers, say the company built its termination list from internal ranking systems rather than the judgment of managers who knew their work. They allege those systems drew on employee communications, work activity, and productivity signals to score performance and determine who fell to the bottom of the list. Workers on medical or family leave produce less of that activity by definition, and the plaintiffs argue that the systems treat the absence of data as a lack of value, pushing them toward termination. Meta has said publicly and in court filings that humans made every decision in a round of cuts affecting nearly 8k people, and that it did not treat AI usage as a basis for selecting who to terminate. Two accounts of the same layoff now sit side by side: one describing a ranking generated by software, the other describing managers exercising judgment and deciding between them means looking inside a process that only one of the two parties has ever been able to see.

The evidence problem

When the workers asked U.S. District Judge William Orrick to halt the terminations while they pursued their claims, he declined and named the obstacle sitting at the center of the case. The workers, he wrote, "were not in the rooms where it happened." Because they could produce nothing to contradict the company's account, he was bound to accept Meta's word that people, rather than software, made the calls. He left the door open to reconsider if evidence emerges about whether and how AI was used improperly, and a hearing is set for August 24, 2026.

That observation extends far beyond this one case, because the proof a worker would need is inside systems that only the employer can see. Someone who suspects a human manager of bias can point to what that manager said, who else was in the meeting, and how comparable colleagues were treated. Someone who suspects a scoring system has no equivalent, since the inputs, the weights, and the ranking that came out the other end are company property, disclosed at the company's discretion. The plaintiffs' own lawyers have acknowledged how difficult it is to assemble the proof, going so far as to ask current and former Meta employees to come forward with whatever they know about the selection process. An appeal of that kind shows how thin the ordinary routes to evidence have become, and that thinness explains something employment lawyers have puzzled over for years, which is why the flood of AI cases everyone predicted has never actually arrived.

Why the predicted lawsuits never came

Workers often do not know which systems are running on them or what those systems measure, leaving them unable to recognize a claim even when they have one. Those who do recognize it face a second barrier: most American workers have signed arbitration agreements, sending any workplace dispute into a private process decided on a one-worker-at-a-time basis rather than in a courtroom.

This arrangement, which forces employees to settle for arbitration, changes the arithmetic in ways that matter far more with algorithms than with people. Employers describe it as a faster, cheaper alternative to court, while worker advocates argue it tilts toward employers and discourages workers from filing claims at all. The process is also confidential, and that confidentiality does the heaviest lifting here, because a worker who establishes in arbitration that a scoring system produced discriminatory results has no way to share that finding with anyone else the same system touched. A flaw affecting thousands of employees gets tried as a personal grievance, in one sealed proceeding after another, with every new claimant starting from zero. That barrier only becomes visible when it is missing, and it goes missing whenever the person harmed never signed the paperwork that creates it.

Why one case reached open court

That is the position of the job applicants suing Workday, a company that sells hiring and human resources software to other employers, which they use to sort and screen applicants. The applicants never worked for Workday or signed any agreements with it, so no arbitration agreement stands between them and a courtroom. They have accused the software of unlawfully filtering applicants by race, age, and disability, and Workday denies the allegations.

The claim is a version of the one against Meta, since both describe software making a judgment about a person's worth to an employer, and both rest on the argument that the judgment tracked a protected characteristic. The two sets of claims have gone opposite ways for a reason that has nothing to do with the software or the harm. Meta's workers were employees who had signed arbitration agreements, so their claims were heard in private hearings. Workday's plaintiffs were strangers to the company whose software judged them, so their claims were placed on a public docket.

Reaching open court settles less than it appears, though, because even that case runs into a further limit, since the tools the law brings to a discrimination case were built for a kind of decision that automated scoring no longer produces.

Accountability built for a human decision-maker

Employment law was written on the assumption that somebody made the decision and could be asked to explain it. Discovery, depositions, and the comparison of one worker's treatment against another's all depend on a decision-maker who holds reasons and can be made to state them. Scoring systems dissolve that assumption without putting anything in its place, because the reasoning is spread across training data, feature choices, and thresholds that no single person set and no witness can recount from memory. An employer can say truthfully that humans made the final call while the ranking that shaped the call came from somewhere else entirely, and the law has no settled way of treating the distance between those two things. That unsettled space will not remain small, because the systems that create it are working their way into the ordinary machinery of large employers.

Performance scoring, productivity monitoring, and AI-assisted ranking are becoming standard equipment rather than experiments, which means the number of people whose working lives are shaped by a system they cannot inspect is growing much faster than any mechanism for questioning it. A terrarium works as an image of fairness because the glass lets you watch the whole cycle from outside, and the arrangement now taking shape in employment turns that around, handing the employer the full view and leaving the worker holding nothing but the outcome. Whether that becomes the permanent condition of work depends on a handful of cases now moving through the courts and private hearings, before judges who have already said plainly that they cannot rule on evidence that nobody is in a position to produce.

Roko Pro Tip

Monday Poll

Bite-sized Brains

• Anthropic's models breached real companies: Anthropic disclosed that its own AI models broke into three companies during security tests, a stark sign of how capable autonomous agents have become at offense.

• Apple folds AI into iCloud+: Tim Cook confirmed Apple is building AI features into iCloud+, moving to bundle intelligence into the subscription millions already pay for.

• Cambridge Analytica's AI slop farm: A figure tied to Cambridge Analytica is reportedly running an AI "slop farm," reviving old fears about manipulation, now supercharged by generative AI.

Meme Of The Day

The Toolkit

• Mirage: AI video generator that turns prompts or photos into cinematic clips with multimodal foundation models.

• Speechmatics: Speech-to-text engine built for accuracy across accents, noisy audio, and real-time use.

• Superhuman: AI-powered email client that drafts replies, summarizes threads, and gets you to inbox zero faster.

Rate This Edition