July 14, 2026 · 9 min read · Sallyann Della Casa

What Went Wrong at Mercor, in Three Acts

Why the only people who care about human in the loop are your AI governance team, and why that should terrify everyone else.

A theatre stage in three acts: rows of identical gig workers dissolving offstage while a single spotlight holds one expert whose glowing thread of judgement stays connected.

Key Takeaways

  • Human-in-the-loop is quietly one of the largest new labor markets on earth. RLHF, evals, expert data. Mercor alone was tracking toward $1 billion in annualized revenue at a $10 billion valuation before its 2026 crisis.
  • AI labs treat human judgement as their scarcest input and pay up to $100 an hour for it. Enterprises treat the exact same thing as compliance paperwork. Both are getting it wrong.
  • Mercor's unraveling came in three acts: contractor wages slashed from $21 to $16 an hour in November 2025, a supply-chain breach exposing roughly 40,000 contractors in April 2026, and a wave of mass offboarding from its largest project in May 2026.
  • The root cause was not one hack. It was the model: renting judgement by the hour, at the lowest clearing price, with no ownership, no continuity, no stake. The gig model that broke drivers is now being applied to PhDs.
  • The alternative is judgement as a documented, attributable asset. Under GLEAC's current Terms, payment is guaranteed for micro-practices scoring above 8/10 completed without any AI usage: $100 per year for up to 10 years for each data-lake dataset licensed to a third party that includes them; no separate revenue share applies.

Have You Noticed Who Actually Cares About Human in the Loop?

Have you noticed the only people who care about judgement and human in the loop are the AI governance people?

Sit in any leadership meeting this year. The CEO wants agents. The CFO wants headcount savings. Product wants speed. The only person in the room asking "where does a human check this?" is the governance lead, and everyone quietly checks their phone when she does.

Here is what that room is missing. While they filed human-in-the-loop under paperwork, it became one of the largest new labor markets on earth.

And in 2026, the biggest company in that market gave us a masterclass in how NOT to run it.

How Big Is the Human-in-the-Loop Labor Market?

Human-in-the-loop work, meaning RLHF, model evaluation, and expert training data, is now a multi-billion dollar labor market employing hundreds of thousands of people worldwide, from data annotators to PhDs and physicians paid up to $100 an hour to teach AI models judgement.

Mercor, the Thiel-backed talent marketplace founded by three 22-year-olds, was tracking toward $1 billion in annualized revenue at a $10 billion valuation. One Meta video-labeling project alone employed over 5,000 contractors at its peak. Another project, codenamed Project L, cycled through more than 11,000.

Read those numbers again. This is not a checkbox. This is an industry.

The frontier labs figured out something most enterprises still have not: the models are commoditising, and the scarcest input left is human judgement. They are paying billions for it.

Meanwhile, your organisation has judgement walking its hallways every single day, exercising itself in every exception, override, and escalation. And the only department that even looks at it is looking for liability, not value.

So how did the market leader stumble? In three acts.

What Went Wrong at Mercor?

Act One: The Wage Cut (November 2025)

Weeks after Mercor's $350 million Series C made its founders the world's youngest self-made billionaires, contractors on its Meta project saw their pay slashed from $21 to $16 an hour as work moved to a new project.

One worker said it perfectly: "I know we are working with AI but we don't work for AI. You don't just dump thousands of people."

Frame that quote. It is the entire labor politics of the AI era in two sentences.

Act Two: The Breach (April 2026)

In March 2026, hackers slipped credential-harvesting malware into LiteLLM, an open-source tool in Mercor's stack. Forty minutes of exposure was enough. Attackers claimed 4 terabytes of data: candidate profiles, employer data, source code, API keys. Roughly 40,000 contractors were exposed. Meta paused its contracts. Five contractors filed lawsuits.

And here is the detail I cannot stop thinking about. Part of the security assurance in that chain traced back to Delve, a compliance startup later accused by whistleblowers of faking its certification data.

The checkbox company was itself a checkbox. Governance theater, certifying governance theater. THIS is what happens when an industry treats trust as paperwork.

Act Three: The Offboarding (May 2026)

Two months after the breach, roughly 800 workers were offboarded in a single wave from Project L. To Mercor's credit, people got paid for onboarding, some handsomely. But the shape of the story was the same shape as Act One: scale up thousands of humans, extract the judgement, release.

Add the fraud pressure documented by Forbes, including North Korean operatives attempting to infiltrate the contractor pool, and you have the full picture of a system straining at every seam.

The Diagnosis: This Is a Model Failure, Not a Hack

Let me be clear about what I am saying and what I am not.

I am not saying Mercor's founders are villains. They built the biggest company in a market everyone else ignored, and they moved fast. Huge respect for the ambition.

I am saying the MODEL is broken. Every act of this story comes from the same root: treating the scarcest input on earth like disposable labor.

Rent judgement by the hour. Source it at the lowest clearing price. Give the humans no ownership, no continuity, no stake in what their judgement builds. Then act surprised when you get churn, fraud, infiltration, and 40,000 people's data treated as an afterthought.

We have seen this movie. It is the gig economy playbook, except this time it is not being applied to delivery drivers. It is being applied to physicians, lawyers, and engineers. The gig model that broke drivers is now coming for PhDs, and it will break the same way, because judgement is not a commodity you mine. It is an asset you steward.

The Enterprise Mirror: You Are Making the Opposite Mistake

Now stop looking at Mercor and look in the mirror, because enterprises are running the same blindness in reverse.

The labs pay billions for judgement but will not let the humans own any of it. Your organisation OWNS oceans of judgement and does not even know it is an asset. Every exception your best operator approves, every policy your team bends for good reason, every escalation resolved by someone's 20 years of instinct: that is exactly the material AI labs are paying $100 an hour to synthesize from strangers.

You are sitting on the thing Mercor sells. And you have filed it under compliance.

Stop reading and do this now. Ask yourself one question: if a frontier lab offered to license the accumulated judgement of your top 20 percent of people, what would it be worth? If you cannot answer, you have never valued your judgement layer. Your governance team is the only team even close to the topic, and they were handed it as a risk, not a treasury.

Frequently Asked Questions

What went wrong at Mercor?

Mercor, the $10 billion AI talent marketplace, faced three crises within seven months: contractor wages cut from $21 to $16 an hour in November 2025, a supply-chain data breach in April 2026 that exposed roughly 40,000 contractors and prompted Meta to pause contracts, and a mass offboarding of around 800 workers from its largest project in May 2026. Five contractors filed lawsuits over the breach.

How big is the human-in-the-loop labor market?

Human-in-the-loop work (RLHF, model evaluation, and expert training data) is a multi-billion dollar global labor market. Mercor alone was tracking toward $1 billion in annualized revenue, with single projects employing 5,000 to 11,000 contractors, and expert contributors earning up to $100 an hour.

Why do AI labs pay so much for human judgement?

Because models commoditise and judgement does not. As AI systems converge in capability, the differentiating input is high-quality human judgement: expert reasoning, evaluations, and decision-making patterns that models cannot generate for themselves. That scarcity is why frontier labs spend billions on human-in-the-loop infrastructure.

Should enterprises care about human in the loop beyond compliance?

Yes. Every enterprise already produces the exact judgement AI labs pay billions for, through daily exceptions, overrides, and escalations by experienced operators. Treating human-in-the-loop purely as a governance checkbox means owning a valuable asset and never putting it on the balance sheet.

What is the alternative to hourly AI training gig work?

Contribution-based models. Under GLEAC's current Terms, payment is guaranteed for micro-practices scoring above 8/10 completed without any AI usage: $100 per year for up to 10 years for each qualifying third-party dataset licence. The payment is tied to the eligible practice and dataset, not the shift; no separate revenue share applies. Apply through the 100-Day AI Accelerator at lovelyhumans.ai.

Your Move

The governance people were right to care about human in the loop. They were just the only ones, and they were handed the topic upside down.

Judgement is not your liability to minimise. It is your scarcest asset to own.

Go ask the licensing question about your top 20 percent. Then come back and tell me the number, because I promise it is bigger than whatever your AI pilot budget was this year.

If curious, tell me what is true.

Sallyann Della Casa, Founder and CEO, GLEAC / Lovely Humans