After days of the usual flood of irrelevant notifications, an offer finally appeared in my LinkedIn feed that actually matched my profile. A well-known AI and data-hiring platform called Mercor was looking for translation specialists to evaluate advanced language models.
The listing advertised a surprisingly good rate—fifty dollars an hour—for an Argentinian Spanish Audio Generalist Evaluator Expert. But it wasn’t the pay that caught my attention. It was something subtler. Right there beneath the title, a small note pointed out that three graduates from my old university were already working at the company.
In a vast internet full of recruitment bots, ghost jobs, and automated emails, seeing the name of your old university creates an immediate sense of familiarity. It doesn’t establish trust. It manufactures the feeling of trust. The recommendation isn’t coming from a person. It’s coming from a system designed to know which small details might lower your guard.
Despite my reservations, I decided to try one more time. I uploaded my résumé, connected my professional profile, and started filling out the registration form.
Before I could get through sign-up, the system stalled. I clicked to log in and was stopped cold. A plain box appeared on the screen: “You need to be signed in to continue.” Below it, in red: “Something went wrong. Please try again later.”
It was a small but revealing irony. A company building AI systems to evaluate human expertise couldn’t reliably confirm who a new applicant even was. At the time, I dismissed it as just another online hiccup. Later it looked less like a temporary glitch than a hint that something deeper had failed.
Once I finally got through, the real test of compliance began. The registration form demanded my date of birth. A date of birth seems harmless on its own, but becomes powerful when combined with names, email addresses, employment histories, and identity documents.
I hesitated. That hesitation itself told the story. In an era when corporate security breaches are just background noise, handing over personal information to a platform I knew almost nothing about didn’t sit right with me. After decades in this industry, I’ve learned to recognize when a risk runs only one way.
It was asking for pieces of my identity that a platform I barely knew hadn’t earned the right to see. I decided how much to reveal, submitted the application, and closed the tab.
But the unease stayed with me. A few days later, I opened a clean browser window, turned again to an AI engine, and started looking into the very system that was supposed to be evaluating me.
What I expected to find were ordinary complaints from frustrated applicants or scattered reviews about the hiring process. Instead, the search results were dominated by something else entirely.
What unfolded on my screen was unsettling—not because of complicated technical language, but because of how far it went. A serious security flaw had exposed the system. On the dark web, four terabytes of stolen data were already being offered for sale.
The first reports described a massive database packed with the login details of tens of thousands of independent contractors. But that was only part of the breach. Other data reportedly included high-definition video interviews and facial biometric information from human experts who had sat in front of their webcams to prove their competence to an automated evaluation system. Major labs had already suspended their work with the company indefinitely as they tried to figure out how bad the damage was.
That was when I realized how close I had come to handing over the kind of identity and biometric data contained in that four-terabyte breach.
The asymmetry was absolute.
I still remembered that plain warning box from days earlier. On the front end, they demand your documents, your compliance, your blind trust, treating you like a suspect before you’ve even done any work. But on the back end, the security of it all looked remarkably fragile. Every verification step flowed in one direction. Applicants had to prove who they were. They handed over passports, interviews, signatures, recordings, and birth dates. Yet almost nothing flowed back. There was no equivalent transparency about how the data would be protected, who could access it, how long it would remain, or what would happen if it all failed.
Every application claimed to value my expertise. Yet with every one, the balance shifted—the work stayed hypothetical while the requests became more concrete. The login failure, the birth-date field, the biometric interview—by then, the original question about language had almost disappeared. Before anyone had asked me to translate a single sentence, I was already deciding how much of myself I was willing to hand over.
Until then, I had assumed I was the one being evaluated. The real evaluation, I realized, was happening in the opposite direction. Expertise had never been the only thing they were collecting. The question was no longer whether I was qualified enough to get in, but whether these systems had demonstrated they deserved my trust.
I opened the platform. Went to my profile settings. Clicked “Delete.”
For the first time in months, the decision belonged entirely to me.
Author’s note: This essay refers to the March 2026 Mercor security incident, which the company attributed to a breach at a supplier—specifically the open-source AI tool LiteLLM. Public reporting and legal filings—including a class-action lawsuit, Ananthula et al. v. Mercor.io Corporation et al., No. 3:26-cv-03362, U.S. District Court for the Northern District of California—describe claims that roughly four terabytes of data were taken, including contractor records, login credentials, source code, and interview materials containing biometric information. What matters for this essay isn’t the breach on its own, but the gap between how much personal information applicants were asked to hand over and how hard it is to find out whether that information was actually protected.
This conversation continues in Part V.
Technology is the setting. Humanity is the subject.