Somewhere in your training pipeline, a person listed as 'senior financial analyst' is grading model outputs about revenue recognition. What that title actually asserts is: someone typed it into a form once, and nobody was in a position to check.
Claims vs. records
There are two kinds of information about a worker's background, and they are not the same species. Claims: resume lines, profile titles, checkbox skills, self-rated fluency. Records: the role written in an employment contract, the months payroll actually ran, the projects that renewed or didn't, the language the person worked in daily because their job required it.
Marketplace vetting is built almost entirely on claims, lightly laundered through qualification tests. Employment data is records. We hold records — for a meaningful share of a 15M-professional network, across 120 countries, accumulated over ten-plus years of being the employer.
Where the gap bites
- Domain evals: a 'nurse' who was actually a hospital receptionist grades clinical safety very differently from an RN with six validated ward years.
- Language work: self-rated C2 fluency is the most inflated metric on the internet; payroll in São Paulo is not.
- Tenure and consistency: the person who stayed four years somewhere behaves differently in week ten of your project than the one who's never stayed four months.
You wouldn't train on a dataset whose labels were self-reported. Your workforce's credentials are labels.
None of this makes marketplaces evil — it makes them unverifiable. For casual tasks that's a fine trade. For the data teaching frontier models medicine, law, and machine safety, it isn't.
We were the employer