Overview of The Journal: How AI Is Being Trained to Do Your Job
This episode of The Journal examines the fast-growing gig economy around AI training, where experts are hired to teach large language models how to perform specialized white-collar tasks. The central example is Mercor, a startup that acts as a middleman between AI companies and human contractors—often professionals with real-world expertise in fields like writing, law, medicine, accounting, design, and consulting. The story shows both the scale of demand for this work and the uneasy reality that people are helping build systems that could eventually replace parts of their own jobs.
What Mercor Does
Mercor, founded in 2023 by three college dropouts, has grown rapidly into a major player in AI training labor.
Business model
- Mercor connects AI companies with human contractors who can help improve models.
- It has worked with major clients such as OpenAI and Anthropic.
- Contractors are hired to evaluate outputs, correct mistakes, and generate examples that help models learn more human-like behavior.
The scope of hiring
- The company posts a huge range of jobs, including:
- investment bankers
- analysts
- dermatologists
- radiologists
- therapists
- accountants
- screenwriters
- voice actors
- consultants
- The goal is to train AI to handle increasingly complex knowledge work.
Carolina Perez’s Story
A major part of the episode follows Carolina Perez, a multilingual former speech and language pathologist who found Mercor through LinkedIn.
How she got the job
- Carolina had been promoting her language and writing skills online.
- Mercor reached out because she spoke Portuguese, Spanish, and English.
- She was offered work as a “writing analyst,” training AI systems in language and writing tasks.
What the work involved
- Her first projects focused on distinguishing European Portuguese from Brazilian Portuguese.
- She had to correct language that sounded wrong for the target dialect.
- Later, she worked on more creative tasks, such as helping AI write memoir-style openings and flagging unnatural, cliché, or illogical writing.
What she noticed
- Contractors in a Slack channel began identifying recurring “AI-isms,” such as:
- overuse of obvious, generic choices
- repetitive structures
- unnatural writing conventions
- One example: AI often made everything happen on a Tuesday, as if “Tuesday” were the most neutral or average day.
Why Companies Need Human Training Data
The episode explains that AI companies have largely exhausted much of the publicly available data used to train models.
Why humans are still needed
- New, higher-quality training data is needed to make models better.
- Humans provide:
- expert judgment
- domain-specific knowledge
- corrections for subtle errors
- examples of nuanced, realistic work
The broader metaphor
- The hosts compare AI development to:
- first phase: reading every book in a library
- second phase: sitting with a tutor
- In other words, AI is moving from broad exposure to more personalized, expert-driven coaching.
Concerns, Controversies, and Risks
The episode also highlights several ethical and legal concerns around this kind of AI training work.
Intellectual property concerns
- Some contractors were reportedly asked to share prior work samples.
- That raised alarms because much of that work may belong to a former employer, not the contractor.
- Visual effects workers, for example, were reportedly asked about high-end production materials they likely did not have the rights to license.
Privacy and surveillance issues
- Mercor was hit by a major data breach and later faced multiple class-action lawsuits.
- Allegations included exposure of:
- recorded job interviews
- facial biometric data
- screenshots of workers’ computers
- The company also allegedly used software that screenshots contractors’ work during tasks.
Mercor’s response
- Mercor said it does not buy intellectual property and only licenses content from people who own the rights.
- It said it does not want materials owned by current or former employers.
- It also disputed the claims in the lawsuits and said it takes privacy seriously.
The Economics of the Work
This type of gig work can pay well at first, but the compensation appears to be declining in some cases.
Carolina’s experience
- She initially felt excited by the work and thought of it as being on the cutting edge.
- Over time, the AI got better quickly, which reduced the need for her corrections.
- Her pay also dropped:
- from about $45/hour
- to $35/hour
- then to per-task pay, around $20 per task
- She eventually quit after doing just a few tasks at the lower rate.
Mixed feelings among contractors
- Some workers see it as a way to make money in a difficult market.
- Others feel uneasy about helping build systems that may automate their own professions.
- Carolina described eventually feeling disillusioned, saying the work felt like it could contribute to a “monster” and a waste of talent.
Key Takeaways
- AI training is becoming a major labor market for professionals with specialized knowledge.
- Mercor and similar companies are scaling fast, showing strong demand from AI firms.
- Human expertise is still essential for teaching models nuance, style, and domain-specific accuracy.
- The work is temporary by nature: as models improve, they may need fewer human corrections.
- There are serious concerns about privacy, IP rights, and how much worker data is being captured and reused.
- The episode’s broader message is that AI companies are actively building tools that can do more of your job, even if the long-term outcome remains uncertain.
Bottom Line
The story frames AI training as the hidden labor behind smarter models: real experts are being paid to teach systems how to write, diagnose, design, and communicate better. But the episode also makes clear that this is not just an opportunity—it’s a preview of the same automation pressure these workers may eventually face themselves.