Skip to main content

Train-Your-Replacement Work

If the job is to generate the footage a robot learns from, the robot is the output. You are the bridge. When it crosses, the bridge comes down.

The Pattern​

A new category of jobs has quietly become one of the largest employers of the AI era. The work looks mundane on the surface: wear a camera while you fold towels, babysit a robotaxi from the passenger seat, take over a drive-thru order when the voice AI gets confused, label 10,000 images so a model learns what a stop sign looks like in the rain. The job description rarely names what is actually being produced. What is being produced is a dataset. The dataset trains a system. The system replaces the job.

This is train-your-replacement work.

Some examples from the current moment:

Humanoid-robot motion capture. Objectways, a 2,200-person annotation company in Tamil Nadu, pays Indian workers roughly $230 to $250 per month to record 60-second clips of household tasks. As of late 2025 they had annotated over 15,000 videos of humanoid robots folding clothes for U.S. clients. One worker, Kavin, told reporters: "In five or 10 years, they'll be able to do all the jobs and there will be none left for us." (TechXplore, Nov 2025; CNN Business, Apr 2026)

"Hand farms" at global scale. Palo Alto contractor Micro1 runs 4,000 camera-wearers across 71 countries and ships 160,000+ hours of training footage per month to undisclosed robotics clients. Scale AI has announced 100,000+ hours of its own humanoid training video. The industry is spending more than $100 million a year on this labor, and the workers rarely know which client is buying their footage or what it is being used for. (MIT Technology Review, Apr 2026)

Robotaxi safety monitors and remote operators. Waymo's Fleet Response and Remote Assistance Operators earn $24 to $27 an hour, often through staffing agencies rather than W2 roles. Tesla Robotaxi safety monitors ride in the passenger seat and have in at least one documented case been forced to climb into the driver's seat to take over. Seven robotaxi operators (Tesla, Zoox, Nuro among them) have disclosed to regulators that humans remain a core part of their purportedly autonomous systems. (CNBC, Dec 2025; Futurism)

RLHF labelers and content moderators. Outlier (Scale AI) pays domain experts $30 to $50 an hour; entry-level labelers make $15 to $20. In Kenya, OpenAI's contractor Sama paid workers roughly $2 an hour, out of $12.50 per worker billed to OpenAI. Meta's Kenyan moderation operation, also run through Sama, ended in 2025 with 1,108 redundancies and a class action from more than 185 former moderators seeking over $1.6 billion. 81% of a 144-person assessed subset showed severe PTSD. (Canadian Affairs; CBS 60 Minutes)

Drive-thru voice AI "escalation agents." Presto's SEC filings disclose that roughly 70% of AI drive-thru orders require human intervention. Bite Ninja runs approximately 12,000 remote workers handling drive-thru orders at pay "roughly comparable to QSR pay." These are the humans the voice AI is learning from. (CNBC)

Warehouse and kitchen teleoperators. Agility's Digit humanoid passed 100,000 totes at GXO's Flowery Branch warehouse in November 2025. The training pipeline blends "teleoperated demonstrations, policy training, reinforcement learning, and simulation." Translation: humans run the robot until the robot doesn't need them. (Robotics & Automation News)

Why Companies Do It (And Why That's Fine)​

Running a human operator in parallel with the model is how every physical-AI system gets good. You cannot train a folding robot without watching humans fold. You cannot ship a self-driving car without a fleet of remote operators catching edge cases. You cannot launch a voice AI without a human escalation path for the 70% of calls it cannot yet handle. The companies doing this are acting rationally. Bridge labor is a standard phase of any automation project.

This page is not an argument against those companies. They are going to do what they are going to do. This page is about what you should do if you find yourself inside one of these roles.

Why This Is Not a Career​

The defining feature of train-your-replacement work is that the job's success metric is the elimination of the job. Every clip you record, every order you escalate, every takeover you perform is a data point that shortens the runway to the version of the system that does not need you. You are not accumulating transferable skills. You are not building relationships that compound. You are not learning judgment that becomes more valuable with time. You are producing training data, at a price set by the client, for a buyer whose explicit goal is to not need you.

Three ways to tell you are in this category:

  1. The job description is defined by the machine's current limitation. "Help the AI when it gets stuck." "Record yourself doing X." "Label the data." When the limitation closes, the role closes with it.
  2. The buyer's spend on your labor is a line item in an R&D budget, not an operations budget. R&D budgets are time-limited by design.
  3. No one can tell you what the role looks like in three years. If the company's roadmap is "and then the AI handles it," you are the bridge.

This overlaps with Robot Mode but is sharper. Robot Mode is "the job reduces you to a machine." Train-your-replacement work is "the job exists to build the machine." The first can be automated around. The second is actively being automated against.

Pick Work Where Humans Are Essential Long Term​

The practical move is to shift into roles that remain valuable on the far side of the current automation wave. What that looks like in 2026:

  • Building and operating the systems, not feeding them. An AGI Whisperer or harness builder goes up in value as models get better, because a better model makes their system more powerful. A labeler goes to zero.
  • Work that requires a body and a specific place. Skilled trades, in-person care, cultural pattern-matching in a particular city. AI learns from text. Tacit and hyper-local knowledge stays earned on the ground. (See RIP To The Career Ladder for the full argument.)
  • Relationship and trust work. Being someone's go-to person, closing deals that depend on someone believing in you, leading a community that shows up because you show up. The ambient level of slop is rising; the premium on a real human relationship is rising with it.
  • Taste, judgment, strategic thinking. The ability to look at a system's output and say this is not quite right, and here is what would make it true. The practitioners pulling ahead are the ones who train this muscle deliberately. (See Strategy Is the New Execution.)
  • Teaching others to make the transition. The Community Leader, the trainer, the activator. Every person who needs to cross the line will need a human to help them cross it.

If your current job is in train-your-replacement work, the move is to use the paycheck to buy the time and tools to build something that is not. Learn a harness. Start a personal agentic OS. Take the Supersuit Up workshop. Find a local chapter. Get Jarvised before the bridge comes down.

A Note on Seeing It From Above​

Most of these jobs do not announce themselves. The worker in a 60-second cleaning clip may not know the clip is flying to a Silicon Valley training cluster. The worker in a drive-thru headset may not know they are the 70% that the voice AI has not figured out yet. One public resource worth knowing about is Conscious Spend, which rates companies on environmental impact, labor practices, and transparency. It is one way to see what the supply chain is doing with its hands when the marketing copy is talking.

If you work in one of these roles, the point of this page is not to make you feel bad. It is to be honest about what the job is so you can make an informed choice. The companies are being rational. Be rational back. The bridge is designed to come down. Make sure you are on the other side when it does.

Further Reading​

  • Robot Mode: The adjacent pattern. Robot Mode is work that reduces you to a machine. Train-your-replacement work is work that exists to build one.
  • The Survivor Economy: The inside-a-company version of the same sort. Adapt or get written out of the org chart.
  • RIP To The Career Ladder: The research-backed view of which rungs are collapsing and what replaces them.
  • Your Two Futures: The daily decision underneath all of this.
  • The Hyperagency Gap: The widening split between people who wrap AI around themselves and people whose jobs wrap around AI.
  • Roles to Workflows: How roles decompose into workflows, and how to see which workflows are on the chopping block.
  • There Is No Demand for Average: Why baseline work commoditizes and where demand concentrates.
  • Liberation Architecture: The positive construction. Design systems that liberate humans into higher-order work.
  • Supersuit Up Workshop: The practical on-ramp for crossing the bridge.
  • Conscious Spend: External resource. Company ratings on environmental impact, labor, and transparency.