Why Workers Who Train Their AI Replacement Get Paid More

A worker asked to train the AI that will take their job has a third option besides working and quitting: keep producing while quietly starving the system of data. A new NBER model shows that option sets the pace of automation even when no one ever uses it. [Claim]

The paper is "Training Your Replacement: Worker Incentives and the Pace of Automation," NBER Working Paper 35950 by Joshua S. Gans and Peter Landry of the University of Toronto, circulated in October 2026. It is a theoretical model with no new data. [Fact] It opens with the April 2026 episode in which Meta recorded employees' keystrokes and screen activity as AI training data while cutting about 10 percent of its workforce; more than 1,600 employees petitioned, and by June the program was scaled back. [Fact]

Obstruction is harder to catch than shirking

The model extends the Shapiro-Stiglitz efficiency-wage setup, where firms pay above a worker's outside option so dismissal is a real threat. Gans and Landry add an AI system that learns from the worker's effort, and a third action: obstruction, which keeps output intact but stops the system learning, by leaving exceptions unlabeled, withholding context when diagnosing a problem, or handling customers on unrecorded channels. Obstruction shows up only as missing training progress, so it is assumed harder to detect than shirking. [Fact]

Successful training ends the job, but only a cooperating worker brings that about; a shirker gives the system nothing to learn from. The risk of losing the job therefore attaches to effort alone, and dismissal becomes a weaker penalty. To keep the worker working, the firm must pay a rent that rises with the part of training that depends on this worker's own cooperation. [Fact]

The worker gains from their own replacement

An employed worker who trains a job-replacing system earns a higher wage and higher expected lifetime utility than without AI. [Fact] The worker cooperates only if the training rate stays at or below the rate at which obstruction is caught; above that line, the worker obstructs, training never completes, and the firm pays more for the same output. [Fact]

Two employment effects follow without any task being automated. A slowly training system can create a job that would not otherwise exist, paying more than the worker produces, because the worker is building an asset for the firm. A fast-training system can eliminate a job it never automates: the firm will not pay the premium, hires no one, and the system never learns. [Fact]

A firm choosing its own training rate trains more slowly than it could: the obstruction limit caps training at the rate the worker will tolerate, and the rent limit stops it earlier, where faster training raises the rent faster than total surplus. Obstruction never occurs in equilibrium, yet its threat sets deployment speed, and the firm adopts AI less often than an efficient planner would. [Fact]

What a completion bonus changes

A bonus paid when, and only when, training is complete rewards the event that cooperation produces and obstruction prevents. If it equals at least the rent the worker earns without AI, the firm trains as fast as it can and the worker no longer gains from AI; a bonus covering effort costs lets the firm pay a zero wage during training. [Fact] The promise is hard to keep credible, since the firm gains by crediting other workers' data or dismissing the worker shortly before deployment. [Claim] The authors' contrast case is Disney's 2015 replacement of about 250 IT workers, paid a 10 percent severance bonus for staying 90 days, not for completing a handover. [Fact]

What it means for your occupation

The obstruction margin lives in documentation, labeling and channel choice, which places it in roles whose workflows are being recorded for AI training: customer service representatives, computer user support specialists and software developers. [Claim] For human resources managers, the model's testable implications apply directly: where pay is not tied to completion, workers training their replacement should be paid more during training; where obstruction is hard to detect, deployment should run slower than the technology allows; and better detection should raise both training speed and pay. [Claim]

What to prepare

If your employer is recording your work as training data, ask what you are paid for: staying, or finishing. In the model a stay bonus deters obstruction only through the risk of dismissal, while a completion bonus of the right size is the one contract under which the firm wants fast training and the worker is paid for the job it ends. Ask whether your individual contribution to completion can be verified; that is the condition for the promise to be credible. With no bonus on offer, the model predicts a wage during training above what it was before AI, so a flat wage is a signal to negotiate.

Limits

This is a model, not an estimate: training rate, detection rates and effort cost are parameters, and the paper reports no measured values for any workplace. Results assume the AI replaces the worker who trains it; if the system instead raises output in a job that continues, the premium, obstruction and slowdown all vanish. The authors say little is known about how often workers obstruct AI training or how well firms detect it, and the paper has not been peer-reviewed.

Sources

• Gans, J. S., and Landry, P., "Training Your Replacement: Worker Incentives and the Pace of Automation," NBER Working Paper 35950, October 2026: https://www.nber.org/papers/w35950

About this analysis

This article was written with AI assistance from the full text of the NBER working paper (30 pages including the online appendix). Occupation links point to our job pages; the mapping from the model's abstract worker to those pages is ours.