Employers often describe AI adoption as a way to make work cheaper. A more difficult question arises when the knowledge needed to build the replacement sits inside the job itself. If a worker has to explain their decisions, label the exceptions and correct the automated system, their cooperation becomes an input into the project that may end their employment. That changes the bargaining problem in ways that a simple comparison of wages and software costs misses.[1][2]
Economists Joshua Gans and Peter Landry examine this problem in an October 2026 National Bureau of Economic Research working paper, Training Your Replacement: Worker Incentives and the Pace of Automation. They construct a theoretical model in which a worker can cooperate with AI training, reduce effort or obstruct the process while continuing to perform ordinary work.[1][2]
One central assumption is unusually stark. Successful training ends the job, and the training cannot be completed without the worker’s cooperation. In an ordinary efficiency-wage model, the threat of dismissal can induce effort: lose the job and lose the benefit of keeping it. Here, the worker’s effort also brings the dismissal closer. The threat is less effective because the very behaviour the firm wants would eliminate the position.[1][2]
Why cooperation can become expensive
A firm may respond by paying the worker more during training. The economic logic is not generosity. It is an attempt to make cooperation preferable to withholding knowledge or quietly slowing the process. In the model, the extra compensation needed to secure effort is a form of worker rent, an advantage above what would otherwise be required for the job.[1][2]
The authors show that the interaction can create surprising equilibria. AI may sustain jobs that pay more than the worker’s marginal product, because the wage is financing progress towards eventual automation. Conversely, it can eliminate jobs that the technology never ultimately automates, because the incentives around the attempted transition alter the employment decision.[1][2]
The model’s incentive mechanisms
| Model condition | Economic consequence under the stated assumptions |
|---|---|
| Training requires worker cooperation | Workers can influence the pace of automationSource 1 |
| Successful training ends the job | Dismissal is a weaker threat for inducing training effortSource 1 |
| Firm pays wages only | Training may slow and worker rents may increaseSource 1 |
| Credible calibrated completion bonus | The authors show a way to remove the worker’s gain and the slowdownSource 1 |
| Real-world generalization | Not measured by the theoretical paper; depends on contracts, institutions and alternativesSource 1 |
These are logical results inside the paper’s assumptions. They are not claims that surveys found workers in large companies receiving specific premiums, or that the model explains every job loss during recent AI adoption. The paper does not provide a measured forecast of how many positions will disappear in a given year.[1][2]
The authors identify another mechanism: a credible completion bonus. In their model, a bonus calibrated to the worker’s rent without AI can remove both the worker’s gain from the replacement option and the resulting slowdown in training. The important adjective is credible. A promise that can be revoked after workers reveal their expertise may not change their behaviour in the way a contractual commitment would.[1][2]
A realistic example of the incentive problem
Imagine a specialist who resolves exceptional insurance claims. Routine cases are well documented, but the worker’s judgment is needed when medical information conflicts or the policy wording is ambiguous. Management wants the specialist to explain these edge cases to an AI team. The worker has two different obligations: keep processing current claims accurately and help build a system that may ultimately remove the job.[1][2]
A conventional rollout plan may assume the same person will volunteer every exception and correct every model error. Yet their private incentive may be to keep expertise informal or to delay the transfer. That is not a moral judgment about workers; it is the predictable consequence of placing the immediate burden and long-term risk on one side of a transaction.[1][2]
The simple example illustrates the model but is not a reported case from the study. Real companies face employment law, collective bargaining, professional duties, reputational costs and alternative training pathways. They may buy data from other firms, use historical records or retrain displaced workers for new roles. Each changes the incentives, so applying the paper’s formula literally would require careful adaptation.[1][2]
What managers should examine before replacing the workflow
The analysis suggests practical questions about ownership and incentives. Which expertise is codified, and which lives in undocumented decisions? Who is expected to train the system? Is that work compensated? Would workers have reasons to report errors and gaps in the automation honestly? Are retraining or redeployment offers credible? Can the firm verify the quality of training data without punishing useful criticism?[1][2]
It also identifies a weakness in many technology cost models. A vendor’s inference bill and an engineer’s implementation estimate are visible. The cost of obtaining the tacit knowledge to make the system dependable may be far less visible, especially when the people who hold it anticipate losing their jobs.[1][2]
The paper’s contribution is a theoretical warning, not an empirical verdict. Its hypotheses should be tested against actual adoption projects and compensation arrangements. But the core insight is strong: automation can depend on human cooperation at exactly the moment it undermines the rewards for giving that cooperation. That is a design problem for the business, not something a larger language model will solve by itself.[1][2]
Sources & further reading
- Training Your Replacement: Worker Incentives and the Pace of Automation Gans and Landry, NBER Working Paper 35950 · Accessed 2 Nov 2026
- Training Your Replacement: Worker Incentives and the Pace of Automation Gans and Landry, SSRN · 16 Oct 2026 · Accessed 2 Nov 2026