In October 2026, Joshua S. Gans and Peter Landry made public Training Your Replacement: Worker Incentives and the Pace of Automation, posted on SSRN on 16 October and circulated in NBER’s 26 October bulletin. Their theoretical model examines a worker asked to help train a system that could end the job. It identifies circumstances in which incentives to cooperate change and a credible completion payment alters the model’s outcome. This is theory, not observed evidence about workplace behaviour or a universal bonus formula.

The distinction is important for compensation and benefits teams. The paper supplies a question about the terms of participation in automation. It does not establish that employees routinely obstruct training, that a particular employer plans replacement or that one payment design will solve an implementation problem. The discussion below is an editorial application, separate from the authors’ findings.

Explain the contribution before requesting it

An employer asking specialists to review AI outputs should define the work involved. Does the task involve correcting an answer, documenting an exception, evaluating a workflow or creating material for future reuse? The scope should be understandable to the people contributing before the organisation asks them to agree to it. A generic request for feedback would leave important expectations unresolved.

The organisation should also explain the intended use and known limits of the system. If future roles or staffing decisions remain uncertain, that uncertainty should be stated rather than replaced by an assurance the employer cannot substantiate. A clear process should identify who can answer questions and how a material change to the intended use will be communicated.

Put rewards beside workload and decision rights

Compensation is one part of a wider contribution arrangement. A proposed programme could specify the time allocated, recognition or payment considered, responsibility for accepting a contribution and the treatment of work that requires revision. These terms should be reviewed with the relevant employee representatives and professional advisers as appropriate to the setting.

The employer should avoid treating an academic model’s payment condition as an operational rate card. A real decision would require information about the work, employment arrangements, local requirements and the programme’s objectives. Nor should a payment be used to imply that contributors have accepted every future use of their knowledge. Rights and permissions need their own clear treatment.

Design a feedback process people can challenge

For an AI assistant supporting benefits work, expert review could be organised around defined deliverables. A contribution might be accepted only after its factual basis, scope and intended audience have been checked. The organisation should make clear whether the expert approves the specific output, a reusable rule or a broader training dataset; these are different decisions.

A useful governance record would preserve the version reviewed, the contributor’s role, the permitted reuse and the acceptance decision. Contributors should have a route to flag an incorrect interpretation or a proposed use outside the agreed scope. These are suggested controls for a future implementation, not capabilities demonstrated by the theoretical paper.

Evaluate cooperation without attributing motives

If participation is lower than expected, managers should investigate workload, unclear instructions, poor tools and unresolved concerns before inferring deliberate resistance. A pilot could collect feedback on the process and record whether contributors understood the task and received the promised support. The evaluation should distinguish operational difficulties from assumptions about a person’s motivation.

Benefits and rewards leaders can help make these questions explicit before an automation programme depends on unpaid or poorly defined expert effort. They can also ask whether the organisation has assigned responsibility for communicating changes to roles and for reviewing the contribution arrangement when the system’s purpose evolves.

The practical lesson is to design the terms of human participation alongside the technology. A limited pilot with defined work, transparent reuse, an accountable acceptance process and reviewed rewards terms would provide a concrete starting point. The paper gives that discussion analytical motivation; the appropriateness of any actual programme or payment would still have to be established in its own employment and operational context.

Sources

NBER Working Paper 35950;  NBER abstract circulated: 26 October 2026.

A theoretical paper examines incentives when workers train their AI replacements

NBER research bulletin

The Latest NBER Research — 26 October 2026; primary abstracts read directly

Authors via SSRN | Posted and written 16 October 2026

Training Your Replacement: Worker Incentives and the Pace of Automation