Automation risk is usually framed as a jobs question: which roles will disappear, and when? A new National Bureau of Economic Research working paper argues that the economic effects may arrive earlier—and show up in a less visible place.
In “Replaceable but Employed: Automation and the Meaning of Work,” economist Joshua S. Gans examines jobs where workers value not only pay and output, but also the knowledge that their own contribution is necessary. In the paper’s model, a credible machine alternative can erode that source of meaning even when the employer continues to employ the worker.
That distinction matters for organizations adopting AI and automation tools. Keeping people on payroll does not, by itself, settle whether a deployment has improved the quality of work.
The cost of being visibly replaceable
The paper describes a “meaning externality”: when a machine is publicly shown to be able to perform work associated with a human role, it can reduce the perceived value of the human alternative. The technology need not actually take over the job.
The proposed mechanism is straightforward. A worker may derive value from producing useful output *and* from being materially responsible for it. If a machine can plausibly generate the same output, the second component weakens. The work remains, but the worker’s sense of contribution may not.
In a labor market where wages adjust fully, the model predicts employers would need to compensate workers for that loss. Where wages adjust only partly, workers absorb some of it themselves. That makes the impact relevant not just to compensation planning, but also to retention, engagement, performance management and the design of AI-enabled workflows.
Quality and salience are not the same thing
One of the paper’s most useful distinctions is between a tool’s technical quality and its public salience.
Better technical quality can improve output. But salience—the fact that a machine alternative is made conspicuous—can weaken the human job’s perceived value even without a corresponding improvement in output. Put differently: demonstrating that an AI can do a task may have organizational consequences separate from whether the AI is the best way to do it.
That creates a practical question for executives. Is a public internal rollout, product launch or executive proclamation necessary to realize the operational benefit of an automation project? Or does it primarily signal replaceability to the teams expected to use it?
The paper also suggests an external developer may have an incentive to publicly demonstrate a machine before licensing it. Such a demonstration can lower the value of the human alternative and thereby increase demand for the technology. In the model, that can make privately profitable development socially harmful.
What operators should do differently
This is a theoretical paper, not evidence that every AI deployment reduces job satisfaction or requires wage premiums. But it supplies a useful lens for implementation.
Leaders evaluating automation should track more than hours saved, output per employee and headcount. They should also ask:
- Which parts of a role still require human judgment, ownership or accountability?
- Does the workflow make those contributions more visible, or obscure them?
- Are workers being asked to supervise output they no longer feel responsible for creating?
- Do compensation, progression and recognition reflect the role’s changed demands?
The answer is not to conceal capable tools. It is to avoid treating communications about automation as neutral. Announcing that a system can replace a function can change the experience of that function—even if no one is laid off.
What to watch next
The central prediction to test is whether exposure to credible automation changes compensation, morale, turnover or demand for particular roles before displacement occurs. Firms rolling out AI now have an opportunity to measure that: compare affected teams’ engagement and retention with productivity outcomes, and distinguish tool quality from the way the tool is introduced.
The debate over automation has long focused on whether machines will replace workers. This paper sharpens a prior question for business: what happens when workers remain, but no longer believe the work depends on them?
