Imagine two people whose output doubles after they begin working with AI. On a spreadsheet, the result looks the same. Both completed more work in less time.
Several weeks later, the system is removed. One person continues to perform better because something transferred into their own judgment. The other loses most of the gain because the capability remained inside the tool. A third person still performs well, but only when the accumulated history of corrections and decisions from that particular relationship remains available.
Those outcomes matter economically, even though the original productivity number cannot distinguish them.
What the number can see
Productivity measures answer an important question: how much output was produced for the time, labor, or capital used? When AI helps a person finish a task faster or improve its quality, that change should be measured. Employment, wages, adoption, consumer surplus, and aggregate growth matter too. The Stanford Digital Economy Lab’s AI Economic Indicators is building a timely picture across those dimensions.
But output is produced at the level of a configuration: a person, a system, a task, an organization, and the conditions governing their interaction. A gain observed there does not automatically tell us what changed in any one component. The spreadsheet records the result. It does not record whether the person learned, lost practice, became more dependent, gained authority, or received any meaningful share of the value.
Capability can end up in different places
Some AI-assisted work may build human capability. A person encounters better examples, receives useful correction, practices judgment, and later carries that improvement into a new situation. If the tool disappears, part of the gain remains.
Other gains may remain system-supplied. The work improves while assistance is present, but later unaided performance stays near its starting point. That does not make the assistance worthless. We routinely rely on calculators, search engines, and other tools without treating independence from them as the only acceptable outcome. It does mean that the observed gain should be described accurately.
There is also a third possibility. Repeated interaction can produce a working relationship with its own accumulated state: corrections, preferences, constraints, decisions, and ways of coordinating. The resulting capability may belong neither to the person alone nor to the model in isolation. It may reside in the relationship.
If that relationship remains visible, correctable, portable, and under human judgment, dependence can function as a governed extension of capability. If the state is hidden, locked to one provider, or reused outside its original authority, the same dependence creates a different economic condition.
The human result has an economic consequence
A worker who develops transferable capability may gain mobility and resilience. A worker whose performance depends on one employer-controlled system may become more productive while losing bargaining power. An organization may capture the output gain while the person absorbs the cost of constant monitoring, reduced discretion, or skill decay.
Average productivity can rise in every case. The distribution of capability, authority, and value does not.
This is why the phrase human-centered AI needs a measurement question attached to it. Good intentions in design cannot show whether people became more capable, more secure, or more able to direct the conditions of their work. We have to observe what persists after assistance, what fails when the configuration changes, and who controls the accumulated resources that made the gain possible.
A measurement problem worth adding
The research direction I am developing at Perfinitive begins with a simple comparison. We can observe a human baseline, performance during AI assistance, and later human performance after assistance is removed or the task changes. The distance between those states may help estimate how much of an assisted gain became independently observable human capability.
That comparison is only a beginning. Human development is larger than task performance. Agency, opportunity, job quality, accommodation, ownership, and the distribution of gains cannot be compressed responsibly into one ratio. Nor should unaided performance become a moral standard. For many people, durable access to a tool or relationship is itself part of substantive capability.
The narrower point is that economic measurement should stop treating the human participant as unchanged simply because the output has been counted. When AI improves the work, we should also ask what changed in the person—and what will remain available to them when the system, employer, model, or relationship changes.