Does using AI to help with work eventually lead to AI just doing the work instead — or does it work the other way around?

This explores whether AI that helps people (augmentation) tends to turn into AI that replaces them (automation), or whether automating some tasks is what makes new kinds of human-AI collaboration possible. The corpus says both happen, and treating them as a single either/or choice is the mistake. Look at one task at a single moment and the two are opposites. Look across a set of tasks over time and they feed each other in a loop: working alongside AI produces the know-how and structured routines that make a task automatable, and automating a task frees people for new work where they need AI's help again Are automation and augmentation really a simple trade-off?.

There is evidence for the 'automation enables augmentation' direction. When a job's tasks depend on each other, so that one weak link drags down the whole result, automating some of them pushes worker attention onto the tasks that are left. Those tasks become higher quality and worth more Does automating some tasks actually increase the value of remaining work?. Usage data shows a related pattern: in Anthropic's Economic Index, early adopters mostly hand off whole tasks, and as use matures it moves toward collaboration and learning Does AI adoption follow wealth and mature over time?. So delegation can be the first step toward a richer partnership, not the end state.

The opposite direction, augmentation sliding into automation, shows up mostly as a risk. A study mapping thousands of workplace AI scenarios found that augmentation is not automatically safe. Leaning on AI agents slowly wears down people's skills and their ability to check the AI's work, until the 'human in the loop' is only there in name Does AI augmentation protect workers from skill erosion?. Older automation research points the same way: once a system moves from presenting information to choosing actions, operators lose the awareness they need to step in when it fails When does automation harm performance after failures?. Augmentation that nobody designs carefully can become automation by drift.

What decides which way the loop turns is mostly economics and institutions, not the technology. Firms earn more by automating expertise than by creating new tasks for workers, so they underinvest in pro-worker AI Why do firms build automating AI instead of pro-worker AI?. Over time, choosing automation can lock in a trap where workers lose the bargaining power to push things back toward augmentation Does automating labor concentrate power more than augmenting it?. At the scale of whole societies, step-by-step replacement weakens a quiet safeguard: institutions stay responsive partly because they depend on people who care how things turn out Does incremental AI replacement erode human influence over society?.

The less obvious point is that the loop can break at the bottom. Raisch and Krakowski's cycle assumes there are always people skilled enough to work on the next frontier. A growth model shows that automating entry-level work can raise output today while slowing long-term growth, because novices stop learning from experts and tacit knowledge stops getting passed on Can automation raise output while slowing growth?. So whether automation enables future augmentation depends on whether we protect the apprenticeship pipeline that produces the people who will do the augmenting.