Does AI mainly erase tasks outright, or just shift them from people to machines — and does that distinction even change wages?

This explores whether AI changes wages mainly by taking over tasks outright, or by shifting which tasks people do and which tasks machines do instead. The corpus suggests the either/or framing misses the point. Automation and reallocation are two steps of one process. Wages depend less on how many tasks get automated than on *which* tasks go and *where the displaced work ends up*.

The sharpest version of this comes from a model showing that wage effects depend on the expertise level of the tasks removed Does automation raise or lower the skills that remaining work demands?. When automation removes a job's easy, routine tasks, what's left takes more skill: wages rise, but fewer people qualify. When it removes the expert tasks, the job gets easier: wages fall, but more people can do it. So the same amount of automation can push wages in opposite directions. Acemoglu's macro model reaches a similar conclusion from a different direction. AI's wage effects are spread more evenly across demographic groups than earlier waves of automation, but the bigger shift is income moving from labor to capital, driven by reallocating tasks between machines and workers rather than by simply eliminating tasks Will AI increase inequality less than past automation?.

Whether workers can actually reallocate depends on how AI exposure is spread across a job. Firm-level data from 2010–2023 shows that high average exposure reduces labor demand. But when exposure hits only a few tasks within a role, workers shift to the tasks AI doesn't touch, and net job losses stay modest Does concentrated AI exposure enable workers to adapt and reallocate?. Timing matters too. Danish administrative records show workers taking on new AI-related tasks within two years of ChatGPT's launch, while earnings and hours barely moved Does AI chatbot adoption change worker pay and hours?. Reallocation appears to come first, and any wage effects come later, if they come at all. Even within a single task, AI reallocates time rather than saving it, moving effort from doing the work toward prompting and checking outputs Does AI really save time, or just change how we spend it?.

The answer also depends on who gets exposed and who gets to use the tools. Where workers have actually handed tasks to AI, it's concentrated in information-heavy jobs, and the wage pattern reverses at the advanced-degree level, which doesn't match older predictions built on routine tasks Where have workers actually delegated tasks to AI?. Exposure also differs by gender. In female-dominated occupations it reaches lower-paid workers as much as higher-paid ones Does AI exposure hit low-wage workers harder in some fields?. On the augmentation side, GPT-4 narrowed the performance gap among BCG consultants, which could compress wages within a role Does AI help lower-skilled workers catch up to top performers?. But those gains showed up only on tasks people already knew how to do, not when they were learning new skills When does AI actually boost worker productivity?.

Here's the part you might not expect: the balance between automation and reallocation isn't set by the technology. It's a choice firms make. Acemoglu, Autor and Johnson argue that firms favor automating expertise over creating new tasks for workers because it pays better for the individual firm, even when it's worse for workers overall Why do firms build automating AI instead of pro-worker AI?. Firms also substitute AI for labor at very different rates, and highly exposed firms do it faster and more cheaply Do firms substitute labor for AI at different rates?. So whether AI's wage effects look more like automation or more like reallocation depends heavily on what firms decide to build.