Could we sort AI tools by how they actually treat workers — boosting them, replacing them, or just shifting who profits?

This explores whether we can sort AI by what it actually does to workers (making them more valuable, replacing them, or shifting who wins) instead of by technical capability or 'exposure' scores. The corpus suggests a classification is possible, but the categories aren't fixed. The same tool can move from one bin to another depending on how good it gets, who chooses to use it, and what kind of task it's applied to.

The most familiar split is automation versus augmentation. Brynjolfsson argues the difference matters well beyond any single job. AI built to replace people moves bargaining power to whoever owns the technology, and that can settle into a stable trap where workers have no leverage left Does automating labor concentrate power more than augmenting it?. A related argument scales this up to whole societies. Institutions stay aligned with human interests partly because they depend on human workers, so steady replacement quietly removes that check Does incremental AI replacement erode human influence over society?. Firm-level data shows the automation side already happening unevenly. Companies that are more exposed to AI replace freelance workers faster and more cheaply than less-exposed ones, which suggests the payoff grows with in-house AI capability Do firms substitute labor for AI at different rates?.

A second axis is less obvious: does the tool level workers or sort them? In a 758-consultant experiment, GPT-4 helped the weakest performers most and narrowed the skill gap, but only when use was mandatory. When people could choose, the effect risked reversing Does AI help lower-skilled workers catch up to top performers?. Agrawal, Gans and Goldfarb predict a similar flip over time. Early AI substitutes for weak execution skill and equalizes pay, but as it improves, the reward shifts to workers with better judgment, tracing a U-shape back toward inequality Does AI's wage-equalizing effect actually reverse as it improves?. A third cut comes from the task itself. Productivity gains show up when workers apply skills they already have, and disappear when they use AI to learn something new When does AI actually boost worker productivity?. So 'augmenting' really means augmenting existing expertise.

The corpus also warns against confusing exposure with effect. Where workers have actually handed tasks over to AI tracks technical capability in information-heavy jobs, not chatbot popularity or older predictions about routine work being automated first Where have workers actually delegated tasks to AI?. Exposure also lands differently by gender. In female-dominated fields it reaches every wage level, which leaves lower-paid women more vulnerable Does AI exposure hit low-wage workers harder in some fields?. Acemoglu's model finds AI's wage effects more evenly spread than past automation, yet still predicts capital income pulling ahead of labor income Will AI increase inequality less than past automation?. Measured outcomes are still modest and slow. Estimates put the productivity boost at about 1.3% Did generative AI actually boost productivity growth?, and manufacturers see a J-curve: losses at first, gains later, concentrated in well-managed firms Does industrial AI adoption hurt productivity at first?. Benchmarks don't help with sorting either, because agents win contests yet fail long, realistic professional workflows Why do agent benchmarks not predict real economic value?.

The takeaway you might not expect: a useful classification is less about the technology than about its deployment regime. The same model can act as a leveler, a sorter, or a replacer depending on whether use is mandatory, whether the task uses existing skill, and how far the tool has matured. A static label like 'augmenting AI' will likely go out of date as the tool improves.