Could policy steer AI toward making workers more capable instead of replacing them — and would that actually work?

This explores whether policy can push AI toward tools that make workers more capable instead of systems that replace them. The corpus makes a strong case for why the choice matters. It has much less on which specific policies would actually work. That gap is itself worth knowing about.

The main argument comes from Brynjolfsson Does automating labor concentrate power more than augmenting it?. He says automation versus augmentation is a question of power as well as productivity. When AI replaces workers, bargaining power moves to whoever owns the technology. That can settle into a stable trap: once workers have no leverage, they also have no way to change the arrangement. A separate line of research on gradual disempowerment Does incremental AI replacement erode human influence over society? takes this further. It argues that institutions stay aligned with human interests partly because they depend on human workers who care how things turn out. Replace that labor bit by bit and you lose a hidden safety mechanism, not just jobs. So augmentation is more than a nicer labor outcome. Keeping humans involved may be part of what keeps the whole system answerable to people.

The less encouraging part is about whether policy can steer this at all. When Amodei proposed slowing AI development for safety, both the US and Chinese governments rejected the idea within days Can AI safety pacing work without government cooperation?. That suggests competition between countries can override even well-designed proposals. The Future of Life Institute argues that companies won't police themselves and calls for binding government limits backed by hardware verification Can companies alone manage the risks of AI systems?. Its focus is safety, though, not labor. Legal scholars add another obstacle. Traditional tools like incentives, monitoring, and enforcement struggle with AI agents that act fast, at scale, and in ways nobody can easily inspect Can traditional governance tools actually control AI agents?. A policy that says "augment, don't automate" has to be checkable, and that's hard when it's unclear what the system is actually doing. A more promising route is to shift liability and require transparency so that deployers are accountable for who gets displaced and who benefits Should AI agents serve one user or balance many interests?.

The surprise in the corpus is that augmentation may be shaped less by regulation and more by interface and training choices that look technical. Cocoa, a research tool, lets users assign each step of a plan either to themselves or to the AI, and change the plan mid-task. Users got more control over the work without it becoming harder to use Can users steer AI agents better by revising plans mid-execution?. Research on proactive agents finds that AI assistants are passive because of how they're trained, not because of what they're capable of. Behaviors like asking clarifying questions can be trained in Why do AI agents fail to take initiative?. In other words, whether AI works alongside people or replaces them is often decided by design choices made long before any regulator gets involved. Policy could target those choices, for example through procurement standards or design requirements, instead of trying to regulate outcomes afterward.

What the corpus doesn't cover: concrete economic levers like taxing automation differently from augmentation, subsidies for complementary tools, or labor-law reforms. If those are what you're after, this collection has the diagnosis but not the prescription.