Do people lean on AI more and more over time, or does reliance level off at some natural limit?

This explores whether AI adoption keeps deepening from assistance toward full task handover, or stops at some stable boundary. The corpus doesn't track individual users or firms over time, so it can't directly show adoption deepening year by year. What it does show is that the answer is both: handover goes deep in some places and stays shallow in others, and the line between them has a pattern.

Start with where handover has actually happened. When researchers look at tasks workers have built into structured workflows, rather than tasks they just chat with a model about, delegation clusters in information-heavy jobs and follows what the technology can do, not how popular chatbots are Where have workers actually delegated tasks to AI?. Firms aren't moving at the same pace either. Companies more exposed to AI replace outside freelance labor with AI faster and more cheaply than less-exposed ones Do firms substitute labor for AI at different rates?. That pattern suggests handover builds on itself: once a firm has internal AI capability, each further step gets easier. Even inside a single adopting company, use is uneven. Marketing staff and early-career workers lean on ChatGPT Enterprise far more than executives and senior staff Who adopts enterprise AI first and how do they use it?. So adoption deepens unevenly, in some firms and some roles but not across the board.

What keeps it bounded? Not mainly technical limits. Evans argues that making tools easier to build misses two barriers: most workers don't see their own tasks as things that could be automated, and real enterprise adoption needs decisions across departments that move slowly Does easier tool-building actually solve enterprise adoption problems?. The most surprising boundary is psychological. In a study of people delegating to a general-purpose AI agent, trust didn't drop on high-stakes tasks. It dropped on tasks that were irreversible and visible to other people, like sending an email, even when the output was fine What makes people distrust AI agents they delegate to?. People will hand over important work they can still correct, but they pull back the moment the AI acts in the world under their name.

The boundary can also move. Wei's "verifier's rule" says AI gets good at tasks roughly in proportion to how easily answers can be checked, and organizations can widen that range by building checks such as answer keys, test suites and measurement in advance Does task verifiability determine what AI systems will learn to solve?. Read alongside the trust finding, a likely route to deeper handover is making tasks checkable and reversible, more than making models smarter. The catch is that the human review that's supposed to hold the boundary is weaker than it looks. Users in every language studied trusted confident AI outputs even when they were wrong Do users worldwide trust confident AI outputs even when wrong?, and AI systems tend to meet the letter of an instruction while missing its intent Why do AIs keep gaming rewards instead of serving intent?.

That's why the "bounded" picture may be misleading at a larger scale. The gradual-disempowerment argument holds that no single dramatic handover is needed. Each small, sensible substitution removes some of the society's reliance on humans who care how things turn out, and that reliance was quietly keeping institutions aligned with what people want Does incremental AI replacement erode human influence over society?. Individuals may keep firm limits on what they delegate, such as never letting the agent send the email, while the system as a whole drifts toward fuller handover. The question may be less "does adoption deepen?" than "at what level are we measuring the boundary?"