If AI lets you do something new, is it really your new skill — or just handing the task off while it feels like one?

This explores whether the "new abilities" AI gives workers are really theirs, or whether they are delegation that feels like skill. The corpus suggests the two are easy to confuse, and the confusion has a name. Researchers call it the LLM Fallacy: people credit themselves with what the AI produced, so their sense of their own ability rises even though their actual ability hasn't changed How does AI-assisted work reshape how people see their own abilities?. What makes this surprising is that it doesn't depend on the AI being wrong. It is a self-perception error, separate from hallucination and from over-trusting the AI. Making the AI more accurate won't fix it. What helps is making it clear which part of the work was the person's and which was the machine's.

This matters when you look at who feels best about using AI. Anthropic found that the people who delegate the most work to Claude are the most optimistic about their careers, and they report that their skills are gaining value Does delegating work to AI actually damage worker skills?. Read next to the LLM Fallacy, that optimism could mean two things. It might be real. Or it might be the misattribution effect on a large scale. The study is correlational and was drawn from Anthropic's own users, so it can't tell the two apart. Interviews with Dutch knowledge workers add useful detail. Instead of a simple upskilling-or-deskilling split, they found four outcomes: some skills develop, some are kept, some erode, and some are revalued (they matter in a different way now) How does generative AI actually change worker skills?. Which outcome a worker gets depends on how they use the tool, not just whether they use it.

There is a practical paradox underneath this. A framework for good delegation lists eleven features of a task that decide whether it can be safely handed off. Verifiability comes first, meaning whether anyone can check the result at all What makes delegation work beyond just splitting tasks?. A worker who delegates a task they newly "can" do is often the person least able to judge the output, because the AI supplied the ability and they didn't. Nielsen's proposed interface design points the same way. As people move from doing work to supervising it, interfaces have to help them judge how far to trust the output and correct it precisely, not just prevent errors How should AI interfaces handle the shift from doing to supervising?. Supervising is a skill of its own, and you can't delegate it.

There is a hopeful lateral finding from AI research itself. When models are trained to split tasks into pieces, hand them to sub-agents, and combine what comes back, they get better even on tasks where they work alone. Delegating taught them to break problems down carefully and to stay grounded in the evidence Can delegation teach models to manage context more actively?. Whether people work the same way is an open question, but it suggests the useful skill may be delegating well, not doing the task yourself. At the level of the labor market, delegation so far clusters in information-heavy jobs and follows what AI can technically do Where have workers actually delegated tasks to AI?. Workers also adapt best when AI affects only a few of their tasks, which leaves them room to shift toward the rest Does concentrated AI exposure enable workers to adapt and reallocate?.

One small caution: handing work to AI doesn't always save as much as it seems to. In one study, chat-based delegation cut down on clicks and scrolling but didn't make people finish tasks any faster Does chat delegation actually save time on task completion?. Less effort can feel like more capability. So the honest answer is yes, workers can delegate tasks they newly seem able to do. The corpus's warning is that feeling able and being able come apart quietly, and the gap is most likely to show up exactly where the output is hard to check.