Perspective
The AI advantage is a skill.
Why I believe in AI as a superpower—and why the edge comes from how you use it.
Potential needs practice.
I believe AI is a superpower. It can expand what a person is able to attempt, build, and learn. But the useful part of that belief is the responsibility it puts on the person using it: develop the skill, choose the task carefully, and check whether the work actually improved.
Access opens the door. The ability to turn that access into a useful result is what interests me. That means knowing when to ask for alternatives, when to supply more context, and when to stop generating and start checking.
The evidence depends on the work.
The 2023 Generative AI at Work working paper reported an average productivity increase of about 14% in a customer-support setting, measured by issues resolved per hour. Gains were larger for less-experienced workers. That is evidence of value in a particular work environment, rather than a universal promise.
A different setting produced a different result. METR’s early-2025 study involved 16 experienced developers completing 246 tasks in familiar open-source projects. With AI available, tasks took 19% longer. The finding was specific to those developers, tools, and tasks.
That result also has a date. In a February 2026 follow-up, METR described evidence suggesting improvement, while explaining why selection effects made the size of the benefit difficult to estimate reliably. An older result should not become a permanent verdict on a changing technology.
My reading is that optimism and measurement belong together. These studies ask different questions in different environments. Neither tells you exactly how much benefit you will get from your next task.
Measure the whole job.
A fast first draft is only one part of the work. The useful comparison includes preparation, review, corrections, and the time it takes to reach an acceptable result.
For example, imagine preparing a research brief. Getting text quickly is helpful only if the sources support it, the important details survive, and the final brief answers the actual question. Ten minutes of generation followed by an hour of repair should be counted honestly.
I would begin with one recurring task and define what “good” means before trying to improve it. Track the time to a usable result, the mistakes that remain, and how much revision is necessary. Repeat the comparison across several examples. Treat the findings as local evidence, not a scientific claim about all AI.
Build capability you can explain.
Learning a tool is valuable. Learning why a particular approach worked is more reusable. Was the task better defined? Did a source document remove ambiguity? Did a clear review rule catch a mistake?
Those questions turn repeated use into learning. They also make the skill easier to transfer when a model or interface changes. A useful habit should survive beyond the button that first introduced it.