"Using AI Well Takes Skill Too": The Crux Is Checking the Output
A post challenges the claim that using AI is simply a form of automation in which people hand work off. Instructing the AI, reviewing its response, and correcting it, the argument goes, is a skill of its own.
A rebuttal arguing that using AI well is itself a skill appeared on a personal blog in October. It is titled “AI is a tool and using it well takes skill.” The author’s case rests on a simple observation: people direct AI toward a task, review what comes back, and fix it. The point of contention here is not how capable the newest models are. It is what role the human plays in the process of using them.
Is handing work to AI different from using a tool?
The post pushes back on an argument made by Rich Hickey. Hickey held that working with large language models looks less like wielding a tool you keep your hands on and more like automation, where the work is handed off. He argued in particular that delegating something like reading a paper to AI costs you the chance to learn by reading it yourself.
The author of the rebuttal counters that automation and human involvement can coexist. Even when AI produces an answer, the person using it can verify the contents as they go. The core of the rebuttal is that handing off a task does not, on its own, make the user’s skill irrelevant.
The two pieces are looking at different things. Hickey focused on the learning that is lost when work is delegated. The rebuttal’s author focused on the work of checking and correcting the output. Faced with the same situation, one points to what disappears and the other to the role that remains.
What do you do after you get an explanation of the code?
The rebuttal uses programming as a concrete example. A programmer asks AI to explain a piece of code, then checks each part of that explanation against the original code. The process does not end with reading the explanation; it includes confirming that the explanation actually matches the code.
When AI proposes a change to the code, the author notes, the person can test it themselves and then accept it or back it out. In this example, AI generates the explanation and the proposed fix, and the user reviews both and decides whether to apply them. The skill the author has in mind extends beyond writing good prompts to testing results and choosing among them.
The author acknowledges that a model’s output is hard to predict in advance, and that models sometimes fail to follow instructions properly. Even with those limits, the author argues, a person can still use the tool while verifying what it produces.
The debate leaves one question for anyone using AI for work or study: in the tasks you hand off, what will you check yourself, and what will you spend your time actually learning? The question going forward is what difference a habit of carefully verifying output makes to real results and learning.