AI Is a Tool. Using It Well Takes Skill.

In his 2026 article, Rich Hickey contends that large language models (LLMs) are not tools but rather forms of automation. While I share his concern about outsourcing work potentially limiting our learning opportunities, I disagree with his conclusion. Even under Hickey's definition, LLMs can be wielded by humans, and using them effectively demands skill, just as operating other tools does.

Mastering a saw or a chisel does not guarantee straight cuts or perfect sculptures; instead, it requires the operator to direct the tool, interpret feedback, recognize errors, and correct mistakes. These same principles apply to LLMs. Their unpredictability, while a significant limitation, is not sufficient to disqualify them from being considered tools.

Instead, factors like predictability, reliability, simplicity, and bounded functionality determine whether something qualifies as a tool. Labeling a tool as such does not negate its usefulness simply because it is challenging to control or prone to misuse. Hickey's distinction between tools and automation also creates a false separation.

Automation can be an essential aspect of a tool. Compilers and spreadsheets, for instance, perform substantial operations without requiring continuous human intervention. We still wield these systems by providing inputs, examining outputs, and deciding on the next course of action. Therefore, why should text generation via LLMs automatically exclude them from being classified as tools, just as code translation or numerical calculations via automation do not?

The claim that LLM usage lacks intention, feedback, direction, and expertise is particularly challenging to substantiate. A programmer can ask an LLM to explain unfamiliar code, challenge the explanation, inspect the source, test suggested changes, reject them, and refine the request. This interaction constitutes a feedback loop directed by a human pursuing a specific goal.

Doing so effectively necessitates knowing what to ask, what evidence is relevant, and when the response is incorrect. Expertise can enhance the tool's usefulness. Hickey's concern regarding reading comprehension is valid for a specific application: substituting a summary for reading a paper may omit much of what reading teaches.

However, one could read the paper and utilize an LLM to scrutinize their interpretation, explore objections, or identify areas for further examination. The potential for bypassing intellectual labor does not imply that every LLM use avoids it. The crucial question is what the individual actually does. What troubles me most is the reliance on the AI industry's anthropomorphic terminology, such as "go rogue," "escape containment," and "conspire."

Adopting these expressions risks accepting the framing that portrays these systems as autonomous entities. While software can generate unexpected outputs, fail to adhere to instructions, or cause harm when granted access to consequential actions, these are genuine engineering concerns. Referencing them as "conspiracy" implies human motives without establishing any.

I would not assert that an LLM consistently performs exactly as directed by a human. It undeniably can fail. However, failing to obey perfectly does not render it any less a tool than successful usage demonstrates that its operator retains their judgment. The critical distinction lies between wielding an instrument with skill and delegating work without adequate understanding or supervision.

This differentiation pertains to AI and offers a more practical alternative to asserting that an entire class of software cannot be considered a tool.