The article contends that large language models (LLMs) are tools that humans can wield, rejecting the claim that they represent pure automation beyond human control. By specifying inputs, examining outputs, and deciding next steps, users direct LLMs just as they would a compiler or spreadsheet, which perform sub‑tasks automatically yet remain under human guidance.
Like any tool, an LLM’s usefulness depends on the operator’s ability to direct it, interpret feedback, recognize mistakes, and make corrections. The author notes that unpredictability is a limitation to manage, not a defining feature that removes LLMs from the tool category, and criticizes anthropomorphic language such as “go rogue” or “conspire” for framing ordinary engineering problems as signs of independent intent.
Effective use requires skill: knowing what to ask, understanding what evidence matters, and recognizing when an answer is wrong. This creates a practical distinction between thoughtlessly delegating work to AI and skillfully wielding it as an instrument to augment human judgment, rather than declaring an entire class of software ineligible to be a tool.