Rich Hickey — though credited only as "Rich" here, the voice and intellectual framework are unmistakable — draws a line that most AI discourse has been too lazy to draw. The word "tool" has become a rhetorical shield: by calling an LLM a tool, users grant themselves the craftsperson's dignity without doing the craftsperson's work. Hickey's argument is that this isn't just imprecise language. It's a category error with consequences. The core move is definitional, and it's sharp. A tool is something you wield — held in hand, powered by your energy, directed by your continuous attention, returning sensory feedback that builds expertise over time. A hammer teaches you about materials. A saw teaches you about grain. A text editor teaches you about the structure of code. The feedback loop between intention, action, and result is where skill lives. LLMs break every link in that chain. You don't direct them continuously. You don't supply the energy. You don't get feedback that maps to the domain you're working in. You get output. The drive-thru analogy lands harder than it first appears. A drive-thru isn't a cooking tool — it's automation that replaces cooking. When you "use" an LLM to read research papers, you didn't read those papers. The LLM read them. What you did was operate a request-fulfillment interface, and what you learned was how to operate that interface — knowledge with the shelf life of the next model update. Hickey calls this "working for a capricious robot" instead of working with the world, and the phrase sticks. The essay's strongest section concerns what's lost in the abdication. Reading papers yourself teaches you how arguments are constructed, which authors to trust, what the edges of a field look like. These are not transferable via summary. The background loading of your mind with raw material for novel thought — what cognitive scientists call incubation — requires the actual encounter with the material. An LLM summary is to reading what a nutrition label is to eating: technically informative, experientially void. Where the piece is weakest is in its implied universalism. The tool/automation distinction is real and useful, but it doesn't resolve every case. A calculator is automation for arithmetic, and mathematicians use them without guilt because the expertise lives elsewhere. The interesting question — which Hickey doesn't fully engage — is where on the tool-to-automation spectrum a given use of LLMs falls, and whether the expertise being bypassed is one the user needs to build. For a junior developer, abdicating code-writing is catastrophic. For a senior architect generating boilerplate, the calculus is different. Still, the essay's rhetorical precision is its real contribution. By anchoring "tool" to "wield" — with all its implications of grip, force, direction, and feedback — Hickey gives readers a single conceptual test they can apply to any technology claim. Does this thing extend my capability while keeping me in the loop? Or does it replace my participation while letting me pretend I'm still the one doing the work? That question, applied honestly, is more useful than a hundred AI ethics frameworks. The piece reads fast, argues clean, and leaves you with a framework that sharpens every future conversation about AI use. It's the kind of essay that changes your vocabulary, which is the most an essay can do.
