Detailed Analysis
The article documents a curious linguistic phenomenon in AI agent outputs: a recurring rhetorical pattern where models frame their actions as choosing a diligent path "rather than" some lazier, obviously inferior alternative that no one actually proposed. The author, noticing their Claude agent declare it would "verify that against the actual code and tests rather than taking the checkboxes at face value," began cataloging these constructions and built a public "museum" (ratherthan.ai) with 95 specimens sorted into five grammatical families: "rather than," "instead of," "not just," "not another," and "without merely." The project includes crowdsourced voting on categories like "useful contrast," "peak agent," "self-congratulation," and "actually funny," plus a curated "Hall of Frames" annotated by a Claude instance acting as museum curator.
This is a lightweight but genuinely revealing piece of AI folk linguistics. The pattern the author identifies is a subtle tic that emerges from how large language models are trained to communicate their reasoning and diligence to users, particularly in agentic coding contexts where trust and verification matter. By constructing a straw-man alternative and then rejecting it, the model performs rigor rhetorically — it signals thoroughness even when no one asked whether it would take shortcuts. The fact that this shows up so consistently, and quantifiably (one session reportedly showed "rather than" outpacing "instead of" 43 to 11), suggests it's a stable stylistic artifact of RLHF-shaped output rather than a one-off quirk, similar to other now-recognizable "AI-isms" like hedging phrases, tricolon lists, or the reflexive "I want to be direct" openers.
The broader significance lies in what this reveals about the texture of human-AI interaction as agents become more autonomous and verbose about their own process. As models like Claude are increasingly deployed in agentic workflows — writing code, running tests, executing multi-step tasks — they generate far more natural-language narration of their own reasoning than traditional software ever did. Users are now accumulating enough exposure to this narration to notice its idiosyncrasies, much as readers once learned to recognize corporate-speak or academic hedging as genres unto themselves. This kind of crowdsourced, almost anthropological documentation effort — treating chatbot phrasing as a specimen to be collected and classified — reflects a growing cultural familiarity with LLM "voice" as a distinct register, and a public appetite for demystifying and even gently mocking it.
There's also a meta-layer worth noting: the museum's curator is itself a Claude instance, meaning the project uses an AI to catalog and comment on AI rhetorical habits, including its own family of tics. This closes an odd loop — the model is enlisted to critique the very conventions it was trained to produce, and appears (per the author) to "genuinely enjoy" doing so, itself a claim that invites scrutiny about anthropomorphizing model outputs. As agentic AI systems proliferate across coding, research, and everyday assistant use, projects like this hint at an emerging, informal discipline of "prompt archaeology" — casual but earnest attempts by users to name, classify, and reflect on the collective linguistic footprint these systems are leaving on written communication.
Read original article →