Drawing on Borges’ allegory of the useless, empire-sized map, this analysis positions Large Language Models (LMs) as modern cartography. These tools have become so effective at summarizing reality that users risk substituting the representation for the actual territory. The core engineering challenge now shifts from building better models to developing better methods for interacting with these powerful, abstract layers mediating professional work.

Baudrillard’s four stages of representation map uncannily onto LM usage, from faithful copies (Stage 1) to representations masking absence (Stage 3). Because models generate smoothed, consensus-based outputs, users can easily stop verifying sources, confusing consumption of averaged explanations for genuine research. The map itself, unlike static ones, is personalized, shifting based on the prompt structure and the user's background.

Developing expertise requires cultivating a tacit skill, akin to Michael Polanyi’s concept of knowing more than one can articulate. This isn't about spotting obvious hallucinations; it’s a subtle intuition, a “code smell,” that signals when the output feels too polished or lacks grounding. Expert interaction involves shifting awareness through the output to the underlying territory, a skill learned via practice, not checklists.

This necessary skill set resists easy codification. While LMs function as powerful means of summarization, enabling novel connections across domains, the danger lies in the cartographers falling in love with their maps. Successfully navigating this requires a delicate balance between trusting the abstraction and remaining tethered to external validation.