# LLM as Average Democratizer

A framing of LLM impact that sidesteps both the hype ("LLMs do expert work") and the dismissal ("LLMs produce slop"). The claim: LLMs' main effect is to collapse the cost of producing *average* output — writing that's serviceable, code that compiles and mostly works, SQL that returns the right columns, drawings that look like drawings. They don't raise the ceiling; they raise the floor.

## The mechanism

Average work used to require real time and skill. Most people, lacking those, produced sub-par work or nothing at all. LLMs let them produce average directly. The normal distribution of output stays the same shape; the left tail gets pulled toward the middle.

This has two downstream effects:

- **The cost of average approaches zero.** Writing an average product blurb, an average SQL query, an average pitch deck no longer needs a professional.
- **The value of average approaches zero.** When everyone can produce average cheaply, average stops being a differentiator, and the market shifts attention to whatever is above average — or to filters that cut through the flood.

The rawquery essay [[average-is-all-you-need]] celebrates the first effect. [[peril-of-laziness-lost]] and [[ai-great-leap-forward]] point at the second.

## Where it holds

Fields where output is heavily determined by "descriptive textual semantics": marketing copy, internal memos, boilerplate legal drafting, translation, tutorial writing, standard analytics queries, CRUD code. These are all places where mediocre-but-correct is a common endpoint, and where the gap between "nothing" and "something useful" is what matters most.

Data analytics is a clean example. A non-technical user has high intuition about their own data but lacks SQL/charting skill. An LLM plus a connector layer (see [[rawquery]]) can bridge the gap without a data team.

## Where it breaks

Several classes of work resist the argument:

- **High-stakes correctness.** Average SQL against production, average legal drafting, average medical summaries — wrong answers have real cost. The [[cult-of-vibe-coding]] critique applies: skipping the audit step is where the failure mode lives.
- **Taste-driven work.** Design, fiction, architecture, strategy. Average output is the enemy because it's indistinguishable from everyone else's average output. Related: [[working-on-products-people-hate]].
- **Essential complexity.** [[no-silver-bullet]] argues most software difficulty is in specification, not implementation. LLMs cut accidental complexity (typing, boilerplate, syntax lookup); they don't cut essential complexity (deciding what to build). See [[no-silver-bullet-llms]].
- **Distributed coordination.** Multi-agent systems amplify, rather than absorb, the variance in average output. See [[log-distributed-llms]].

## The second-order effect

When average is free, attention becomes the scarce resource. The distribution of outputs gets flatter; the distribution of trust, provenance, and curation gets steeper. Personal wikis, curated feeds, and hand-maintained lists appear more valuable precisely because of the glut. This is part of the argument for the [[llm-wiki-pattern]] and for [[filesystem-is-graph-database]]-style personal knowledge work: if the model produces average text cheaply, the durable artifact is a structured, human-curated graph that the model can draw on later.

## Related

- [[ai-assisted-workflow]] — workflow that uses the LLM for average tasks while preserving human judgement at decision points
- [[vamp-ai-frontend]] — making the machine-readable surface explicit so the LLM can produce average-quality frontend code that actually composes
- [[clean-code-coding-agents]] — keeping average LLM output maintainable by giving it less to parse