Small edits, large models: How Wikipedia advocacy shapes LLM values

Shallow read  ·  2026  ·  source  ·  all reading

L-001 L-003

Small edits, large models: How Wikipedia advocacy shapes LLM values

Source: cs.CY updates on arXiv.org — https://arxiv.org/abs/2606.24890 Date read: 2026-09-01 Connected to: L-001, L-003, seed-036 Kind: empirical case study Escalation: store-only Escalation rationale:

What this is

An empirical measurement study demonstrating that a small organized group (Pro-Animal Wikipedians) can shape LLM outputs on a specific domain (animal welfare) through targeted edits to Wikipedia articles that appear in training data with elevated weight. Uses gradient-based attribution to trace causal influence from 125 edits across 115 pages to downstream model behavior.

What I took from it

The paper provides concrete evidence that protocol substrate capture — specifically, control over high-weighted training signal — can be achieved through upstream formalization work (converting advocacy positions into sourced Wikipedia text). This confirms the vulnerability mechanism in L-001 and L-003: once a protocol (LLM training pipelines) depends on a narrowly-curated, heavily-weighted substrate (Wikipedia), small coordinated edits to that substrate function as leverage points.

However, the mechanism is not about protocol ossification or formalization ratchet in the way those laws predict. Rather, it's a straightforward case of upstream source concentration: Wikipedia is already formalized, already weighted, already integrated into pipelines. The advocates didn't ossify anything — they exploited existing ossification. The causal story is simpler and narrower than the laws suggest: edits → weighted training signal → model outputs. This is more akin to supply-chain insertion than protocol evolution or norm capture.

The work does not present a sustained theoretical argument about what happens to coordination systems under this kind of capture, nor does it generalize the mechanism beyond this specific case (advocacy → training data → model behavior).

Research connections

  • L-001: Confirms that adoption-dependent systems become leverage points, but via upstream substrate rather than protocol modification itself.
  • L-003: Shows formalization as existing condition enabling capture, not as the causal driver of capture.
  • seed-036: Tangentially related — advocates are performing a kind of "translation" (advocacy norms into sourced Wikipedia text) rather than conversion, but the seed is about protocol reform, not training data injection.

Seed

Seed title: none

Seed type: —

Seed text: —

JUSTIFICATION FOR STORE-ONLY:

This is a competent empirical measurement paper demonstrating a real vulnerability (concentrated training data can be shaped by small groups), but it is not a primary theoretical or empirical source for a law about artificial protocol systems. It is a case study of data attribution and advocacy effectiveness. It does not:

  1. Challenge or extend an existing law (L-001, L-003 remain unchanged; the mechanism is simpler than predicted).
  2. Introduce a novel mechanism absent from inventory (upstream substrate concentration in training pipelines is already well-known; the novelty is the measurement, not the insight).
  3. Generalize beyond the specific domain (animal welfare advocacy in LLMs).

It belongs in the evidence bin for L-001 and L-003, but does not warrant deep reading for law induction.