ENTITY

Anthropic

Properties1
tagsanthropic, claude, ai-research, training-data, study

2 min read

Anthropic

Type: organization (AI lab) | Domain relevance: Builder of Claude; author of the “250 documents” study that underpins the 250 Authority Protocol.

Summary

Anthropic is the AI lab behind Claude. In this wiki it matters chiefly for a study — conducted with the UK AI Security Institute and the Alan Turing Institute — showing that ~250 documents (~0.00016% of training data) were enough to establish a behavioral pattern in models from 600M to 13B parameters, with the threshold not scaling with model size. Caleb Ulku uses this as the namesake and justification for his 250 Authority Protocol. Important nuance (which Caleb partly acknowledges): the study tested a data-poisoning trigger, not brand/marketing influence, so its application to GEO is an extrapolation, not a proven result.

Key Facts

  • Builds Claude; partnered with UK AI Security Institute + Alan Turing Institute on the study.
  • Study finding: ~250 documents can establish a pattern across 600M–13B-parameter models; defenses don’t scale with size.
  • Study scope: a poisoning trigger mechanism — not a test of brand marketing influence.
  • Used (with caveats) as the evidentiary basis for the 250 Authority Protocol.

In the Sources

SourceContext
A small number of samples can poison LLMs of any sizeThe primary study (with UK AISI + Alan Turing Institute) — now ingested; its scope quotes rebut the SEO extrapolation.
TRAIN Any AI To Recommend You (This Study Proves How)Cites the 250-documents study as the basis for the protocol.

Entities: Claude, Claude Code, ChatGPT, UK AI Security Institute, Alan Turing Institute Concepts: Data Poisoning & LLM Backdoors, 250 Authority Protocol, GEO / AEO (Getting Recommended by AI)

Notes

Backfill resolved: the primary study is now ingested as A small number of samples can poison LLMs of any size. It confirms the ~250-doc figure but for a narrow gibberish backdoor under direct training-set control — which does not support the marketing 250 Authority Protocol (see Data Poisoning & LLM Backdoors). Anthropic also builds Claude and Claude Code.