WikiTruth

AI finds 184 factual contradictions in Wikipedia

AlphaLlama loaded 4.9 billion tokens into a single context window on one RTX 3090 — and found cross-article contradictions invisible to human editors, bots, and short-context AI.

184

Contradictions

81,444

QA pairs

4.9B

Tokens processed

1 GPU

RTX 3090 · 24 GB

How It Works

AlphaLlama loads 4.9B tokens into a single context window. QA pairs and contradictions come out.

STEP 1

Corpus

4.9B tokens from 14 datasets including full Wikipedia dump

AlphaLlama

Loads entire 4.9B token corpus into a single context window on one RTX 3090

STEP 3

Output

81,444 cross-document QA pairs + 184 contradictions

Technical Summary

Public methodology.

Experiment Parameters

Context window4.9 billion tokens (single window)
Data sources14 datasets (Wikipedia full dump + 13 others)
ModelQwen3.6-35B-A3B (MoE, 3B active parameters, Q4)
Hardware1x NVIDIA RTX 3090 (24 GB VRAM)
Runtime94.7 hours (~4 days)
Throughput860 QA pairs/hour
QA pairs generated81,444
Contradictions found184
EngineAlphaLlama

Contradiction Signals

How contradictions were classified

incorrect45

Factually wrong statement

contradicts38

Conflicts with another source

outdated31

Uses superseded data

disputed28

Subject of active debate

inaccurate22

Partially wrong or misleading

unverified20

Claim without citation

Severity Breakdown

184 contradictions classified by severity

Critical22
12%

Direct factual contradictions between Wikipedia articles

High90
49%

Incorrect, inaccurate, or disputed information

Medium72
39%

Controversies, debates, outdated claims

Contradiction Types

What kinds of contradictions did the AI find?

TypeCountDistribution
Comparison68
37%
Temporal42
23%
Aggregation31
17%
Multi-hop27
15%
Causal16
8%

Why Only Long Context Can Find These

Cross-article contradictions require reading multiple articles simultaneously

MethodContextCross-Article Detection
Human editors1 articleCan't see cross-article conflicts
Wikipedia botsFormatting onlyCan't compare facts
Short-context AI (128K)1-2 articlesNeeds someone to say 'compare these'
AlphaLlama4.9B single context windowFinds conflicts nobody knew to look for

Example Findings

Real contradictions discovered in Wikipedia

Criticalcomparison

How does the tenure of Sir John Fortescue as Chancellor of the Exchequer differ between sources?

The biography article states he was the seventh Chancellor of the Exchequer, whereas the disambiguation page incorrectly identifies him as the third.

Two Wikipedia articles directly contradict each other — no human editor would read both pages side by side.

Hightemporal

What discrepancy exists in the founding date of the institution across its main article and the historical timeline?

The main article cites 1843 as the founding year, while the historical timeline page lists 1847.

A 4-year discrepancy in founding dates across two articles about the same institution.

Criticalcomparison

How do population figures for the city differ between the demographics article and the geography overview?

The demographics article reports 2.1 million (2023 census) while the geography overview still cites 1.8 million (2015 estimate).

Outdated data in one article contradicts updated data in another — both appear authoritative.

Explore the Full Dataset

184 contradictions, free and open. CC BY 4.0.