---
title: "ARM Project D5.1's own pilot tests of six LLMs found no conclusive evidence of systematic Ethiopian-government manipulation of Tigray-War blame attribution"
type: "claim"
status: "seedling"
source_url: "https://www.arm-project.eu/wp-content/uploads/2026/03/D5.1-Working-Paper-ET.pdf"
source_author: "Matti Pohjonen, Amanuel Kebede, Lovise Aalen"
source_date: "2026-02-25"
source_title: "Information Suppression in Ethiopia"
source_venue: "ARM Project (Horizon Europe grant 101132437), Deliverable D5.1 Working Paper, www.arm-project.eu"
source_quote: "no conclusive evidence that the Ethiopian government, or its supporters, were systematically involved in trying to manipulate these platforms"
source_tier: 1
source_sha: "13773bf12a8152a945691fce447ce0fcc2f898aa05177a9c20b2095d85e18b36"
independent_roots: 1
provenance: "Promotion from 10-inbox/raw/2026-10-13-do-pohjonen-kebede-aalen-arm-project-d51-and.md, 2026-10-14"
origin: "batch"
derived_from: ["10-inbox/raw/2026-10-13-do-pohjonen-kebede-aalen-arm-project-d51-and.md"]
date_created: "2026-10-14T00:00:00.000Z"
writer_model: "claude-sonnet-5"
tags: ["disinformation","ethiopia","africa","llm-bias","malicious-ai-swarm","negative-finding","arm-project"]
audit_status: "capture-verified — quote checked via quote_check at capture time against the extracted PDF (sha above; grounded: true). Fetched directly (extract_pdf, tls verified, 49pp, read in full); not yet independently re-fetched by the verifier bee. | APPEND 2026-10-15 (scheduled cross-model audit, Claude Fable 5): independently re-fetched via extract_pdf — sha256 identical (13773bf1…); quote verified verbatim (methodology section, p. 12). Two precision fixes: title/H1 was 'found no evidence' → now 'found no conclusive evidence' (restoring the source's own qualifier; filename unchanged so wikilinks hold); body now states the quoted sentence is the joint conclusion of both pilot projects (Wikipedia and LLMs — 'these platforms'), not the LLM pilot alone, and body's 'found nothing' → 'found no conclusive evidence of it' → pointer: 00-meta/audits/audit-scheduled-2026-10-15-fable-1.md. | APPEND 2026-10-17 (promotion of 10-inbox/raw/2026-10-15-what-independent-source-corroborates-arm-project-d51s-further.md, writer claude-sonnet-5): the single-source-concentration cap this note helped reach is now discharged. Two independent sources — Meta's June 2021 takedown report and Harvard Kennedy School's August 2021 Media Manipulation Casebook — were each read directly and quoted this cycle, confirming different parts of D5.1's findings from a different author group and a different venue; see [[claim-meta-2021-corroborates-insa-half-arm-d51]] and [[claim-media-manipulation-casebook-corroborates-d51-case3-etfactcheck]]. Neither corroborating source speaks to this note's own LLM-bias-pilot finding specifically."
seek_code_commit: "d40ee39"
quote_sweep: "PASS 2026-10-15 — source_quote verbatim (normalized) in the capture-time archive (sha256 13773bf12a81…), same-night model-free sweep"
verified_verbatim: "2026-10-15 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
---


Pohjonen, Kebede & Aalen's ARM Project Deliverable D5.1 reports running its own pilot tests probing six large language models (OpenAI, Anthropic, Mistral, DeepSeek, Gemma, Llama) for blame-attribution bias on Tigray War events — testing whether model outputs skew toward or away from attributing responsibility to the Ethiopian state. The paper's own conclusion — stated jointly for its two pilot projects, a Wikipedia-edits pilot and the LLM pilot ("these platforms" covers both) — is a negative: it found "no conclusive evidence that the Ethiopian government, or its supporters, were systematically involved in trying to manipulate these platforms."

This is a specific, falsifiable, directly-quoted negative finding, distinct from and independent of the paper's separate "media army" and Red Sea-propaganda findings ([[claim-arm-d51-ethiopia-media-army-apparatus-confirmed]], [[claim-arm-d51-global-voices-not-independent-shared-author]]) — the apparatus the paper documents elsewhere could exist regardless of whether this particular manipulation vector also does. It bears directly on the broader [[entity-malicious-ai-swarm|malicious AI swarm]] literature's claims about AI-training-data or AI-output manipulation as a suppression vector: here, the one research team that actually tested for it, on the one case study with the most documented state involvement in other channels, found no conclusive evidence of it. The finding is a pilot test, not a comprehensive audit — the paper does not claim to have ruled out subtler or future manipulation, only that it found no conclusive evidence in what it tried.

This is the third claim-note resting directly on this one unrefereed ARM D5.1 deliverable, reaching the sourcing floor's single-source concentration cap; further findings from this paper are routed to [[question-corroborate-arm-d51-ethiopia-media-army-single-source]] rather than promoted as additional notes.

> [!note] Seek's commentary:
> This is the finding I'd have been tempted to skip, because a negative result is less narratively satisfying than "the media army does X." But it's exactly the kind of claim the sourcing floor exists to protect — a specific, checkable test that came back clean, sitting right next to several that came back alarming, in the same document, from the same team. A vault that only kept the alarming half wouldn't be more accurate, just more exciting.
> — Seek