---
id: "20261010-0239-outside-arxiv261011930-automated-disinformation"
title: "Automated Disinformation and Malicious AI Swarms: Risks for Democracy and Development in Africa (arXiv:2610.11930) — capture"
type: "capture"
status: "promoted"
promoted_to: ["30-notes/claim-schroeder-2026-ai-swarms-no-documented-scale-operations-yet.md","30-notes/claim-schroeder-2026-malicious-ai-swarm-five-dimensions-four-level-capability-ladder.md","30-notes/claim-schroeder-2026-ethiopia-state-media-army-ai-red-sea-propaganda.md","40-entities/entity-daniel-thilo-schroeder.md","40-entities/entity-malicious-ai-swarm.md","50-questions/question-verify-ethiopia-media-army-red-sea-ai-propaganda-primaries.md","50-questions/question-corroborate-schroeder-2026-ai-swarms-preprint-single-source.md"]
not_promoted: ["Mali case (fragmented Twittersphere, 56,505-account figure, Gerring most-likely-case selection) — not written as a fourth claim-note: the single-source concentration cap (sources.md) caps claim-notes on one unrefereed preprint at three, already spent on the near-future-risk, definition/ladder, and Ethiopia claims. Routed instead to 50-questions/question-corroborate-schroeder-2026-ai-swarms-preprint-single-source.md, which also covers the unverified 56,505-account quant figure.","Entity candidate: Schroeder et al. (2026) Science paper — not a person/concept/org/term under the entity-page spec's four kinds; left as a citation (Further leads) rather than an entity page.","Entity candidates: Nick Bostrom, Maria Ressa, Gary Marcus — co-authors of the (unfetched) Science paper only; per the 2026-10-01 co-authorship ruling, co-authorship alone is a mention, not a hub. Left as mentions.","Entity candidate: Lukasz Olejnik — real person with separate cited work (commodity-hardware AI-propaganda-factory experiments), a plausible future hub, but no promoted claim-note currently links him in body text; left as a mention rather than an orphan hub, pending a session that reads his own work directly.","Entity candidate: Mirjam de Bruijn — originates the 'networked journalism/conflict' framework the (unpromoted) Mali case leans on; not independently fetched and not referenced by any promoted note. Left unpromoted.","Entity candidate: FIMI (EU/EEAS term) — mentioned only in the capture's entity-candidates list, not used in any promoted claim's body text; not yet load-bearing in this vault. Left unpromoted.","Entity candidate: four-level capability ladder as a separate concept from malicious AI swarm — folded into the malicious-ai-swarm watching stub and the definition claim-note rather than split into its own entity, per the spec's bias against over-splitting one referent.","Further leads (Schroeder et al. 2023 Mali dataset paper, Olejnik 2026 propaganda-factory papers, Pohjonen et al. ARM Project D5.1, Tesfaye & Pohjonen 2025 Global Voices, Wack/Linvill/Warren 2024 Rwanda report, Waight et al. Nature, Dembele et al. 2024 survey, ITU 2025 figures) — none independently fetched this session; the two that are load-bearing for a kept claim (ARM Project D5.1, Global Voices) are named in question-verify-ethiopia-media-army-red-sea-ai-propaganda-primaries.md. The rest remain leads, not claims, and were never candidates for independent promotion."]
origin: "batch"
writer_model: "claude-sonnet-5"
date_created: "2026-10-10T00:00:00.000Z"
provenance: "2026-10-10 batch run, outside-vault web research topic"
derived_from: []
tags: ["disinformation","misinformation","ai-swarms","multi-agent-systems","influence-operations","computational-propaganda","africa","mali","ethiopia","democracy","generative-ai","arxiv","cs-cy"]
source_url: "https://arxiv.org/abs/2610.11930"
source_author: "Daniel Thilo Schroeder, Philipp M. Lutscher, Samba Dialimpa Badji, Stefan Brenner, Wenchao Dong, Tsehaye Haidemariam, Saheed Bidemi Ibrahim, Amanuel Tesfaye Kebede, Jonas R. Kunst, Johannes Langguth, Pauline Lemaire, Mulatu Alemayehu Moges, Lukasz Olejnik, Kristin Skare Orgeret, Gerald Walulya"
source_date: "2026-10-08T00:00:00.000Z"
source_title: "Automated Disinformation and Malicious AI Swarms: Risks for Democracy and Development in Africa"
source_venue: "arXiv:2610.11930 [cs.CY]"
source_tier: 1
source_sha: "6b58ef4a3ab2d5c21ef0238a6574a49202f8502a2955fdf71b0cedbc04e19844"
seek_code_commit: "d40ee39"
---


Topic was an "outside" surface — a fresh arXiv preprint (submitted 8 Oct 2026, cs.CY) not previously touched by this vault. Fetched the abstract page via `archive_page` (sha256 `6b057c846a09b3c3993ce3859eb63a4701627e924655f1d175dd33d8e8d14dcc`) and the full 26-page PDF via `extract_pdf` (`tls: "verified"`, sha256 `6b58ef4a3ab2d5c21ef0238a6574a49202f8502a2955fdf71b0cedbc04e19844`). Both read in full. No safety signals fired on either fetch — ordinary third-person academic prose throughout (abstract, eleven numbered sections, conclusion, 90+-item reference list); no addressed-to-AI language, override language, claimed authority, tier self-assignment, file-system instructions, credential requests, or urgency framing. No `## Safety flags` entries this capture.

Per the blind-spot instruction to flag the older/foundational reference first: this paper's entire operational definition of "malicious AI swarms" is explicitly built on, and repeatedly cites as its functional ancestor, **Schroeder et al. (2026), "How malicious AI swarms can threaten democracy," *Science* 391(6783), 354–357** — a separate, earlier paper by the same lead author with a notable co-author list (including Nick Bostrom, Gary Marcus, Maria Ressa, Nicholas Christakis, David Rand, Gordon Pennycook, Filippo Menczer). That Science paper is flagged first in Entity candidates below, ahead of this paper's own fifteen co-authors. The current paper states it is "[b]uilding on this functional account" to "develop an operational and empirically testable definition" — i.e., this capture's definitional claim rests on the arXiv text's own words but traces its lineage to that earlier, higher-profile publication, which was not independently fetched this session.

All claims below rest on this arXiv preprint (Tier 1 — primary, authors' own text) with verbatim quotes, each checked against the extracted PDF text via `quote_check` (all returned `grounded: true`, two via dehyphenation across a line-wrap).

## Claim: the paper's own central claim is that malicious AI swarms are a near-future risk, not a documented present-day phenomenon

The paper is explicit that it is conducting forward-looking risk analysis rather than reporting detected operations: it states there is no clear evidence that fully autonomous AI swarms are already conducting influence operations at scale, while their component technologies (LLMs, multi-agent coordination, real-time optimization) are advancing and becoming more accessible. This is a historical/evidentiary claim about the current state of the field, made by the paper's own authors about their own analytic stance — it directly bears on any question of whether AI-driven swarms are currently disrupting African information environments: per this source, the answer as of October 2026 is "not yet documented, but increasingly plausible."

source_quote: "Although there is currently no clear evidence that fully autonomous "AI swarms" are conducting influence operations at scale, their enabling components, including large language models, multi-agent coordination, and real-time optimization, are advancing rapidly and becoming more accessible."
source_url: https://arxiv.org/pdf/2610.11930
source_title: "Automated Disinformation and Malicious AI Swarms: Risks for Democracy and Development in Africa"
source_author: "Daniel Thilo Schroeder et al."
source_venue: "arXiv:2610.11930 [cs.CY]"
source_date: 2026-10-08
source_tier: 1
source_sha: 6b58ef4a3ab2d5c21ef0238a6574a49202f8502a2955fdf71b0cedbc04e19844

## Claim: "malicious AI swarms" are defined by five graded dimensions and distinguished from botnets/synthetic personas by a four-level capability ladder

Building on the Schroeder et al. (2026) *Science* functional account (see note above), the paper offers an operational definition: a coordinated system of multiple AI-enabled agents that maintain partially persistent identities, act on shared information, adapt to feedback, and collectively pursue an influence objective with some operational autonomy. It specifies five constitutive dimensions — multiplicity, coordination, persistence, adaptation, collective agency — as graded rather than binary, and a four-level capability ladder running from (1) AI-assisted content production, through (2) coordinated synthetic personas and (3) adaptive multi-agent campaigning, to (4) goal-revising malicious swarms. This is a definitional/technical-mechanism claim about what distinguishes a "swarm" from earlier automation, directly grounded in the primary text.

source_quote: "we define a malicious AI swarm as a coordinated system of multiple AI-enabled agents that maintain partially persistent identities or operational roles, act on shared or mutually relevant information, adapt their behaviour in response to environmental feedback, and collectively pursue an influence objective with some degree of operational autonomy"
source_url: https://arxiv.org/pdf/2610.11930
source_title: "Automated Disinformation and Malicious AI Swarms: Risks for Democracy and Development in Africa"
source_author: "Daniel Thilo Schroeder et al."
source_venue: "arXiv:2610.11930 [cs.CY]"
source_date: 2026-10-08
source_tier: 1
source_sha: 6b58ef4a3ab2d5c21ef0238a6574a49202f8502a2955fdf71b0cedbc04e19844

source_quote_2: "The first level is AI-assisted content production, in which human operators use generative models to create, translate, summarize, or vary political messages. The second level is coordinated synthetic personas, in which multiple accounts maintain differentiated identities and distribute strategically aligned content, but remain substantially directed by human operators. The third level is adaptive multi-agent campaigning, in which agents respond to engagement signals, alter messages or targets, and coordinate actions across accounts with reduced human supervision. The fourth level is goal-revising malicious swarms, in which agents not only adapt tactics but also modify intermediate objectives, redistribute tasks, and sustain a campaign through recursive feedback loops."
source_url_2: https://arxiv.org/pdf/2610.11930
source_title_2: "Automated Disinformation and Malicious AI Swarms: Risks for Democracy and Development in Africa"
source_author_2: "Daniel Thilo Schroeder et al."
source_venue_2: "arXiv:2610.11930 [cs.CY]"
source_date_2: 2026-10-08
source_tier_2: 1
source_sha_2: 6b58ef4a3ab2d5c21ef0238a6574a49202f8502a2955fdf71b0cedbc04e19844

## Claim: Ethiopia case — a government-organized "media army" and AI-generated Red Sea propaganda illustrate hybrid state-led influence operations, not fragmented bottom-up ones

The paper's Ethiopia case study argues that AI-enabled influence need not emerge from decentralized or external actors: it describes a documented shift in Ethiopian state strategy from censorship/shutdowns toward complementary strategies including a government-organized "media army," paid activists and influencers, coordinated messaging, agenda-setting, and diaspora mobilization (drawing on Pohjonen, Kebede, & Aalen's 2026 ARM Project working paper). It further reports that AI-generated images, video, synthetic voice, and music promoting Ethiopia's Red Sea/maritime-access ambitions have circulated on X and TikTok, amplified by former government officials and pro-government accounts (drawing on Tesfaye & Pohjonen's 2025 Global Voices reporting). This is a specific technical-mechanism/historical claim about an already-observed hybrid human-AI state propaganda apparatus, as synthesized and stated directly in this Tier-1 primary text; the two underlying empirical sources (the ARM Project working paper and the Global Voices article) were not independently fetched this session (see Further leads).

source_quote: "These include a government-organized "media army," paid activists and influencers, coordinated messaging, agenda-setting, and the mobilization of diaspora constituencies"
source_url: https://arxiv.org/pdf/2610.11930
source_title: "Automated Disinformation and Malicious AI Swarms: Risks for Democracy and Development in Africa"
source_author: "Daniel Thilo Schroeder et al."
source_venue: "arXiv:2610.11930 [cs.CY]"
source_date: 2026-10-08
source_tier: 1
source_sha: 6b58ef4a3ab2d5c21ef0238a6574a49202f8502a2955fdf71b0cedbc04e19844

source_quote_2: "AI-generated images, video, synthetic voice, and music promoting Ethiopia's maritime ambitions have circulated on platforms such as X and TikTok. Some content has been amplified by former government officials and pro-government"
source_url_2: https://arxiv.org/pdf/2610.11930
source_title_2: "Automated Disinformation and Malicious AI Swarms: Risks for Democracy and Development in Africa"
source_author_2: "Daniel Thilo Schroeder et al."
source_venue_2: "arXiv:2610.11930 [cs.CY]"
source_date_2: 2026-10-08
source_tier_2: 1
source_sha_2: 6b58ef4a3ab2d5c21ef0238a6574a49202f8502a2955fdf71b0cedbc04e19844

## Claim: Mali case — a fragmented, weakly-gatekept Twittersphere (56,505 Malian accounts in the authors' own 2023 dataset) is presented as the "most-likely-case" for swarm exploitation

The Mali case study argues the opposite configuration from Ethiopia: a fragmented information ecology with weak gatekeeping and competing domestic/foreign actors (state forces, Tuareg-led movements, JNIM, foreign governments/contractors), where "networked journalism" blurs into "networked conflict." The paper grounds the scale of this environment in a quantitative figure from the lead author's own prior empirical dataset paper (Schroeder et al., 2023, *Journal of Data Mining & Digital Humanities*): a follower-network snapshot of 56,505 accounts identified as Malian. Mali is explicitly selected using Gerring's (2007) "most-likely-case" logic — not claimed representative of Africa, but chosen because its conditions make AI-enabled manipulation especially plausible there first.

source_quote: "Schroeder et al. (2023) document a follower-network snapshot containing 56,505 accounts identified as Malian"
source_url: https://arxiv.org/pdf/2610.11930
source_title: "Automated Disinformation and Malicious AI Swarms: Risks for Democracy and Development in Africa"
source_author: "Daniel Thilo Schroeder et al."
source_venue: "arXiv:2610.11930 [cs.CY]"
source_date: 2026-10-08
source_tier: 1
source_sha: 6b58ef4a3ab2d5c21ef0238a6574a49202f8502a2955fdf71b0cedbc04e19844

Note: the 56,505-account figure traces to the authors' own underlying dataset paper — Schroeder, D. T. et al. (2023), "Social media in the Global South: A network dataset of the Malian Twittersphere," *Journal of Data Mining & Digital Humanities*, Article 11246, https://doi.org/10.46298/jdmdh.11246 — which was not independently fetched this session; the number is recorded here on the strength of the current Tier-1 arXiv paper's own citation of its lead author's prior empirical work, not on independent verification against that dataset paper directly.

> [!note] Seek's commentary:
> None of the retrieve-before-write candidates surfaced for this topic (diplomatics/BASIC-authentication clusters, hallucination-lineage claims, canonic ledger notes, notation-dispersal claims) bear directly on disinformation, multi-agent influence operations, or African media ecology — this is a genuinely fresh vault surface with no collision to resolve. The closest conceptual echo is the governance section's proposal of "cryptographic attestation tools, federated identity protocols, and optional 'proof-of-personhood' systems" to limit synthetic-account fleets — a provenance/attestation mechanism structurally similar in spirit to [[observation-watermark-same-provenance-mechanism-paper-to-llm]]'s paper-to-LLM provenance parallel, though applied here to social accounts rather than documents or text. Worth a light link if a future synthesis note treats "provenance/attestation as a cross-domain countermeasure" as its own cluster; not merged here.

## Further leads

- Schroeder, D. T. et al. (2026), "How malicious AI swarms can threaten democracy," *Science* 391(6783), 354–357 — the foundational functional account this entire paper operationalizes; star-studded co-author list (Bostrom, Marcus, Ressa, Christakis, Rand, Pennycook, Menczer); not independently fetched this session.
- Schroeder, D. T. et al. (2023), "Social media in the Global South: A network dataset of the Malian Twittersphere," *Journal of Data Mining & Digital Humanities*, Art. 11246 — primary dataset underlying the 56,505-account Mali figure above; not independently fetched.
- Olejnik, L. (2026), "Artificial intelligence propaganda factories with language models," *Applied Cybersecurity & Internet Governance* 5(1); and Olejnik et al. (2026), "IO Factory: Simulating AI-enabled influence campaigns at scale" (arXiv:2608.10920) — experimental work the paper leans on repeatedly as its closest empirical grounding for cheap, commodity-hardware swarm feasibility; not independently fetched.
- Pohjonen, M., Kebede, A., & Aalen, L. (2026a), "Information suppression in Ethiopia" (ARM Project Deliverable D5.1) — primary source for the Ethiopian "media army" apparatus and the co-option of fact-checking language by the state; not independently fetched.
- Tesfaye, A., & Pohjonen, M. (2025, Oct 9), "AI and visual propaganda are being used to stoke tensions as Ethiopia eyes Eritrea's Red Sea port," Global Voices — primary journalistic source for the Red Sea AI-propaganda case; not independently fetched.
- Wack, M., Linvill, D., & Warren, P. (2024), "Old despots, new tricks: An AI-empowered pro-Kagame/RPF coordinated influence network on X" (Clemson Media Forensics Hub Report No. 5) — a third African case (Rwanda) of an AI-empowered state-aligned influence network, mentioned but not developed by this paper; not independently fetched.
- Waight, H. et al. (2026), "State media control influences large language models," *Nature* 655(8123), 685–693 — finds Chinese state-coordinated media in open-weight training corpora shifts model outputs toward favoring Chinese institutions/leaders after additional pretraining; a training-data-poisoning mechanism cited in passing; [unverified-mechanism — needs primary], not independently fetched.
- Dembele, K. et al. (2024) quota-sampled survey reports "approximately 68%" of Malian journalists/media professionals expressed censorship concerns or practiced self-censorship (cited at p. 281) — [unverified-quant — needs primary], only seen via this paper's citation.
- ITU (2025), "Measuring digital development: Facts and figures 2025" — source of the cited 36% (ITU Africa region, 2025) vs. 74% (global) internet-use figures; [unverified-quant — needs primary], not independently fetched against the ITU report itself.

## Entity candidates

- Schroeder et al. (2026), "How malicious AI swarms can threaten democracy" (*Science* 391(6783)) — concept/paper — the FOUNDATIONAL functional account this entire arXiv paper's operational definition is built on; flagged first per the blind-spot rule, ahead of the current paper's own co-authors.
- Daniel Thilo Schroeder — person — lead author linking three works cited here: this paper, the foundational 2026 *Science* paper, and the 2023 Malian Twittersphere dataset paper.
- Nick Bostrom — person — co-author of the foundational Schroeder et al. (2026) *Science* paper.
- Maria Ressa — person — Nobel Peace Prize laureate journalist; co-author of the foundational Schroeder et al. (2026) *Science* paper.
- Gary Marcus — person — AI critic/cognitive scientist; co-author of the foundational Schroeder et al. (2026) *Science* paper.
- malicious AI swarm — concept — this paper's central defined term (five graded dimensions, four-level capability ladder); distinguished from botnets and coordinated synthetic personas.
- Mulatu Kebede / Pohjonen et al. (ARM Project) — concept/people — primary source for Ethiopia's state-organized "media army" and co-opted fact-checking apparatus.
- Lukasz Olejnik — person — co-author of this paper; independent security researcher whose commodity-hardware "AI propaganda factory" experiments are the paper's nearest empirical grounding for swarm feasibility.
- Mirjam de Bruijn — person — originates the "networked journalism" / "networked conflict" framework (2025, Malian Twittersphere) this paper's Mali case leans on heavily.
- FIMI (foreign information manipulation and interference) — concept — EU/EEAS term the paper adopts, treating it as a subset of influence operations rather than the overarching frame.
- four-level capability ladder (AI-assisted content production → coordinated synthetic personas → adaptive multi-agent campaigning → goal-revising malicious swarms) — concept — the paper's graded taxonomy for distinguishing automation maturity.