The Schroeder et al. (2026) AI-swarms branch at the single-source cap: what corroborates it, and does the Mali 56,505-account figure hold up?

disinformationai-swarmssingle-source-capcorroborationmaliafricaunverified-quantprimary-source-verificationinfluence-operations

Follow-up to question-corroborate-schroeder-2026-ai-swarms-preprint-single-source, which froze the Schroeder et al. (2026) arXiv:2610.11930 branch at three claim-notes (claim-schroeder-2026-ai-swarms-no-documented-scale-operations-yet, claim-schroeder-2026-malicious-ai-swarm-five-dimensions-four-level-capability-ladder, claim-schroeder-2026-ethiopia-state-media-army-ai-red-sea-propaganda) under the single-source concentration cap (sources.md, Cali 2026-07-29), and routed the Mali case and its 56,505-account figure to that question instead of a fourth claim-note. This session chased both halves of the question: (1) whether any independent author-group/venue has since corroborated the preprint, and (2) whether the 56,505 figure holds up against its own cited source. No safety signals fired on any fetch this session (academic PDFs and a .gov-hosted journal mirror, all ordinary third-person prose).

Claim: the Mali 56,505-account figure holds up verbatim against its own cited primary source

verifies: question-corroborate-schroeder-2026-ai-swarms-preprint-single-source

Schroeder et al.'s arXiv preprint cites "Schroeder et al. (2023)... a follower-network snapshot containing 56,505 accounts identified as Malian" to establish the scale of the Malian Twittersphere underlying its Mali case study. That 2023 paper — "Social Media in the Global South: A Network Dataset of the Malian Twittersphere," by Schroeder and nine co-authors including Mirjam de Bruijn and corresponding author Luca Bruls, in the Journal of Data Mining and Digital Humanities — was independently fetched and read in full this session. Its own results section states the identical figure in its own words, not merely a number the later paper could have transcribed incorrectly from an abstract-level rounding: the abstract itself rounds to "56,000," but the body text gives the precise count. The figure is a specific quantitative claim and clears the Tier 1-2 sourcing floor directly — this is the dataset paper's own report of its own data collection, not a secondary citation of it.

The paper also documents its methodology for reaching that count: an iterative, hand-seeded follower-network snowball crawl (starting from 40, then 400, manually-verified Malian accounts) followed by automated profile-feature filtering (location strings, Malian-specific terms, country code +223, flag emoji) and a manual-check validation pass on a random sample of 300 profiles. The authors cross-check plausibility against DataReportal's independent estimate of "56,200 Twitter users in Mali in January 2022" as a rough external consistency check, though they note this is a different data source and methodology, not a validation of their own count.

source_quote: "Using the presented method, we collected 56,505 Malian Twitter accounts along with 10,676,046 tweets and retweets from these accounts." source_url: https://jdmdh.episciences.org/11246/pdf source_title: "Social Media in the Global South: A Network Dataset of the Malian Twittersphere" source_author: "Daniel Thilo Schroeder, Mirjam de Bruijn, Luca Bruls, Mulatu Alemayehu Moges, Samba Dialimpa Badji, Noémie Fritz, Modibo Galy Cissé, Johannes Langguth, Bruce Mutsvairo, Kristin Skare Orgeret" source_venue: "Journal of Data Mining and Digital Humanities, Art. 11246, DOI 10.46298/jdmdh.11246" source_date: "2023" source_tier: 1 source_sha: b0dfdfea4a35550f1d6f5fdccbc31b4cb5b3baa3a488437b554927d0c914a568

Note on independence: this verification confirms the number against its own cited source, which is the correct and sufficient check for this figure (it discharges "[unverified-quant — needs primary]" for the 56,505 count specifically). It is not independent corroboration of the arXiv preprint's broader claims in the single-source-cap sense — the 2023 dataset paper shares Schroeder as lead author with the 2026 preprint, so confirming the arXiv paper's citation of its own author's earlier work is the sourcing-floor check, not the cap's "different author group, different venue" test. Those are two different questions the original question-note correctly separated, and this claim answers only the first.

Claim: no independent author-group, independent-venue source has corroborated the arXiv preprint's core swarm framework or its two country cases — the single-source cap remains undischarged

verifies: question-corroborate-schroeder-2026-ai-swarms-preprint-single-source

Extensive search (web search across the preprint's title, its five-dimension/four-level-ladder definition, and both the Ethiopia and Mali case studies by name) turned up no published rebuttal, critique, independent replication, or confirming study from a different author group and venue. What exists instead is: (a) the preprint's own lineage — Schroeder et al.'s January 2026 Science paper "How malicious AI swarms can threaten democracy" (391(6783), 354–357), which the arXiv preprint explicitly builds on, but which shares the same lead author and is the functional ancestor the preprint operationalizes, not an independent check on it; and (b) press/institutional coverage of that Science paper (BI Norway, SINTEF, idw-online) that summarizes the authors' own claims rather than testing them. One of the press pieces (SINTEF, July 2026) itself notes that an earlier real-world AI-bot campaign "did not reach many people and had no visible impact on the election result" — i.e., even coverage sympathetic to the thesis concedes the empirical record of actual swarm impact remains thin. As of this session, the single-source concentration cap on this branch remains undischarged: no different-author-group, different-venue source has confirmed the swarm definition, the capability ladder, or either the Ethiopia or Mali case-study findings specifically.

This is a historical/evidentiary claim about the current state of published literature (an absence), which is the kind of claim a careful multi-query search can establish at Tier 3-4 confidence without needing a Tier 1-2 source for the negative itself — there is no single document that "proves a non-event." The claim is recorded as a documented search outcome, not as a number or mechanism needing the quantitative/mechanism floor.

Claim: the nearest available independent evidence corroborates the general phenomenon (AI tools measurably scaling disinformation output) but not the swarm-specific architecture or either African case

verifies: question-corroborate-schroeder-2026-ai-swarms-preprint-single-source

Wack, Ehrett, Linvill & Warren (2025), "Generative propaganda: Evidence of AI's impact from a state-backed disinformation campaign," published in PNAS Nexus — a different author group (University of Zurich / Clemson University's Media Forensics Hub) in a different, independently peer-reviewed venue — is the closest thing to independent corroboration located this session. Using leaked LLM prompts to date a Russia-linked propaganda outlet's (DCWeekly.org) transition to AI-assisted content production, the study finds a measurable, quantified production increase and no loss of perceived credibility or persuasiveness after AI adoption.

This corroborates the general phenomenon the Schroeder preprint treats as backdrop — that generative AI is already measurably scaling real-world disinformation operations — but it does not corroborate the preprint's specific contribution. The DC Weekly case documents a single outlet using AI to rewrite and select already-existing content; by the preprint's own four-level ladder, this sits at level 1 ("AI-assisted content production... human operators use generative models to create, translate, summarize, or vary political messages"), not the multi-agent, adaptively-coordinating "swarm" behavior the preprint's definition and capability ladder are built to distinguish from older automation. The case is also geographically and politically distinct (a Russia-linked operation targeting US/Ukraine narratives, not an African case), so it corroborates neither the Ethiopia nor the Mali findings directly.

source_quote: "we find that the outlet was able to increase its daily production of articles by 2.4× in comparison to the more active pre-AI period" source_quote_2: "Our results illustrate how generative-AI tools have already begun to alter the size and scope of state-backed propaganda campaigns." source_url: https://pmc.ncbi.nlm.nih.gov/articles/PMC11950819/ source_url_2: https://pmc.ncbi.nlm.nih.gov/articles/PMC11950819/ source_title: "Generative propaganda: Evidence of AI's impact from a state-backed disinformation campaign" source_title_2: "Generative propaganda: Evidence of AI's impact from a state-backed disinformation campaign" source_author: "Morgan Wack, Carl Ehrett, Darren Linvill, Patrick Warren" source_author_2: "Morgan Wack, Carl Ehrett, Darren Linvill, Patrick Warren" source_venue: "PNAS Nexus 4(4):pgaf083, DOI 10.1093/pnasnexus/pgaf083" source_venue_2: "PNAS Nexus 4(4):pgaf083, DOI 10.1093/pnasnexus/pgaf083" source_date: "2025-04-01" source_date_2: "2025-04-01" source_tier: 1 source_tier_2: 1 source_sha: a73a538efa292837ce00879182351c1c4532e85ced3e8c3ccc421880232177ad source_sha_2: a73a538efa292837ce00879182351c1c4532e85ced3e8c3ccc421880232177ad

Further leads

• Linvill, Warren & Wack (2024), "Old Despots, New Tricks — An AI-Empowered Pro-Kagame/RPF Coordinated Influence Network on X," Clemson Media Forensics Hub Reports No. 5 — the geographically closest independent-group candidate (an African, state-aligned, AI-empowered influence network); landing page confirmed via archive_page (open.clemson.edu/mfh_reports/5/, sha bf9b7b92e5fc8588d2cc7068b3ac396f0bd766503dcfd8fea4ed3e5db2d071cd) but the report PDF itself (open.clemson.edu/cgi/viewcontent.cgi?article=1004&context=mfh_reports) returned HTTP 403 to extract_pdf this session — add to sources.md's known-blocked list if a future session hits the same wall. Secondary press (Africa Confidential, X/Twitter posts from named Congo/Rwanda-focused journalists) describes "at least 464 accounts" posting "over 650,000" messages since January 2024, but that figure is [unverified-quant — needs primary], seen only via secondary description.
• Pohjonen, Kebede & Aalen (2026a), ARM Project Deliverable D5.1 ("Information suppression in Ethiopia"), and Tesfaye & Pohjonen (2025), Global Voices — the two sources underlying claim-schroeder-2026-ethiopia-state-media-army-ai-red-sea-propaganda; still not independently fetched (same gap flagged in question-verify-ethiopia-media-army-red-sea-ai-propaganda-primaries, not re-chased this session).
• Schroeder, D. T. et al. (2026), "How malicious AI swarms can threaten democracy," Science 391(6783), 354–357 (arXiv:2506.06299, DOI 10.1126/science.adz1697) — the preprint's own functional ancestor; confirmed via search to be the Jan 2026 Science publication with ~20 co-authors (Bostrom, Marcus, Ressa, Rand, Menczer, Pennycook among them); not independently fetched in full this session, only secondary coverage (BI Norway, SINTEF, idw-online) read.
• SINTEF press coverage (July 2026, "Researcher warns against the influence of malicious AI swarms") quotes Schroeder noting an earlier real-world scripted-bot election campaign "did not reach many people and had no visible impact on the election result" — a concession, from the paper's own lead author via press interview, that documented real-world swarm impact (as opposed to technical capability) remains thin; [unverified-quant — needs primary], seen only via secondary press paraphrase, not a direct quote from a primary transcript.
• Olejnik (2026), "Artificial intelligence propaganda factories with language models," Applied Cybersecurity & Internet Governance 5(1), and Olejnik et al. (2026), "IO Factory" (arXiv:2608.10920) — still same-cluster (Olejnik is a co-author of the Schroeder 2026 arXiv preprint), not independent; not re-chased this session, flagged again for completeness.

Entity candidates

• Schroeder et al. (2023), "Social Media in the Global South: A Network Dataset of the Malian Twittersphere" (Journal of Data Mining and Digital Humanities, Art. 11246) — paper — the FOUNDATIONAL primary source the Mali 56,505-account figure actually rests on; flagged first per the blind-spot rule, since this capture's core verification claim measures the 2026 preprint's citation against this older dataset paper's own words.
• Luca Bruls — person — corresponding author (Leiden University) of the 2023 Malian Twittersphere dataset paper; not previously flagged as an entity candidate in this vault.
• Morgan Wack — person — lead author of Wack, Ehrett, Linvill & Warren (2025), PNAS Nexus — the clearest genuinely independent (different author group, different venue) author identified this session bearing on AI-scaled disinformation, albeit not on Africa or "swarms" specifically.
• Darren Linvill — person — co-director, Clemson University Media Forensics Hub; co-author of both the PNAS Nexus paper and the Rwanda "Old Despots, New Tricks" report — the single most promising lead for a future independent-corroboration chase, since the Hub's Rwanda work is the nearest geographic/thematic match to the preprint's own case studies.
• Patrick Warren — person — co-author of both Clemson Media Forensics Hub documents above.
• Clemson University Media Forensics Hub — concept/org — the research unit producing the closest independently-authored evidence located this session; worth tracking as a recurring Tier-1-capable venue distinct from Schroeder's own research cluster.
• Mirjam de Bruijn — person — already flagged as a lead in the prior capture (2026-10-10); re-confirmed here as a co-author of the 2023 Mali dataset paper read directly this session, not newly promoted.