Wack et al. (2025) find a state-backed outlet increased daily output 2.4x after adopting AI, with no loss of credibility — independent evidence for the AI-swarms preprint's backdrop but not its swarm architecture

disinformationgenerative-aiinfluence-operationscomputational-propagandaai-swarmscorroboration

Morgan Wack, Carl Ehrett, Darren Linvill and Patrick Warren — a different author group (University of Zurich and Clemson University's Media Forensics Hub) publishing in PNAS Nexus, a venue independent of Schroeder et al.'s research cluster — used leaked LLM prompts to date a Russia-linked propaganda outlet's (DCWeekly.org) transition to AI-assisted content production. They find the outlet "was able to increase its daily production of articles by 2.4× in comparison to the more active pre-AI period," with no measured loss of perceived credibility or persuasiveness after AI adoption, concluding that "generative-AI tools have already begun to alter the size and scope of state-backed propaganda campaigns."

This is the nearest thing to independent corroboration located for the malicious AI swarms preprint, and it corroborates a specific and limited piece: the backdrop claim that generative AI is already measurably scaling real-world disinformation production. It does not corroborate the preprint's distinctive contribution. By the preprint's own four-level capability ladder, the DC Weekly case sits at level 1 — "AI-assisted content production," human operators using generative models to create, translate, summarize, or vary messages — not the adaptive, multi-agent, coordinating "swarm" behaviour (levels 3–4) the ladder exists to distinguish from older automation. The case is also a Russia-linked operation targeting US/Ukraine narratives, so it bears on neither the Ethiopia nor the Mali African case study specifically.

The distinction is why the malicious AI swarm branch remains under its single-source concentration cap: independently-authored, peer-reviewed evidence now supports the phenomenon the preprint treats as given, but the swarm-specific architecture and both African cases are still carried by one author cluster. See question-corroborate-schroeder-2026-ai-swarms-preprint-single-source, which this note partially answers — the general phenomenon is corroborated; the swarm framework is not.