An unexplained death at a Siberian laboratory working on plague was enough to trigger, in less than 48 hours, half a dozen conspiracy narratives in the French-speaking online sphere. Detected by the monitoring tool developed within PROACT EU-Response, this episode shows why tackling misinformation must be regarded as an integral part of preparedness for health crises.
PROACT EU-Response: preparing before the crisis
Building on the EU-RESPONSE project, born out of the COVID-19 crisis, PROACT EU-Response1 aims to equip Europe with a permanent network of adaptive clinical trials able to rapidly evaluate treatments against emerging pathogens. It relies on a network of clinical sites, a network of virology laboratories and teams of methodologists, as well as on a network of social science researchers, and civil society engagement. The core idea: to turn preparedness into a standing capability rather than relying on an improvised reaction.
WS7: trust as infrastructure
Workstream 7 is dedicated to capacity building, community involvement and communication. It trains trial sites, sets up patient and community advisory boards, involves patients and citizens in trial design and in the dissemination of results, and develops communication strategies to counter misinformation and build trust. One of its deliverables, developed by project partner VLS (Vaccination et Lien Social), is an algorithm to detect weak signals of misinformation on social media when a worrying health phenomenon emerges. The system continuously collects messages from French-speaking online channels, sorts them by topic and automatically classifies them as disinformation, facts or uncertainty; its alerts are then confirmed or rejected by a human annotator.
Irkutsk, October 2026: a textbook case
Daria Chipilova, a 28-year-old laboratory technician at the Irkutsk Anti-Plague Institute, died during the night of 1-2 October after being admitted to hospital in Shelekhov. A local media outlet suggested a pneumonic plague contracted after a test tube broke, while the Russian authorities referred to “pneumonia of undetermined aetiology” and stated that no pathogen linked to her work had been detected. The World Health Organisation (WHO) assessed the risk as “low” for Russia and “very low” for Europe, while calling, together with the European Union and the United States, for greater transparency. The diagnosis has yet to be publicly confirmed. All the ingredients of an infodemic are there: a pathogen steeped in history, a high-security laboratory, incomplete official information and a tense geopolitical context.
The monitoring tool
The tool being developed by VLS is a software application that monitors online channels discovered automatically by exploring interactions from a seed of hand-picked channels. Each message received on these channels is prefiltered, on the basis of a wide range of keywords, to determine whether it addresses a health-related topic. Those addressing a health topic are then rated through a SetFit classifier to estimate their likelihood of carrying misinformation. Currently, flagged messages are crossed-checked by a human to further improve the classifier quality. The collected messages and their metadata (propagation, number of reads, etc.) are then used for downstream analysis.
When the suspected plague outbreak occurred, the tool was offline, and was still a toy model, with only a few channels being monitored. Seeing an opportunity to test it during a real-life event, project partners decided to deploy it online, and to tweak it to specifically target plague-related messages (this is why the collection timeframe is 4-13 October, and why the overall message volume is low).
What the tool revealed
The system identified 49 messages posted across 24 channels in French-speaking online media (data collected between 4 and 13 October 2026). The first factual news reports circulated on 4 and 5 October; by 6 October, the first detected conspiracy framings appeared; on 7 October, six narratives coexisted. The most widely shared narrative overall twisted an accurate fact – an mRNA plague vaccine, tested in mice, with studies published between 2023 and 2025 – to suggest that the outbreak had been “premeditated” for a new global vaccination. Other narratives denied the event altogether (“fake plague”, “false flag”) and reinterpreted the official denial as proof of a homicide, or accused the WHO of trying to launch a new “plandemic”, or Kyiv of carrying out a bacteriological attack. The hoax of a “Putin arrest warrant against the bioterrorist Bill Gates who created the plague in a Ukrainian AI lab”, a variant of a recurring and already debunked rumour, became the most widespread narrative and resurfaced, rewritten, on 13 October, after the news flow had dried up.
Three facts emerge. First, these narratives were largely imported from the English-language ecosystem of X and then translated: French-speaking Telegram acted as an amplifier, not as a source. Second, propagation relied on a handful of nodes – prolific relay channels, translators and the discussion groups of large channels, already structured by the COVID pandemic. Finally, most messages did not contain any false statement: it is the framing that misleads, through reversed chronology, irony or the recycling of COVID-19 tropes (mRNA vaccine, intensive care “protocol”, doxycycline in place of hydroxychloroquine).
The “plague in Russia” case in figures :
- Dataset of 49 messages, from 24 French-speaking Telegram channels linked to the topic between 4 and 13 October 2026.
- 6 misinformation narratives about the event, emerging around 48 hours after the first news reports and peaking on 7 and 8 October.
- 11,557 views in less than a week for the dominant narrative, “the vaccine was ready before the outbreak” (8 messages).
- 7 messages on 6 channels, 395 forwards for the hoax of a “Putin arrest warrant against the bioterrorist Bill Gates who created the plague in a Ukrainian AI lab”, revived on 13 October.
What the episode teaches the tools
The audit carried out by the VLS team is also instructive about the limits of automation. In its first version, the classifier picked up only slightly more than half of the misinformation messages: irony (“False Flag” followed by laughing emojis), all-caps texts and insinuations without any explicit falsehood slipped through, while factual relays using anxiety-inducing vocabulary were wrongly flagged. The retrained model performs markedly better, but on a sample too small to draw conclusions. Human validation remains essential, and the team plans to enrich training with ironic examples, to better group variants of the same message and to track hoaxes that have already been identified.
That said, the episode, which detected weak signal despite a small sample size and a very basic classifier, also shows that the tool is working, finding narratives that were not present in mainstream media, but spreading quietly among the conspiracy ecosystems. In an emergency, with the help of this tool, we would be able to try to debunk these narratives before they reach the general population.
This episode will help the VLS team to drastically improve the tool by broadening the scope of monitored channels, allowing to map out the French-speaking conspiracy ecosystem.
Why this matters for preparedness
- A window for action exists. Around 48 hours separated the news from its conspiracy reinterpretation. This suggests that, in similar situations, there could be a critical window of time to prepare debunking content if the alert comes early.
- Information voids are fuel. Genuine uncertainty about the diagnosis and official communication perceived as opaque fed narratives that were contradictory yet converged on one goal: undermining trust in health institutions.
- Narratives are recycled from one crisis to the next. The conspiracy tropes forged during COVID-19 are ready to be reused. Knowing them makes it possible to anticipate their return at the next health emergency, and prepare responses to counter them, including around vaccines and treatments evaluated in trials.
- Trust is a precondition for research. A clinical trial in a crisis depends on the support of patients and healthcare workers. Unaddressed misinformation about a vaccine or a treatment can slow enrolment in trials and jeopardise the generation of the evidence about efficacy and safety of treatment that decision-makers need.
The Irkutsk episode did not lead to an outbreak, and its volume remained modest. That is precisely what makes it a good real-life exercise: it shows that a monitoring tool, combined with human expertise, can detect weak signals early, characterise narratives and their relays, and prepare a response before rumour becomes an obstacle to the public health response.
1 PROACT EU-Response is funded by the Horizon Europe programme (grant agreement No 101156304) and coordinated by ANRS Emerging Infectious Diseases and Inserm. The project, running from January 2025 to January 2030, brings together 26 partners from 10 countries, with a budget of €17 million.
