Annex
Annex brings propaganda classification into continuous information-environment monitoring, where teams can connect detected techniques to narratives, actors and patterns of spread.
Explore AnnexCatch AI-generated propaganda before it shapes your audienceWASPer by Identrics
Open-source models that flag synthetic text and classify propaganda techniques across user-generated content - multilingual, DSA-aligned, ready for production scale.
A four-level taxonomy and pre-trained / fine-tuned LLMs make your moderation more reliable and online discourse harder to manipulate.
Live comment stream
wasper.v1
- @marta_v2m
Just got back from the market - prices on tomatoes are wild this week.
- @news_alert_bg1m
BREAKING: officials refuse to comment on the growing scandal - the public deserves answers now, before it's too late.
- @voice_of_reason_2025just now
Everyone is finally waking up - the entire establishment has been lying to us for years. Only the brave few dare to speak.
AI-Generated · 94%Propaganda: Bandwagon
F1 0.85 · multilingual1 flagged · 2 clean
0.85
F1 score
Multilingual
Languages
Open source
Since 2025
NGI SEARCH
Grant #101069364
Open-sourced 2025. Built under NGI SEARCH grant #101069364.
What WASPer Does
Stop AI propaganda at its source.
The scale and subtlety of AI-driven manipulation has outpaced manual moderation. WASPer closes the gap.
Unmoderated Content
What your team is fighting today
- Manual review can't keep pace with the volume of user-generated content.
- 457% more AI-generated misinformation per recent studies on platforms known for spreading disinformation.
- Coordinated campaigns and troll networks are easy to miss when only humans are watching.
- Trust and safety budgets get eaten by reactive firefighting instead of prevention.
VS
With WASPer
What your team ships instead
- Detects synthetic content with very high accuracy across long and short form text.
- Filters propaganda automatically with an LLM-based binary classifier.
- Analyses techniques across a four-level hierarchical taxonomy.
- Adapts across multiple languages, starting with English and Bulgarian.
How WASPer Works
A multi-step pipeline built for the precision and scale you need.
State-of-the-art AI models and a robust hierarchical taxonomy work together so even the subtlest manipulation in your feed gets flagged.
- 01
Human vs AI-Generated Detection
Identifies whether content is human-generated or AI-generated with very high accuracy across long and short form text.
AI96%Human4% - 02
Binary Propaganda Detection
An LLM-based model determines whether the synthetic content carries propaganda or is benign.
Propaganda1Benign0 - 03
Multi-Label Propaganda Classification
Breaks down detected propaganda into distinct techniques using a hierarchical taxonomy.
BandwagonFear AppealWhataboutismLoaded Language - 04
Multilingual Processing
Excels at multilingual content, starting with English and Bulgarian, making WASPer a globally adaptable solution.
MultilingualENBGDEFRESITPLROELHU+1
Powers in Production
From algorithmic capability to operational intelligence.
WASPer remains an open-source project for synthetic-text and propaganda classification. Its research also informs an operational workflow in the Identrics product ecosystem.
Technological Highlights
A four-level taxonomy systematically categorises propaganda, from general detection to specific techniques. Level 1 determines whether text is human- or AI-generated. Level 2 identifies the presence of propaganda. Level 3 groups it into broader categories such as self-identification or defamation. Level 4 pinpoints specific techniques like Bandwagon, Fear Appeal, or Whataboutism.
Who Uses WASPer
Built for teams protecting the integrity of public discourse.
Each panel shows a real-world example of how WASPer fits into the workflow.
Audience · 01
For researchers and innovators
Our open-source resources provide a foundation for innovation in AI against synthetic propaganda. You can collaborate with us to:
- Expand the capabilities of WASPer's models.
- Develop new features and applications tailored to your research goals.
- Contribute to the ongoing refinement of the hierarchical taxonomy.
Example scenario
A university team fine-tunes WASPer on a Balkan-language corpus, publishes updated weights back to the community and cites the model in their next peer-reviewed paper.
Audience · 02
For media platforms
WASPer easily integrates with media outlets, online publishers, and content moderators. By embedding its AI-powered tools, platforms can:
- Automate the detection of synthetic and propagandistic content.
- Enhance trust with their audience by delivering verified, reliable information.
- Scale moderation efforts without compromising quality.
Example scenario
A national news portal routes every user comment through WASPer before publication, auto-holding AI-generated propaganda for review and cutting moderator load by half.
Audience · 03
For businesses and organisations
WASPer's customisable nature can be tailored to meet the unique challenges of your industry - whether combatting disinformation, monitoring public sentiment, or facilitating content analysis.
- Deploy self-hosted for full data control.
- Adapt the taxonomy to your risk categories.
- Feed detections straight into your existing SOC or trust and safety stack.
Example scenario
A financial services firm screens branded social mentions for coordinated smear campaigns, escalating flagged clusters to comms and legal within the same hour.
Use Cases & Real-World Impact
Identifying disinformation campaigns
Disinformation campaigns often leverage AI to produce persuasive, targeted content at scale. WASPer identifies these operations by detecting synthetic propaganda and analysing the techniques used, allowing platforms and organisations to respond promptly.
Enhancing UGC moderation
With user-generated content increasing exponentially, traditional moderation methods are no longer sufficient. WASPer automates harmful content identification, ensuring faster and more consistent moderation and freeing human moderators for high-priority tasks.
Building trust in journalism
WASPer safeguards editorial workflows by detecting synthetic and propagandistic content, enabling publishers to maintain the credibility of their reporting and deliver accurate, reliable information.
Protecting digital spaces
WASPer identifies not just manipulative content but also the strategies behind it, equipping platforms to build safer, more respectful online environments and foster community trust.
Optimising media listening platforms
For platforms monitoring media sentiment and engagement, WASPer offers valuable insights by analysing sentiment, detecting propaganda, and categorising content at scale.
Supporting research and policy
Academic teams, think tanks and regulators use WASPer to study propaganda techniques at scale, ground policy recommendations in evidence, and stress-test platform compliance with DSA obligations.
Team Behind WASPer
Meet the people building WASPer.
Yolina Petrova, PhD
Chief Operations & AI Officer
Boryana Kostadinova
Data Scientist & Linguist
Bogomil Katanov
AI Engineer
Open source · DSA-compliant
Want WASPer running inside your stack?
Deploy it on your own infrastructure, or partner with us to push open-source disinformation defence further.
NGI SEARCH
© 2022-2025 NGI SEARCH. Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or European Commission. Neither the European Union, nor the granting authority can be held responsible for them. Funded within the framework of the NGI SEARCH Project under grant agreement No. 101069364.