An auditing system for LLMs designed to identify narrative risks, propaganda, and algorithmic censorship.
Report + benchmarked datasets with IRR
Research example:
🔎EU-funded · Weaponised Algorithms: Auditing AI in the Age of Conflict and Propaganda
Methodology example:
📗EU-funded · Weaponised Algorithms (Methodology)
What it does
– Tests LLMs using multilingual prompts – Measures factual accuracy and propaganda alignment across three key metrics – Documents censorship and manipulation patterns – Provides comparative analysis of model behavior across different global providers
Problem context
AI broadcasts disinformation and alter "versions of truth" depending on the user's language
What it is not
– An automated fact-checking tool – A content moderation system – A predictive model for AI developer intent
Current use
– Completed EU-supported audit of 6 global LLMs (OpenAI, Google, Anthropic, xAI, DeepSeek, Yandex) – Available for custom institutional deployment (elections, conflicts, narrative monitoring)
Collaboration
Request-based
Limitations
– Results methodology and baseline facts adoption for each case – Results are model-version specific – Requires regular updates to reflect rapid AI development