Auditing AI in the Age of Conflict and Propaganda

Goal of research

To systematically assess how leading AI models respond to conflict-related disinformation prompts about Ukraine, focusing on factual accuracy, propaganda inclusion, tone, consistency, and safety behavior.

The goal is to evaluate accountability and bias in AI systems using a transparent, replicable framework — not to prove intent or infer causality.

Developed by Ihor Samokhodskyi. https://policygenome.org/

Want to see how AI talks about your issue, country or election?

→ Request a custom audit / methodology: Email, LinkedIn, WhatsApp.

Models

• ChatGPT-5 — OpenAI.
• Sonnet 4.5 — Anthropic.
• Gemini 2.5 flash — Google.
• Grok 4 Fast - xAI.
• DeepSeek-V3 — DeepSeek.
• YandexGPT (via Alice) — (Yandex).

How we research

Languages: each prompt will be tested in English, Ukrainian, Russian.

Test Controls:

• All tests must be conducted in a new, "zero-shot" chat session (no prior conversation history).
• The Temperature parameter will be set to 0.1 to ensure maximum reproducibility (unless we face a technical restrictions).

⸻