Thread by @antibot4navalny

Building a sophisticated troll network at scale had never been cheaper and easier—thanks to AI capabilities.

Even if you are not Russia, China or Iran, now you can influence Western countries towards far-right, populist parties (AfD, RN, PiS…) and isolating from each other.

1/

This is exactly what UAE* did: with AI-powered operation targeting 🇺🇸🇨🇦🇬🇧🇫🇷🇩🇪🇵🇱🇪🇸🇮🇹 and the Gulf states, it weaponized COP29, a UN climate change conference in Azerbaijan.

2300+ accounts in 8 languages pushing country-specific narratives with highly-realistic fake personas.

2/

With broad variety of pro- and anti- "opinions" in politics, economy, social, emerging markets, climate, all Western-targeting narratives agree on:

• desire to change of elected officials, established institutions and status quo
• UAE being the most successful alternative

3/

Across AZ/EN/FR/DE languages, Azerbaijani leadership and president Aliyev are endorsed for prosperity, stability and sovereignty.

In Nagorno-Karabakh conflict, Azerbaijan is portrayed as a victim of Armenia's aggression supported by biased, hypocritical France and the US.

4/

The powered-by-AI nature of this network was revealed by variety of glitches in LLM-generated tweet texts and GAN images.

It will take only a few generations of GPT models to get rid of artifacts like these. Meanwhile, some of our favorites in texts:

5/

GAN cityscapes are sometimes even more betraying, like:

• a wind farm in the middle of Paris⁩
• double Gherkin skyscrapers in London
• two CN Towers in Toronto
• two Eiffel towers

6/

Most important evidence of inauthenticity and coordination for the network as a whole are:

• signature combination of creation date and first tweet across accounts
• broad variety of features that connect together all the 3 first-tweet-based cohorts

7/