The evolution of the AI workflow in propaganda #
In previous years, a typical influence campaign required a small army of human operators to write posts, manage accounts, and engage in arguments. Now, the deployment of Large Language Models (LLMs) has fundamentally changed the cost-benefit analysis for these actors. They can now run an entire ecosystem of personas from a single server.
The modern playbook generally follows a specific pattern:
Persona Cultivation: Instead of creating accounts that post nothing but political slogans, they use AI to generate months of "lifestyle" content—photos of food, travel, or hobbies—to build credibility before the political pivot.Micro-Targeting via Sentiment Analysis: They use LLM agents to scan social media trends and identify specific pockets of social tension. Once a topic is identified, the AI generates hyper-specific content designed to inflame that exact demographic.Language Nuance: This is where it gets dangerous. Older botnets were easy to spot because of broken grammar. Modern models can handle slang, regional dialects, and even sarcastic tones, making the "uncanny valley" effect almost non-existent for the average user.
Why traditional detection is failing #
Most people think of bot detection as looking for high-frequency posting. If an account posts 500 times an hour, it's a bot. That's a beginner-level approach. The new wave of covert influence is much more patient. These accounts might post once a day, just like a real human, but their entire "personality" is a mathematical construct designed to steer a specific conversation.
We are seeing a move toward "hybrid" campaigns. This involves a human-in-the-loop system where an AI generates ten different variations of a narrative, and a human operator picks the one that is currently gaining the most traction. This makes the content feel organic because it is, in a sense, responding to real-time human emotion.
Identifying the fingerprints of automated influence #
If you are looking for ways to spot these operations, don't look for the volume; look for the narrative synchronization. When multiple, seemingly unrelated accounts across different platforms (X, Reddit, Facebook) start using the exact same unique phrasing or "new" conspiracy theories within a very tight window, you are likely looking at a coordinated deployment. The technical challenge for security researchers is that as we build better classifiers to detect AI text, the actors are using those very same tools to "humanize" their output. It’s a recursive arms race. We need to move toward analyzing the metadata of influence—the way information spreads through network clusters—rather than just analyzing the text itself. The text is becoming too perfect to trust.
Nvidia chips are showing up in Russian missile guidance systems 9d ago
Nvidia chips are showing up in Russian missiles again 11d ago Private companies can now launch authorized cyberattacks under 12d ago
OpenAI is holding back Astra because of cybersecurity risks 18d ago
[AI Era Web Safety: How Chrome Is Beefing Up the Web 26d ago](/en/news/4476/)
[Claude Code: My Take on the Rogue Agent Incident 27d ago](/en/news/4328/)
Next Bill Gates thinks we are flying blind with AI development →