Invisible Persuasion: The Shift From Content to AI Retrieval in the Age of Disinformation

Artificial intelligence is shifting the nature of disinformation from the creation of deceptive content to the manipulation of how machines retrieve, summarize, and recommend information. According to a report from The Cipher Brief, the primary target for influence is no longer the human reader but the large language model (LLM) acting as the information gateway, a transition that moves the information environment from an “attention economy” to a “cognitive economy.”

This shift means that influencing what an AI assistant retrieves is more valuable than influencing what millions of people read directly. Adversaries are expected to move away from “flooding the zone” with fake articles to influence individual people and instead seek to influence the systems that provide answers to everyone.

The Rise of Sovereign AI and Model Diversity

The global AI landscape is diversifying as nations pursue “cognitive sovereignty” through the development of foundational models that reflect their own national interests, values, and historical contexts. The Center for a New American Security (CNAS) Sovereign AI Index reports that more than 130 active national sovereign initiatives exist across more than 60 countries.

This proliferation of models creates a fragmented information environment where different versions of reality emerge based on how specific models are taught to reason and prioritize. The Hugging Face Model Hub, the world’s top repository, currently tracks over 2.9 million machine learning models, each acting as an editor with its own training corpus and retrieval strategy.

Generative Engine Optimization and Invisible Persuasion

Bad actors are transitioning from search engine optimization (SEO) to generative engine optimization (GEO). This process involves engineering content specifically to influence what AI systems retrieve and cite, a tactic described by The Cipher Brief as “invisible persuasion.” Unlike traditional propaganda, machine-led information thrives on appearing ordinary and mundane while quietly removing the human from both ends of the media system.

The risk extends to the technical infrastructure of these systems. The report notes that retrieval systems can be compromised through Retrieval-Augmented Generation (RAG) attacks and the manipulation of agent memories.

Declining Human Engagement with Original Sources

Data indicates a migration of trust from original publishers to AI summarizers. A Pew Research Center report found that users clicked a source cited inside an AI summary only 1% of the time, based on 68,879 google searches conducted by 900 US adults.

Further data from SparkToro and Datos Group shows that 60% of US google searches ended without a click during the first four months of 2026. Consequently, the Reuters Institute expects search referrals to nearly halve over the next three years.

As AI models begin to cite other AI models, the original human reporting becomes increasingly distant. The report warns that the original source may eventually disappear entirely behind multiple layers of machine summarization.

The Redistribution of Editorial Power

Editorial judgment is centralizing inside machines, moving away from traditional newsrooms. This power is now distributed across the AI stack, including:

• Model developers who decide guardrails.
• Publishers licensing training data.
• Platform owners determining retrieval rankings.
• Governments building sovereign AI models.
• Open-source communities releasing foundation models.
• Enterprises curating the knowledge bases their AI agents consult.

The report concludes that the future information environment will be shaped by organizations that understand how machines learn, retrieve, reason, and remember.