AI Swarms Could Automate Disinformation at Machine Speed

The next wave of AI-driven disinformation won't just flood social media with synthetic content—it could autonomously manipulate both human decision-making and the data that trains future AI systems, according to researchers at the Geneva Centre for Security Policy.

While generative AI has already contributed to an estimated 50% of newly published articles and synthetic content appearing on up to 74% of new webpages, agentic AI represents a qualitative shift. Unlike static content generators, AI agents can perceive their environment and act autonomously to achieve goals with minimal human oversight.

The proliferation is rapid: enterprise use of AI agents jumped from 28.6 million in 2025 to a projected 2 billion by 2030, according to the GCSP analysis first reported by the World Economic Forum.

From generation to automation

The critical difference lies in coordination. When specialized AI agents operate in swarms, they can conduct adaptive, end-to-end influence operations that evolve in real time. A recent University of Southern California study demonstrated this potential in a simulated social media environment where 500 AI agents independently coordinated to promote a political candidate without human direction.

For businesses, the implications extend beyond brand reputation. Companies face exposure through supply chain disinformation, market manipulation via coordinated campaigns on stock-sensitive news, and deepfakes impersonating executives. These attacks can inflict reputational damage within days that takes years to repair.

The threat operates at two levels. Agentic AI can target human cognition through hyper-personalized disinformation and fabricated synthetic consensus at scale. But it can also compromise machine cognition by poisoning the datasets that train AI models.

Data poisoning at scale

Research shows that as few as 250 poisoned documents can embed hidden vulnerabilities into a large language model. As synthetic data proliferates and human-generated data grows scarce, models increasingly train on AI-generated content, risking what researchers call "model collapse"—self-reinforcing errors and biases that degrade model performance over generations.

Agent swarms could rapidly fill data voids with synthetic content, distorting both model outputs and public perception before credible information establishes itself. This creates a dual vulnerability: corrupted models produce flawed outputs that influence human decisions, while those decisions generate more corrupted data.

Why it matters

As enterprises deploy billions of autonomous AI agents for decision-making, the line between human and machine cognition blurs. Compromising either threatens the reliable information and undistorted judgment essential to business operations, public policy, and institutional trust. Unlike previous disinformation campaigns that required human coordination, agentic AI can operate at machine speed and scale with attribution challenges that make attacks easier to launch and harder to trace.

Building cognitive security

Defending against autonomous influence operations requires treating both human and machine cognition as critical infrastructure, according to the GCSP researchers. For organizations, this means implementing AI auditing and red teaming protocols, training models on verified datasets, and establishing verification processes before information influences decisions.

Broader defenses include cross-platform monitoring to predict narrative spread, defensive AI agents to detect coordination patterns, and progress in mechanistic interpretability—understanding how models reach conclusions. The researchers emphasize that building resilience demands not just technical solutions but institutional capacity, cross-sector research, and genuine multi-stakeholder governance.

The analysis was published by Dr. Jean-Marc Rickli, Head of Global and Emerging Risks at the Geneva Centre for Security Policy, and Tobias Knappe, Senior Project and Research Officer, with details first reported by the World Economic Forum.