IO Factory: Simulating AI-Enabled Influence Campaigns at Scale

IO Factory is an AI-driven framework for simulating information and influence campaigns as fully integrated, traceable processes. The paper's starting observation is that the threat of digital manipulation has moved beyond persuasive text from single models to AI swarms — persistent groups of coordinated agents that keep stable identities, adapt to platform feedback, and disguise organized campaigns as ordinary social interaction. Because such campaigns cannot be identified from isolated messages, they must be studied across a continuous spectrum of planning, platform action, exposure, interpretation, measurement, and adaptation.

The framework models the campaign lifecycle as its core unit. A ten-phase reference lifecycle (reconnaissance, narrative design, infrastructure, content production, laundering, integration, amplification, absorption, adaptation, evaluation) controls which actors can act, what becomes visible, and what evidence is recorded. Three actor categories populate the simulation: civilians (the measured audience), IO operators (public-facing intervention accounts), and a non-public manager that assesses the run and issues bounded guidance.

The measurement pipeline is the paper's centerpiece. A post has no effect merely because it exists — it must become visible to a civilian, be recorded as an exposure, pass through an LLM judge (returning relevance, stance, confidence, persuasiveness), and survive a deterministic construct-update rule that factors in source trust, repetition decay, and update rate. Influence is measured as directional lift: movement of civilian construct values (e.g., trust in a source) toward the campaign target in an active run, minus movement in a matched baseline run with no IO operators.

Empirically, the authors run 13 matched replicates of 10,000 civilians (active runs add 1,000 IO operators) plus a 100,000-civilian scale test, using Gemma 4 31B on H200 nodes. All three primary endpoints move in the target direction with significant directional lift (0.132 for eating-insect support, 0.130 for trust in Russia, 0.336 sign-adjusted for declining public-institution trust). The paper is explicit that these are simulator-scale outcomes under declared assumptions, not estimates of real-world persuasion.