Is AI shifting from content creation to strategy in influence operations?

The most significant finding in Anthropic's March 2025 misuse detection report is not that AI was used for influence operations — that was expected. It is the kind of use: Claude was employed not just for content generation but as an orchestrator deciding what actions social media bot accounts should take based on politically motivated personas. The AI determined when to comment, when to like, when to re-share posts from authentic social media users.

This represents a qualitative shift on the machine agency spectrum. Prior influence operations used AI as a tool (generate content, human deploys it). This operation used AI as a decision-maker (AI assesses context, selects action, executes timing). The bottleneck moved from "what to say" to "when to act" — and AI was assigned the strategic layer.

The scale matters: the operation engaged with tens of thousands of authentic social media accounts across multiple countries and languages. This is not a proof-of-concept. It is a professional "influence-as-a-service" operation at production scale.

A second pattern from the report: generative AI raises the capability floor for less sophisticated actors. An individual with limited technical skills developed malware that would typically require more advanced expertise. The pattern generalizes beyond influence operations — AI compresses the skill gap between amateur and expert-level misuse, making previously complex operations accessible to actors who could not have executed them before.

Since Does machine agency exist on a spectrum rather than binary?, influence operations are pushing AI toward the cooperative end of the spectrum — not in the prosocial sense, but in the structural sense that AI is taking on planning and decision-making roles within adversarial systems. And since Does incremental AI replacement erode human influence over society?, the automation of influence operation strategy (not just content) removes a layer of human judgment that previously constrained the scale and sophistication of such operations.

A second, orthogonal threat: displacement through comprehensiveness. Influence operations as typically framed involve coordinated inauthentic behavior — orchestrated campaigns of fake accounts. But a distinct mechanism threatens the same social-proof economy without any coordination: AI-generated single posts displace real influencer content. A single AI-written post that looks comprehensive, confident, and on-topic crowds out the human commentary it resembles — not because it is part of a campaign, but because it is available at scale and optimized for the reader's attention. Since Why do different LLMs generate nearly identical outputs?, the displacement compounds across models: many accounts posting AI content produce a hivemind of similar takes without any coordination. False punditry and influence operations therefore represent convergent threats to the same social-proof function of social media — one through coordination, the other through comprehensiveness-without-reply. Defending against one does not defend against the other; they require different detection strategies because they have different structural signatures.

The detection side is equally notable: Anthropic's team used Clio and hierarchical summarization to analyze large volumes of conversation data for misuse patterns. The same capabilities that enable influence orchestration also enable influence detection — but the asymmetry favors the attacker, who only needs to succeed once at scale while the defender must detect patterns across millions of conversations.

Source: Social Media