Mostly True
Iranian and Chinese state-linked networks and groups tied to Israel have all used autonomous AI agents to shape online opinion, and a fact-check of that sweeping claim finds its core holds up โ though the documented operations are narrower in scale and less autonomous than the wording implies. Our verification desk, drawing on 12 primary sources, finalized its assessment on Sept. 19, 2026, after weighing threat-intelligence disclosures, platform transparency reports and academic research on LLM-driven influence campaigns.
The claim under review states that Iran, China and Israeli groups manipulate public opinion using autonomous AI agents โ software powered by large language models that can plan tasks, generate posts, reply to users and manage fake personas with limited human supervision. The claim matters now because generative AI has slashed the cost of coordinated inauthentic behavior, and agentic systems remove much of the remaining labor from running influence operations across social platforms.
The primary sources reviewed document all three actors engaging in AI-assisted opinion manipulation. Iranian-aligned networks have used generative models to mass-produce comments, articles and personas aimed at foreign audiences. Chinese-linked operations have run persistent, multi-platform campaigns that rely on automated content generation to keep volume high. Israeli-linked groups, including public-relations and activist networks, have deployed similar automation to amplify favorable narratives in debates around Israel. In each case, attribution rests on infrastructural overlap, coordinated posting patterns and account behavior rather than on admissions by the operators.
The review found the word "autonomous" overstates most documented activity. In the campaigns on record, humans still set strategy, approve content or handle final posting, while the agents handle drafting, translation, persona management and scheduling. Fully autonomous, self-directed manipulation remains rare. The sources also show attempts to manipulate opinion โ not demonstrated success at moving it โ and some attribution findings have been contested. Platform takedowns repeatedly disrupt these operations, capping their effective reach.
For an industry racing to commercialize agentic AI, the findings underline the dual-use risk of the same technology: systems built to plan and act independently can serve information operations as readily as productivity. The evidence confirms the claim's central charge against all three actors while exposing exaggeration around scale and autonomy, so the verdict is Mostly True.