How Soft Power Learns to Speak to AI

When the internet was still new, I once told my grandmother about people I had met online. She knew almost nothing about computers and listened to my explanation with some suspicion. Then she asked a question that made us laugh: “Are any of them actually alive?”

More than two decades later, her question sounds considerably less naive.

We have grown accustomed to an internet populated by bots, synthetic images and increasingly convincing artificial personas. Foreign influence operations now go a step further and obscure both ends of the conversation: the apparent speaker may be a person who does not exist, and the first audience for the message may no longer be human at all.

A recent investigation by the Foundation for Defense of Democracies (FDD) addresses the less familiar of the two problems. FDD identified a cluster of websites presenting themselves as independent watchdogs, news outlets and research organizations while consistently promoting narratives favorable to Qatar and hostile to its regional rivals. Direct Qatari sponsorship has not been established, but the sites focused heavily on allegations surrounding the UAE, Israel and critics of Islamist organizations, often presenting themselves as neutral European or international sources.

Their independence was largely an illusion. FDD traced five of the main sites — NGO Report, Brussels Watch, AML Network, Inform Europe and Evening Star — through a trail of technical and operational overlaps. All five domains were registered on the same day in February 2023, four within 16 seconds of one another. Files across all five contained the same name, “Anisur,” in their metadata, while further traces led investigators to Bangladesh-based web developer and digital marketer SM Asif Hassan, whose former company had advertised NGO Report as a client and who had previously advertised his work creating Evening Star.

The content followed an equally recognizable pattern. Brussels Watch repeatedly sought to recast the 2022 European Parliament corruption scandal known as Qatargate — not to be confused with Israel’s more recent scandal of the same name — as “Belgiumgate,” alleging that Belgium and the UAE had conspired to frame Qatar. NGO Report built blacklists of organizations it classified as pro-Israel, pro-Saudi or pro-Emirati and portrayed criticism of the Muslim Brotherhood and political Islam as part of a UAE-backed campaign. None of the sites disclosed a connection to the others; none identified an owner or named editorial staff, and as of August 2026 FDD could find no associated legal entities behind them.

By conventional measures, parts of the operation hardly looked successful. Brussels Watch attracted just over 600 visitors in July, according to Similarweb data cited by FDD; NGO Report drew slightly more than 10,000 and AML Network more than 15,000. Yet NGO Report was cited by Chatham House and the Royal United Services Institute. FDD also tested ChatGPT with a specialized question closely matching an AML Network article, and the chatbot cited AML Network several times in its answer. Grokipedia, the AI-generated Wikipedia alternative launched by xAI, had cited NGO Report, AML Network and Brussels Watch at least 45 times. And in November 2025, according to Similarweb data reviewed by FDD, all measured referral traffic to NGO Report and Brussels Watch came from ChatGPT.

Six hundred readers make for an unimpressive media outlet, but only if the readers who matter are human.

A traditional propaganda outlet needs an audience; a website aimed at an AI system needs only to be indexed, retrieved and treated as a sufficiently credible source. The person asking the eventual question may never visit it. Search engines traditionally presented users with sources and left them to choose what to open, whereas generative systems increasingly read across those sources and return an answer. A user asking whether an organization is linked to the Muslim Brotherhood may never encounter the obscure “watchdog” on which part of that answer rests, much less investigate who created it.

A commercial industry devoted to exploiting this mechanism already exists. Generative engine optimization, or GEO, does for AI answers what SEO did for Google, trying to make particular brands and sources appear in responses produced by ChatGPT, Gemini, Perplexity and other systems. In China, where GEO services are sold commercially, CCTV demonstrated the principle by inventing a fitness tracker called Apollo-9 and flooding the web with fabricated reviews and expert commentary. Within hours, AI systems were recommending a product that did not exist.

The Qatar-aligned network points to the disappearance of the human reader; an Iranian operation recently uncovered by Meta illustrates the opposite end of the conversation, where the speaker disappears.

In its latest Adversarial Threat Report, Meta described an influence network originating in Iran and targeting audiences primarily in Nigeria and South Africa, as well as Kenya, Ethiopia and Sudan. The company removed 23 Facebook accounts, 13 pages, a group and 11 Instagram accounts and traced the operation to actors in Iran. Those actors, however, were not presenting themselves as Iranians. They built false identities tailored to African audiences — researchers, nurses and women’s-rights advocates — using images of local people and joining genuine local activist groups. The network also used AI to produce polished graphics and relied on fake engagement to make its activity look more popular than it was.

According to Meta, the network promoted anti-Western and anti-imperialist narratives, supported pro-Hamas activism and opposed African countries normalizing relations with Israel. The agenda fits Iran’s broader interests on the continent, where Tehran competes with Israel for political and diplomatic space and has long sought common ground through the language of anti-colonialism and resistance to Western dominance.

Iran also delivers such messages openly. Its embassy in Ghana has developed a taste for particularly outrageous social-media content: Benjamin Netanyahu was depicted as Voldemort during his address to the UN General Assembly, while a recent post promising to show the world “without Zionism and imperialism” produced a smiling planet on which the continents had been moved around almost beyond geographical recognition. Ghana is not an exception: Iranian diplomatic accounts in South Africa, Zimbabwe, Indonesia and Tajikistan have ventured into similar territory, particularly where host governments seem disinclined to demand much diplomatic restraint. Such posts leave little ambiguity about who is speaking. A supposed Nigerian researcher or South African women’s-rights activist does not. The message remains foreign, while the messenger disappears.

The two operations are very different: one manufactured the appearance of independent sources, the other manufactured identities. Taken together with the emerging GEO industry, however, they point in the same direction. Synthetic speakers can produce political content for systems that ingest it and eventually return it to human users, with much of its original context lost along the way.

The old war of narratives was a contest over which story people would believe. The new one is also a contest over who appears to be telling the story, and over what machines learn to recognize as reality.