Geo-politics, fake journalists: the new fronts of AI influence

The election campaign is no longer played out only at the ballot box, in the media or on social networks: it is also being prepared in the answers of ChatGPT, Gemini and Perplexity. Candidates have understood that before convincing the voter, it may be useful to make themselves known to their assistant.

In France, according to RTL [1], Édouard Philippe’s team already devotes a full-time position to this new battle. Gabriel Attal’s team assigns one to two people to it. The objective: to improve the candidates’ visibility in AI answers, notably by adapting the content published on the Internet.

This practice has a name: GEO (Generative Engine Optimization), the equivalent of SEO applied to conversational assistants. It is no longer just a matter of appearing at the top of Google results, but of becoming a source that the AI will retain and reformulate.

Influence can, however, take far less above-board forms. On October 8, OpenAI [2] announced that it had disrupted two covert operations using ChatGPT. The first, Bogus Bylines, run from Iran, according to the lab, allegedly placed around a hundred articles in online media under seven fake journalist identities. The Washington Post [3] confirmed the infiltration of several publications.

The second, Dark Clark, attributed to operators located in Russia, relied on a purported Latin American research center, some of whose collaborators allegedly did not know who was really behind it. In both cases, AI was used to fabricate or consolidate borrowed credibility: that of a journalist, a media outlet or a research institute. The modus operandi nevertheless remains more conventional than the attempt to influence LLMs: “What is most striking about these operations is how closely they resemble the complex influence operations of the pre-AI era, with AI simplifying certain processes,” OpenAI explains. Note that while the lab did identify the geographic origin of these operations, it remained silent on any possible state direction behind them.

Three-cushion shots, with no guaranteed result

In any case, because of their complexity, such operations have the drawback of being easier to spot. Some have clearly understood this. Thus, on October 14, Anthropic [4] acknowledged a flaw in Claude’s election safeguards: while the assistant generally refuses to take part in an explicitly malicious influence operation, it can contribute to the same operation when it is broken down into seemingly harmless tasks, such as audience research, article writing, website creation or content distribution. An admission that suggests the company has already spotted some clever operators who, rather than trying to break through its AI’s safeguards, presented it with the pieces of the puzzle separately.

In Quebec, an audit by McGill’s Media Ecosystem Observatory, reported by Agence Science-Presse [5], illustrates another problem. A recent election website, created with the help of AI and containing several errors, appeared at the end of September in 63% of the ChatGPT answers studied for election recommendation questions. At this stage, no intentional manipulation has been established. The finding is nonetheless worrying: an unreliable source can gain considerable visibility in an assistant’s answers.

These examples illustrate the mechanisms detailed in our investigation of LLMs as a cognitive attack surface: influencing information upstream can be enough to steer what the user receives downstream. Still, being cited by an AI does not automatically mean convincing a voter. While the political effectiveness of these strategies remains hard to measure, their promoters are clearly ready to gamble.

Sources and references

[1] RTL, October 5, 2026
How AI will influence the presidential election

[2] OpenAI, October 8, 2026
Disrupting “front” operations using AI

[3] The Washington Post, October 9, 2026
A Florida local newspaper ran an op-ed. It was an Iranian AI fake

[4] Anthropic. Preparing for the 2026 midterm elections
https://www.anthropic.com/news/2026-us-midterms

[5] Agence Science-Presse, October 8, 2026
AI and elections: risks, but what impact?