The AI Election Threat Is Here. Europe Is Not Ready.

In December 2025, Romanian authorities identified over 30,000 AI-generated social media posts designed to suppress voter turnout in the presidential runoff. The posts were traced to a network operating from servers in Moldova and Russia, using large language models to produce hyper-localised content in Romanian dialects. Many were indistinguishable from genuine citizen commentary.

Romania was not alone. German intelligence agencies documented at least four synthetic news websites — complete with AI-generated journalist bylines and fabricated source quotes — that published hundreds of articles targeting swing constituencies in North Rhine-Westphalia ahead of state elections.

The uncomfortable truth is that detection technology is losing the arms race against generation. Current AI-generated text detection tools achieve roughly 62% accuracy — barely better than a coin flip when adversarial techniques are applied. Deepfake video detection is more reliable for now, but the gap is narrowing with each model generation.

The EU's Digital Services Act requires platforms to label AI-generated content, but enforcement depends on platforms' own detection capabilities. And the most sophisticated operations do not use mainstream platforms at all — they operate through encrypted messaging apps, niche forums, and newly created websites designed to be indexed by search engines.

Experts broadly agree on three interventions. First, mandatory watermarking of AI-generated content at the model level — a requirement that the EU AI Act includes but has not yet enforced. Second, real-time threat sharing between national intelligence agencies and platforms, which currently operates on an ad-hoc basis. Third, media literacy programmes that go beyond school curricula to reach the demographics most vulnerable to manipulation: elderly voters and first-generation internet users.