How Synthetic Media Took Over a Major German Election
UncovAI collected 33,964 tweets posted around the 6 September 2026 Saxony-Anhalt state election and scored every attached image. Synthetic pictures beat real photographs on all eight days measured. The widest-reaching content in the entire corpus was AI-generated.
The study in four numbers
The AfD won almost 44% of the vote. The information space around that result is what we measured, using a hand-coded stance sample, a keyword classifier and AI image detection across the full corpus.
The highest-reach post came from a refugee who does not exist
An account created in February 2026 posed as a Syrian refugee promising to leave Germany if the AfD won. Two tweets, one carrying an AI-generated photo, reached over 7 million views combined — more than any politician, outlet or activist in the dataset. Copycat accounts running the same format pulled hundreds of thousands more.
The account discloses "AI-Satire" in its bio. Not in the tweets. A bio label does not survive a screenshot, and screenshots are how this content actually travels.
The day-by-day image counts, detection heatmaps and all eight narratives are in the report.
Download the Full Report →The disinformation was domestic, not foreign
Election-fraud narratives drove it. A false story about a Dessau nursing home ran to 283 tweets and 851,000 views, and kept circulating after prosecutors closed the case without evidence. A fabricated poll beat every debunk of it by four to one.
Meanwhile the operation attributed to the Russia-linked Storm-1516 network reached roughly 1,600 views. Links to Russian state media appeared 18 times in the whole corpus. The inauthentic networks were easy to spot and reached almost nobody; verified domestic accounts did the damage.
Want the rest?
The full report covers the day-by-day AI-versus-real image counts with detection heatmaps, all eight disinformation narratives with view counts and peak days, the 14:1 engagement amplification gap, and the pro-Russian and anti-EU narrative breakdown.
How were images classified as AI-generated?
Every image attached to a tweet was scored by UncovAI's detection models, which return a confidence score and a heatmap showing which regions drove the decision.
Does AI-generated mean disinformation?
No. Detection tells you what a file is, not what it was for. Much of this imagery is satire or illustration. The risk is that at 3.5:1, synthetic imagery stops being a signal and becomes background noise.
Is the dataset public?
Yes — 33,964 tweets, published on Hugging Face as UncovAI/2026_Saxony-Anhalt_Election_X, with the full methodology described in the report.
Get the full report
Every figure, chart and narrative breakdown from the Saxony-Anhalt analysis. Free.
Download the Full Report →