AI in the Bundestag: Share of AI-Written Speeches Rises 58 Percent Within a Year
PR agency Twenty-Eight has a total of 4,471 plenary speeches from the current legislative period individually checked for the use of AI with the AI detection tool Pangram
Average AI share in Bundestag speeches rises from 3.8 percent (2025) to 6 percent (2026)
SPD members of parliament write the most AI-heavy speeches, government members and the Greens the most restrained
For 69 percent of members of parliament, no detectable AI share could be found in their speeches
While in 2025 around one in 13 longer speeches in the German Bundestag (Germany's federal parliament) showed linguistic traces of AI, in 2026 it is already around one in nine. That is the finding of a study by the PR agency Twenty-Eight, which analyzed 4,471 plenary speeches of 500 words or more from the 21st legislative period, from the constituent session on March 25, 2025 to July 9, 2026. Each speech was individually checked with the AI detection service Pangram. Pangram analyzes texts for linguistic patterns that could indicate the use of language models and assigns them to one of three categories: "Human," "Mixed" or "AI."
Clear rise in AI use within one year
For the analysis, an average AI score was calculated. To do so, the percentage share of AI-generated content within each speech was determined; content classified as "Mixed" was counted as 50 percent AI-generated. Across both years covered, this average AI score stands at 4.7 percent. Within one year, the figure rose from 3.8 percent in 2025 to 6 percent in 2026. That corresponds to an increase of around 58 percent.
In addition, the study evaluated how many speeches Pangram classifies as AI-influenced at all. Of the 4,471 plenary speeches examined, the tool rates 92 speeches (2.1 percent) as fully AI-generated. A further 314 speeches (7 percent) are classified as "Mixed" and thus as AI-assisted. Taken together, 406 speeches show linguistic features that, according to the analysis, point to the use of language models. That corresponds to around one in eleven speeches examined.
SPD communicates in the most AI-heavy way, government members and the Greens most restrained
There are considerable differences between the parliamentary groups: Speeches by SPD members show the highest average AI score at 7.69 percent, followed by the AfD (6 percent), CDU/CSU (4.4 percent) and Die Linke (4.2 percent). The figure is significantly lower for the Greens, whose speeches come in at an average AI score of 1.8 percent. In the analysis, speeches by members of the government, i.e. ministers from the CDU/CSU and SPD, were reported separately and not attributed to the respective parliamentary group figures, since not all of them are members of parliament. Notably, speeches by government members reach the lowest figure of all groups examined, with an AI score of 1.2, compared to those of parliamentary group members.
The year-on-year comparison also shows that low overall figures do not necessarily mean stagnation. It is precisely the groups with the lowest AI shares that record the strongest relative increases between 2025 and 2026, with the AI score for government speeches rising from 0.5 to 2.5 percent and for the Greens from 1.3 to 2.5 percent. Both groups thus remain at a low level in comparison, but show the strongest development over time.
Individual members of parliament stand out
Marked spikes also appear at the level of individual members of parliament. The highest AI scores in the entire evaluation are reached by three AfD members: Tobias Teich with 100 percent across two speeches checked, Christoph Grimm with 100 percent for one speech and Claudia Weiss, likewise with 100 percent for one speech. Given the small number of speeches checked, however, the significance of these individual figures is limited by comparison. Sascha Lensing of the AfD reaches an AI score of 76.2 percent across five speeches checked (fourth place). Christos Pantazis of the SPD reaches an average AI score of 66.6 percent across eleven speeches checked (fifth place).
For the majority of the members of parliament examined, however, Pangram found no indications of AI use. For 401 of the 582 members evaluated, or 69 percent, no detectable AI share could be found in the speeches checked.
Lukas von Zittwitz, founder and managing director of Twenty-Eight, comments:
"AI detectors do not prove authorship in the first place. They measure probabilities. And because language models were trained on human texts, there is no such thing as the one AI language. Patterns that look suspicious today can therefore also be found in speeches given long before ChatGPT. What makes an evaluation like this reliable is therefore the volume and the trend, with more than 4,400 speeches over two years examined using the same method.
That members of parliament and their teams use AI tools for support is not particularly surprising at first. Language models are now part of everyday life in many offices and can make daily work considerably easier. For parliamentarians in particular, hands-on experience with AI is important, since they also have to deal with the societal consequences, legal frameworks and possible limits of this technology. It is all the more remarkable how quickly AI-assisted writing is gaining ground in political communication as well. Within one year, the AI share in Bundestag speeches has risen by more than half, which suggests that LLMs are being used more and more in the preparation of parliamentary speeches.
For members of parliament and their teams, AI can be a helpful tool for structuring speeches, refining wording and making complex content easier to understand. What is decisive, however, is that aspects such as the political message, the stance or the personal signature remain recognizable. Political communication thrives not only on linguistic precision but also on people conveying a voice of their own. The more intensively AI is used in speech preparation, the more important a conscious approach to these tools becomes."
About the study
The study examined 4,471 plenary speeches, 2,598 from 2025 and 1,873 from 2026, of the 21st legislative period of the Bundestag (since the constituent session on March 25, 2025 and up to July 9, 2026) with at least 500 words. Each speech was individually checked via the AI detection service Pangram. Pangram classifies each text as "Human," "Mixed" (AI-assisted) or "AI" (fully AI). As a consistent metric, a weighted AI score was formed per speech, counting AI in full and "Mixed" at half (AI score = share AI + 0.5 × share AI-assisted, in percent). The average AI score was calculated per parliamentary group (SPD, CDU/CSU, AfD, Greens, Die Linke, plus government speeches separately) and per speaker (582 people), each additionally broken down by 2025 and 2026.