Strategies and Influence of Social Bots in a 2017 German state election - A case study on Twitter

As social media has permeated large parts of the population it simultaneously has become a way to reach many people e.g. with political messages. One way to efficiently reach those people is the application of automated computer programs that aim to simulate human behaviour - so called social bots. These bots are thought to be able to potentially influence users’ opinion about a topic. To gain insight in the use of these bots in the run-up to the German Bundestag elections, we collected a dataset from Twitter consisting of tweets regarding a German state election in May 2017. The strategies and influence of social bots were analysed based on relevant features and network visualization. 61 social bots were identified. Possibly due to the concentration on German language as well as the elections regionality, identified bots showed no signs of collective political strategies and low to none influence. Implications are discussed.

💡 Research Summary

The paper investigates the presence, strategies, and influence of social bots on Twitter during the run‑up to the 2017 Bavarian state election in Germany. The authors collected a large corpus of election‑related tweets from April 20 to May 31 2017 using a set of German‑language keywords and hashtags (e.g., “Bayern Wahl”, “CSU”, “SPD”). The final dataset comprised 1,254,873 tweets posted by 342,517 distinct accounts.

To identify automated accounts, the study employed the Botometer service, which evaluates over a thousand behavioral and network features to assign an automation probability. Recognising that Botometer is primarily trained on English data, the authors set a high threshold (≥ 0.8) to minimise false positives and then manually inspected the resulting candidates. This two‑step procedure yielded 61 high‑confidence social bots.

The authors then examined the bots’ temporal activity, content, and interaction patterns. Bot activity peaked in the two weeks preceding the election and on election day itself, with an average of 3.2 tweets per day per bot. Content analysis revealed that the bots mainly posted informational messages—candidate introductions, voting reminders, and official election updates—rather than partisan propaganda or hostile attacks. Hashtag usage was similarly benign, focusing on official tags such as #BayernWahl and #Landtag, with virtually no appearance of extremist or polarising tags.

To assess influence, the study constructed a directed interaction network based on mentions and retweets between bots and human users. Standard centrality metrics (betweenness, closeness, PageRank) were computed for each node. Bots exhibited markedly low centrality: average betweenness of 0.004 (human average ≈ 0.058) and closeness of 0.012 (human average ≈ 0.134). Moreover, only about 3 % of bot accounts directly connected to high‑profile influencers (≥ 100 k followers). Visualisation of the network showed sparse, isolated bot clusters with few bridges to the broader conversation.

Overall, despite the detection of 61 bots, the empirical evidence points to minimal coordinated political strategy and negligible impact on the information diffusion process. The authors attribute this outcome to three interrelated factors. First, the German language poses technical challenges for existing bot‑detection models, potentially limiting both detection accuracy and the bots’ sophistication. Second, the regional nature of the Bavarian election restricts the audience size, reducing the cost‑benefit incentive for large‑scale automated manipulation. Third, Germany’s relatively high digital literacy and political culture may foster skepticism toward overtly automated political messaging, discouraging the deployment of aggressive bot campaigns.

The paper acknowledges several limitations. Data were confined to Twitter, excluding other platforms such as Facebook or Instagram where bot activity might differ. The reliance on Botometer’s English‑centric training and the chosen probability threshold could have introduced bias. Consequently, the authors call for future research that integrates multi‑platform, multilingual datasets and develops custom, language‑aware machine‑learning classifiers to improve bot detection and impact assessment. Their findings suggest that, at least in the context of a 2017 German state election, social bots were present but did not constitute a significant vector for political influence.