Social bots distort the 2016 U.S. Presidential election online discussion

Key claims

• Bot prevalence: ~400,000 accounts (15% of the active population) were classified as likely bots, responsible for 3.8 million tweets (19% of total conversation volume).
• Ineffective replies: Bots failed to engage humans meaningfully—they interacted with other bots more than with humans via replies, suggesting insufficient sophistication to generate engaging discussion.
• Effective rebroadcasting: Despite poor reply engagement, humans and bots retweeted each other at similar rates, indicating bots are effective at spreading information to the human population.
• Geographic spoofing: Bot-generated tweets exhibited different geographic patterns than human tweets, with unusual concentration in the Midwest and South, while human-generated discussion tracked populated states (California, Texas, Florida, New York).
• Sentiment skew: Trump-supporting bots produced overwhelmingly positive sentiment (~2/3 of non-negative tweets), generating "a stream of support at staggering odds with respect to the overall negative tone" of the campaign, potentially creating false impressions of organic grassroots enthusiasm.
• Network embeddedness: As nodes became more embedded in the retweet network (higher k-core), the proportion of bots increased fourfold from ~3% to 12%, suggesting bots occupy increasingly central positions in information diffusion.

Methodology

The study collected Twitter data via the Search API using 23 manually curated election-related hashtags and keywords (e.g., #donaldtrump, #hillaryclinton, #trumppence16, #imwithher) across three observation periods covering the three 2016 Presidential debates (September 16–October 21, 2016). Bot detection employed the BotOrNot machine-learning system, which extracted >1,000 features spanning account metadata, usage statistics, temporal activity, content, network structure, and sentiment, trained on thousands of known bot instances with >95% accuracy. For computational efficiency, analysis focused on the top 50,000 most-active users (2% of the population but generating 60% of all tweets). Sentiment analysis used SentiStrength, a tool optimized for social media text, to assign positive and negative polarity scores. Partisanship assignment employed a conservative hashtag-adoption heuristic: users whose top 10 hashtags were predominantly supporting one candidate were classified accordingly.

Results and findings

• Detection: Of the top 50,000 users, 7,183 (14.4%) were classified as likely bots (bot score >0.5), responsible for 2.3 million of 12.6 million tweets from this cohort. Extrapolating to the full population yields ~400,000 bots generating ~3.8 million tweets total.
• Temporal dynamics: Baseline periods showed steady bot contribution (~10% of hourly tweets). During debate-related spikes, human activity dominated the peaks (60,000–80,000 tweets/hour vs. 10,000 tweets/hour for bots), but bots maintained continuous elevated activity, explaining their large volume share.
• Interaction patterns: Humans engaged other humans at significantly higher rates than bots (one order of magnitude difference in replies). Conversely, bots predominantly interacted with other bots via replies, failing the fundamental goal of engaging the human audience.
• Retweet effectiveness: No significant difference between rates of human retweets of bot-generated vs. human-generated content, suggesting bots exploit rebroadcasting as their primary influence channel.
• Sentiment asymmetry: Trump-supporting bots generated almost exclusively positive tweets, with no appreciable negative sentiment. Clinton-supporting bots showed more balanced positive/negative distribution. The #NeverHillary hashtag (Trump-supporting) had significantly more negative tweets (204,418 negative vs. 171,877 positive), while #NeverTrump (Clinton-supporting) showed near parity (105,906 positive vs. 118,661 negative).
• Network centrality: k-core decomposition revealed that the fraction of bots in higher k-cores grows steadily: from 3.1% in 10-core to 12.3% in 100-core, indicating that the most-retweeted, most-influential accounts show a disproportionate concentration of bots.

Significance and limitations

This study was among the first empirical demonstrations of large-scale bot activity during a U.S. presidential election, quantifying the scope of automated manipulation and its specific effects on information diffusion. The findings show that bots do not create political discussion in isolation—they operate alongside humans—but distort the observed support, sentiment, and geographic distribution of political expression.

Limitations include: (1) analysis limited to one month before election day, missing pre-primary dynamics and post-election effects; (2) bot detection via BotOrNot is probabilistic (bot scores can be ambiguous near the 0.5 threshold); (3) lack of knowledge about bot operators (state, organization, individual—unknowable from detected behavior); (4) small annotated dataset for partisanship assignment (7,112 Clinton supporters, 17,202 Trump supporters), limiting scope of sentiment analysis; (5) inability to determine causal effect on voter behavior or election outcome.

Notes

This paper was seminal in raising academic and public awareness of bot-driven manipulation in U.S. elections. The BotOrNot system used here became widely adopted; a later paper (Davis et al. 2016) describes it in detail. The study's limitation to the top 50,000 accounts by activity is a practical necessity given Twitter API constraints, but it means the inferences about the full population rest on extrapolation. The finding that bots fail to engage humans via replies is noteworthy: it suggests that sophisticated bot farms are designed more for information amplification (retweets, sentiment manipulation) than for social engineering (replies, convincing discourse). The geographic clustering in the Midwest and South warrants follow-up investigation—it could reflect either bot farm geography or deliberate targeting strategies.