Bots, #StrongerIn, and #Brexit: Computational Propaganda during the UK-EU Referendum

Contributions

• Characterizes political bot deployment during a high-stakes electoral event (Brexit referendum)
• Demonstrates strategic vs. incidental bot use: some bots mechanically retweet to amplify, others repurposed from previous campaigns (e.g., pro-Palestine accounts)
• Quantifies message concentration: identifies that top 100 users generate 8% of traffic, top ~2,000 users (less than 1% of 313K accounts) generate 32% of all referendum-related tweets
• Operationalizes "heavy automation" as >50 tweets/day and shows differential automation by perspective (5.7% in neutral accounts vs. 33% in mixed hashtag accounts)
• Provides ground-truth analysis of coordination: accounts like @ivoteLeave, @ivotestay, and @Col_Connaughton use pure retweet algorithms with no original content generation

Method

• Data collection: 1.5M tweets June 5–12, 2016, using 25 pre-selected hashtags (10 pro-leave, 10 pro-remain, 5 neutral)
• Scope: 313,832 distinct Twitter accounts identified
• Bot detection: Defined heavy automation as accounts posting ≥50 tweets/day (≥350 tweets in sample period); also searched for explicit "bot" in account description
• Hashtag analysis: Counted hashtag occurrences (allowing multiple per tweet) across three ideological buckets
• Limitations: Twitter Streaming API returns max 1% sample; no geolocation data; sentimement analysis not performed

Results

Hashtag dominance:

• Leave hashtags: 993,176 occurrences (54%)
• Remain hashtags: 363,217 occurrences (20%)
• Neutral hashtags: 475,233 occurrences (26%)
• Most popular leave hashtag (#voteleave) used 3× more than most popular remain hashtag (#StrongerIn): 341,839 vs. 110,653

Message concentration:

• Top 100 accounts: ~121,000 tweets (8% of all referendum traffic)
• Top 2,000 accounts (0.6% of 313K): 32% of all referendum tweets
• Accounts tweeting 100+ times during week generated 32% of traffic despite <1% of population

Automation levels by perspective:

• Neutral hashtags only: 5.7% heavy automation
• Mixed hashtag usage: 14–19.6% heavy automation
• All hashtags combined: 14% heavy automation, 0.8% self-disclosed bots (n=1,531)

Known bots identified:

• @ivoteLeave, @ivotestay: pure mechanical retweeters, no original content
• @Col_Connaughton: previously pro-Palestine bot repurposed for Brexit
• @Rotenyahu: repurposed to retweet Col_Connaughton Brexit content

Notes

Strengths:

• High-quality ground-truth dataset from real high-stakes referendum
• Identifies specific named accounts and their coordination patterns
• Novel measure of "heavy automation" (posting velocity) as practical bot proxy
• Documents temporal coordination signatures (multiple accounts posting same content simultaneously)

Limitations:

• Sample limited to 1% Twitter stream; microblog universe during referendum was larger
• Doesn't measure influence or persuasion: high tweet volume ≠ impact on voter decisions; Brexit votes still went 52-48 despite bot-dominated pro-leave messaging
• Self-disclosure ("bot" in bio) catches <1%; actual bot count likely 2–3× higher based on behavioral criteria
• No content analysis: can't distinguish between propaganda bots, activist bots, and news-feed bots purely from posting velocity

Significance: Demonstrates that political bot use during elections is real, detectable, and differentially deployed by campaigns. The finding that <1% of accounts generate 30% of traffic is economically and strategically important: small coordinated groups can distort Twitter discourse without majority support, shifting perceived prevalence and momentum. Published as COMPROP research note (non-peer-reviewed) but cited ~800 times by 2026; foundational for bot research during 2016 election and beyond.