An Analysis of Socialbots Activity and Influence in Modern Japanese Social Media

This study analyzes socialbot activity in modern Japanese social media, revealing that their influence exceeds that observed during the 2016 U.S. presidential election and that human-retweeted socialbots are difficult to distinguish from genuine human accounts.

The Gist

Imagine the internet, specifically Twitter (now X), as a giant, bustling town square. In this square, people are shouting news, opinions, and rumors. Usually, you expect the voices to be real people. But recently, researchers discovered that a lot of the shouting is actually coming from robots (called "socialbots") that look and sound exactly like humans.

This paper is like a detective report written by two Japanese investigators, Shuhei Ippa and Masaki Hashimoto. They wanted to see how loud these robots are in Japan compared to the rest of the world, and whether real people are accidentally listening to them.

Here is the breakdown of their investigation, using simple analogies:

1. The Setup: The "Fake News" Factory

In the past, if you wanted to spread a rumor, you had to write a letter or call a newspaper. Today, anyone (or anything) can shout a rumor into the digital town square instantly.

• The Problem: Sometimes, bad actors use robots to shout lies (disinformation) about politics, business, or conspiracy theories to make people panic or get angry.
• The Old Way of Catching Robots: Years ago, robots were easy to spot. They were like clumsy dancers who spun in circles and shouted the same thing 1,000 times a minute. You could tell they weren't human because they were too fast and repetitive.
• The New Reality: These days, robots are like method actors. They don't shout 1,000 times; they whisper once a day. They have profile pictures, bios, and they wait for the right moment to speak. They are much harder to catch.

2. The Investigation: Three Big Cases

The researchers picked three specific "events" in Japan where rumors were flying around to see how many robots were involved. Think of these as three different festivals in the town square:

• Case A: The State Funeral of a Former Prime Minister.
  • The Vibe: A very emotional, divisive event.
  • The Find: They found that 17.4% of the people shouting were robots. Even more surprisingly, 30% of all the messages shouted came from robots.
• Case B: The Unified Local Elections.
  • The Vibe: A massive election for local mayors and council members.
  • The Find: The robot presence was even higher here. 23.8% of the accounts were robots, and they were responsible for 37.4% of the tweets.
• Case C: The "PASCO Cricket" Scandal.
  • The Vibe: A conspiracy theory that a bread company was raising crickets for government subsidies (which was proven false).
  • The Find: This was the robot hotspot. 28.9% of the accounts were robots, and they generated 37.1% of the noise.

3. The Big Surprise: Japan vs. The World

The researchers compared their findings to the famous 2016 US Presidential Election (a famous case where robots were used to influence voters).

• The Result: The robots in Japan were louder and more active than in the 2016 US election. In the US, robots made up about 15% of the crowd; in Japan's recent events, they were often nearly 30%.

4. The "Human" Mistake: Why We Can't Tell the Difference

Here is the most critical part of the study. The researchers asked: "Do real humans know they are retweeting (re-sharing) a robot?"

They looked at the profiles of the robots that humans shared. They checked for "tells" like:

• Do they have a profile picture?
• Do they have a bio?
• How long have they been on Twitter?

The Shocking Discovery: The robots looked exactly like real humans.

• They had profile pictures.
• They had bios.
• They had been around for years.
• They didn't spam; they posted slowly.

It's like a wolf in sheep's clothing. The robots were so well-dressed that when a real human saw a tweet from a robot, they thought, "Oh, that's a nice person sharing news!" and they shared it too. The study found that humans were retweeting robots just as often as they retweeted other humans.

5. The "Middlemen" (Influencers)

In the "Cricket" case, the researchers noticed something interesting. The robots weren't just shouting into the void; they were often being amplified by "middle media" (web magazines or smaller news sites).

• The Analogy: Imagine a robot whispers a rumor to a popular local radio station (the middle media). The radio station repeats it, and suddenly, thousands of real people hear it and think it's true. The robot didn't need to shout; it just needed to get the radio station to speak for it.

The Bottom Line

This paper tells us three main things:

1. Robots are everywhere in Japan: They are more active in Japanese social media than they were in the famous US election cases.
2. They are masters of disguise: They don't look like spammy bots anymore. They look like normal people.
3. We are their best friends: Real humans are unknowingly helping these robots spread their messages because we can't tell the difference between a robot and a real person.

The Takeaway: The town square is noisier than we thought, and the loudest voices might not be who we think they are. We need to be more careful about what we share, because the "actors" in the crowd are getting very good at their roles.

Technical Summary

Here is a detailed technical summary of the preprint paper "An Analysis of Socialbots Activities and Influence in Modern Japanese Social Media."

1. Problem Statement

The proliferation of disinformation on social media is a critical global issue, with politics, conspiracy theories, and business-related topics being primary vectors for rapid and deep spread. A significant driver of this phenomenon is the use of socialbots (automated accounts). While extensive research exists regarding socialbot activity in English-speaking contexts (e.g., the 2016 U.S. Presidential Election), there is a scarcity of data regarding their activity and influence in Japan.

Previous studies on Japan (e.g., Mintalra et al., 2017) suggested lower socialbot prevalence (~2%) compared to Western democracies (9–15%). However, these studies often lacked the use of advanced, universal detection tools like Botometer, potentially leading to underestimation. This study aims to clarify the actual scale, activity patterns, and influence of socialbots in contemporary Japanese social media, specifically investigating whether they are more active than in typical English-speaking cases and how they interact with human users.

2. Methodology

The study employed a quantitative approach using Botometer v4 to analyze Twitter data across three distinct Japanese case studies representing different domains of disinformation:

• Case 1: State Funeral of Former Prime Minister Abe (July–Sept 2022)
  • Context: A highly polarized political event.
  • Data: 980,894 accounts, 21.3 million tweets.
• Case 2: Unified Local Elections (March–April 2023)
  • Context: A major nationwide political event.
  • Data: 143,440 accounts, 440,601 tweets.
• Case 3: PASCO Cricket Disinformation (Feb–March 2023)
  • Context: A business/conspiracy theory case (false claim that a bakery glorified crickets for subsidies), verified by the Japan Fact Check Center.
  • Data: 49,467 accounts, 98,297 tweets.

Analysis Techniques:

• Bot Detection: The Botometer v4 tool was used to assign a "bot score" (0 to 1) to each account based on over 1,200 features. A threshold of 0.5 was applied to classify an account as a socialbot, consistent with prior literature.
• Comparative Metrics: The study compared the proportion of bot accounts and bot-generated tweets against the 2016 U.S. Presidential Election benchmarks.
• Human-Bot Interaction Analysis: To understand how humans interact with bots, the authors compared 10 specific profile attributes between socialbots that were retweeted and human accounts that were retweeted. These attributes included:
  • Daily tweet volume
  • Profile/Background images
  • Verification status (Certification)
  • Self-introduction, Residence, URL info
  • Account age (Duration of activity)
  • Following/Follower counts

3. Key Results

A. Prevalence of Socialbots

The study found that socialbot activity in Japan is significantly higher than previously estimated and exceeds the benchmarks of the 2016 U.S. Presidential Election:

• State Funeral: 17.4% of accounts were bots; bots generated 30.1% of all tweets.
• Unified Local Elections: 23.8% of accounts were bots; bots generated 37.4% of all tweets.
• PASCO Case: 28.9% of accounts were bots; bots generated 37.1% of all tweets.
• Comparison: In all three cases, the percentage of bot accounts and bot-generated tweets exceeded the ~15% account and ~20% tweet benchmarks often cited for the 2016 U.S. election.

B. Human-Bot Interaction (Retweeting)

• Indistinguishability: Humans frequently retweeted socialbot content. The analysis of retweeted accounts revealed that socialbots do not differ significantly from human accounts in terms of profile completeness.
  • Both groups had high rates of profile pictures (90%+ for bots, 96%+ for humans) and self-introductions.
  • Both groups had similar distributions of account age and follower counts.
• Low Activity Volume: Contrary to early socialbot models that posted massive volumes of content, the socialbots in this study posted less than one tweet per day on average. They rely on human amplification rather than sheer volume.

C. Specific Anomalies

• Unified Local Elections: Unusually, 10.6% of the detected "bots" in this category were verified (authenticated) accounts, compared to only 2.5% of human accounts. These were identified as false positives (official government, politician, or media accounts) that Botometer flagged due to high automation-like behavior or specific feature sets.
• PASCO Case: Two accounts with over 100,000 followers were identified as bots. These were web media outlets, suggesting a role for "middle media" in amplifying disinformation.

4. Key Contributions

1. Quantitative Baseline for Japan: Establishes that socialbot prevalence in Japan (17–29%) is higher than in the 2016 U.S. election, challenging the assumption that Japan is less affected by bot-driven disinformation.
2. Behavioral Shift Identification: Demonstrates that modern Japanese socialbots have evolved from "spamming" accounts to low-volume, high-simulation accounts that mimic human profiles closely to evade detection and gain human trust.
3. Human Amplification Mechanism: Confirms that the primary vector for bot influence is not the bot's own posting, but human retweeting. Humans cannot distinguish between these sophisticated bots and real users based on profile metadata.
4. Tool Validation: Validates the use of Botometer v4 (Universal Score) for Japanese-language contexts, showing its effectiveness despite being trained on English-centric data.

5. Significance and Implications

• Policy and Countermeasures: The findings suggest that current countermeasures focusing solely on high-volume spam accounts are insufficient. Detection systems must evolve to identify low-volume, high-fidelity bots that blend into human networks.
• Disinformation Dynamics: The study highlights that disinformation in Japan spreads through a symbiotic relationship between sophisticated bots and unaware human users. The "middle media" (web media acting as bots or bot-like entities) plays a crucial role in bridging the gap between automated generation and human consumption.
• Future Research: The authors propose developing specific "influence indices" to better measure the impact of these accounts and call for similar analyses in other non-Western contexts to build a global understanding of socialbot evolution.

In conclusion, the paper argues that socialbots in Japan are a more potent force in information diffusion than previously recognized, operating through sophisticated mimicry that renders them nearly indistinguishable from human users, thereby facilitating the rapid spread of disinformation.