Source-linked AI summary
People are Strange when you're a Stranger: Impact and Influence of Bots on Social Networks
Luca Maria Aiello, Martina Deplano, Rossano Schifanella, Giancarlo Ruffo
TL;DR
The potential of bots to manipulate online social networks remains uncertain, despite their possible effects on users and communities. This paper tests an unprofiled bot in a social experiment, finding that simple social probing can build popularity that users may follow into connectivity choices. The bot’s activity also exposed polarization and prompted privacy concerns.
Problem
The potential of bots to manipulate online social networks is still widely unknown, including whether a bot without public information can become popular and influential.
Source: §1 · p. 1Method
Results
Simple social probing let the bot acquire high popularity, and users more aware of its presence were more inclined to follow its friendship suggestions; its activity also triggered polarization and privacy concerns.
Source: §6 · p. 7Takeaways & Limitations
The findings have implications for how security and privacy are conceived in social networks.
Source: §6 · p. 7The bot’s account was suspended and its recommendation messages canceled, so feedback from all recommended users could not be recorded.
Source: §4 · p. 5
Abstract
from arXiv · showhide
Bots are, for many Web and social media users, the source of many dangerous attacks or the carrier of unwanted messages, such as spam. Nevertheless, crawlers and software agents are a precious tool for analysts, and they are continuously executed to collect data or to test distributed applications. However, no one knows which is the real potential of a bot whose purpose is to control a community, to manipulate consensus, or to influence user behavior. It is commonly believed that the better an agent simulates human behavior in a social network, the more it can succeed to generate an impact in that community. We contribute to shed light on this issue through an online social experiment aimed to study to what extent a bot with no trust, no profile, and no aims to reproduce human behavior, can become popular and influential in a social media. Results show that a basic social probing activity can be used to acquire social relevance on the network and that the so-acquired popularity can be effectively leveraged to drive users in their social connectivity choices. We also register that our bot activity unveiled hidden social polarization patterns in the community and triggered an emotional response of individuals that brings to light subtle privacy hazards perceived by the user base.
1 Introduction
As online and offline life become increasingly entangled, social media can affect real-world events and behavior. The paper investigates whether a bot with no public information can become popular and influential, finding that simple automated activity can achieve both.
1 Introduction
- Online communication has affected political opinion formation and the spread of earthquake alerts, illustrating the Web’s direct effects on human lives.The cited examples include the 2011 Arab Spring, political opinion formation, and earthquake alarms [Sakaki, Okazaki, and Matsuo 2010].Source: §1 · p. 1
- The real potential of bots to manipulate online social networks remains widely unknown, motivating an experiment on trust, popularity, and influence.The authors note that social media has been used to boost political consensus through automatically generated fake tweets.Source: §1 · p. 1
- The study uses an online social experiment to test whether a bot with no public information can become popular and influential.The authors frame the question around a bot that does not aim to reproduce human behavior.Source: §1 · p. 1
- Simple automated activity can make an untrustworthy individual relevant and influential, while the bot’s social reactions reveal how automated activity may affect network structure.The authors also report that these reactions provide insights into possible automated attacks on the social substrate.Source: §1 · p. 1
Contributions and roadmap
The paper uses an interacting bot to explore trust, popularity, and influence, and reports contributions spanning popularity, recommendation, and network dynamics. Its experiment also allows the authors to examine event causality from an interacting agent’s perspective.
Contributions and roadmap
- The reported outcomes address popularity and influence, recommender effectiveness, explicit feedback on recommendations, and network changes including social polarization.The authors also describe implications for social recommenders or spammers and for recommendation techniques tested on a real user base.Source: Contributions and roadmap · p. 1
- Interacting with users lets the authors explore event causality from an agent’s perspective rather than infer it only from temporal sequences.The anonymized data collected during and immediately after the experiment are available upon request.Source: Contributions and roadmap · p. 2
- The paper describes the experiment’s social-media setting and bot, then reports popularity and influence results before discussing interaction side effects.Table 1 is identified as reporting basic structural quantities in the social and communication networks at the experiment’s beginning and end.Source: Contributions and roadmap · p. 2, Table 1 · p. 2
2 Dataset and experimental setup
The experiment takes place on aNobii, a book-focused social network with directed social ties and a separate message-based communication graph. A crawler’s profile visits unexpectedly prompted user reactions, leading to a two-part experiment on popularity and influence.
2 Dataset and experimental setup
- aNobii users share book libraries and interact through thematic groups, directed social ties, and public profile shoutboxes.The social ties are either friendship or neighborhood links, which the study treats together; messages form a separate directed, weighted communication network.Source: §2 · p. 2
- Over the 1.5-year experiment, the network roughly doubled in size while retaining high social-link reciprocation and relatively low average degree versus general-purpose social media.Table 1 reports structural statistics at the beginning and end; Figure 1 presents degree and message-weighted degree distributions.Source: §2 · p. 2, Figure 1 · p. 2
- The study began in September 2009 with periodic crawling to track social-link creation, using a breadth-first search every 15 days to reach all accessible nodes.An empty stub account was used to access some profile sections restricted to subscribed users.Source: §2 · p. 2
- Figure 2 groups users by received-message count and presents their average book counts and sent-message counts.The figure caption describes the quantities shown but does not state their outcome.Source: Figure 2 · p. 2
- After a site-setting change in July 2010, each logged-in visit left a trace in the visited profile’s private guestbook, exposing the crawler among the last 30 visitors.The crawler consequently left traces in reached profiles approximately twice a month.Source: §2 · p. 2
- Unexpected reactions to those visits prompted a two-part experiment: continued crawling measured user responses, then a modified bot messaged users asking them to act.The later phase’s results are discussed in the popularity and influence sections.Source: §2 · p. 2
3 Gain of popularity
Lajello gained substantial popularity through profile visits and simple automated activity, despite lacking trust or human-like behavior. Responses recurred with crawling, while the bot reached the tail of several network popularity distributions.
3 Gain of popularity
- 98th, 99th, 98th, and 96th percentiles were lajello’s positions across the four Pagerank and HITS distributions.The paper describes the scores as lying in the very tail of the distributions.Source: Figure 4 · p. 3, §3 · p. 4
- Message bursts followed crawling rounds, but incoming communication largely stopped during inactivity and resumed at the prior pace when crawling restarted.The cumulative number of distinct users writing messages grew roughly linearly over time.Source: §3 · p. 3
- 2,435 public messages from 1,263 users, more than 66,000 profile visits, and 211 incoming social connections made lajello one of aNobii’s most popular profiles.By December 2011, it ranked second for distinct shoutbox contributors and 28th for received messages, and was among the 0.3% of users with the highest in-degree.Source: §3 · pp. 3–4
- Continuous social probing—visiting other profiles—can gain popularity even when the bot does not emulate human behavior.The experiment isolated profile visits as a minimal social activity to test their effectiveness in gaining attention.Source: §3 · pp. 3–4
4 Influence
The experiment tested whether an untrusted bot could influence users’ decisions to add social contacts. Among users who made a connection after receiving a recommendation, 52% followed the bot’s suggestion, while the results also indicate that automated activity can make an untrusted individual influential.
4 Influence
- The recommender combined profile and social-graph features in a Rotation Forest classifier, which achieved 0.82 AUC on 10-fold cross-validation.It used a balanced set of 20,000 user pairs two hops apart; the selected features came from Chi Squared selection.Source: §4 · pp. 4–5
- The experiment sampled Italian followers and non-followers, assigned recommendation or random suggestions, and added reciprocal recommendations for a subset, yielding about 3,000 recommendations.Messages suggesting a contact were sent to users’ public shoutboxes.Source: §4 · p. 5
- 52% of the 361 users who created a social connection in the 36 hours after the recommendation followed the bot’s suggestion.The classifier-based recommendations were at least twice as effective as random recommendations; suggestions to followers were more effective than those to non-followers, and reciprocal suggestions were most likely to succeed.Source: §4 · p. 5, Figure 6 · p. 5
- Bidirectional suggestions were more likely to succeed, in part because they could prompt communication between the two users and agreement to form a tie.A non-negligible 11% of successful recommendations belonged to the random categories.Source: §4 · p. 5
- Although the experiment was limited in time and to a relatively small user group, the authors conclude that popularity gained through automated activity can make an untrustworthy individual influential.The bot account was suspended and recommendation messages were canceled, so feedback from all recommended users was not registered.Source: §4 · p. 5
- The findings suggest that unauthorized external robots could effectively steer user opinions and actions, prompting a call for greater care in social-media security design.Source: §4 · pp. 5–6
5 Emotional reactions and social polarization
The bot’s activity elicited varied emotional responses and coincided with visible polarization in groups and social connections. Network analysis found that group membership aligned with detected clusters, while users also formed new communication links and bot-like profiles.
5 Emotional reactions and social polarization
- Messages to the bot ranged from veneration to hatred, and many users expressed discomfort at an unknown user periodically visiting their public profiles.Some users described fear of control or a strong privacy violation, despite the bot viewing only information available to community members.Source: §5 · p. 6
- The authors identify more systematic sentiment analysis as future work because fine-grained analysis of users’ emotional responses was beyond this study’s scope.Source: §5 · p. 6
- Manual annotation found 59% of messages positive and 16% negative, while sentiment shifted over time and opinions polarized most visibly in the last 400 messages after the recommendations.The analysis used positive, neutral, and negative labels; finer-grained sentiment analysis was outside the study’s scope.Source: §5 · p. 6
- Seven groups formed around the bot, including opposing groups of 102 supporters and 72 critics, whose members tended to form distinct clusters in social and communication networks.The cluster analysis compared within-cluster and between-cluster links against randomized networks and used OSLOM to assess group alignment.Source: Table 3 · p. 6, §5 · p. 7
- OSLOM clusters correctly classified 80% of social-network nodes and 72% of communication-network nodes by group membership, indicating that faction structure was rooted in social connections but had been hidden until the bot’s perturbation.The authors note that detectable clusters could also provide edges along which a network might be disrupted.Source: §5 · p. 7
- Tracking messages containing “lajello” revealed new communication links and bot-like profiles that attracted messages from other users.The authors describe these clones as empty profiles that visited users in the same way as lajello.Source: §5 · p. 7, Figure 9 · p. 7
6 Conclusions
The experiment shows that simple social probing can give an automated agent substantial popularity, which can translate into influence over users’ friendship choices. The bot also triggered social and emotional reactions, including polarization and privacy concerns, with systematic sentiment analysis left for future work.
6 Conclusions
- Simple social probing gave an automated agent surprisingly high popularity, and users more aware of the untrustworthy bot were more inclined to follow its friendship suggestions.The study also validates modern link-recommendation techniques with explicit feedback from users.Source: §6 · p. 7
- The bot was widely mistaken for a human, prompted new groups and communication channels, and triggered privacy concerns, fear of control, and polarization around latent clusters.Its activity also originated emulous users.Source: §6 · p. 7
- The findings have implications for how security and privacy in social networks are conceived, while systematic sentiment analysis of user messages remains future work.The authors identify sociological and emotional aspects as important to investigate.Source: §6 · pp. 7–8