TL;DR
This paper presents the first large-scale characterization of Telegram bots, revealing their critical role as infrastructure for both legitimate services and illicit activities. The authors introduce a novel dataset and methodology to analyze bot functionality, usage, and community engagement, highlighting the urgent need for better bot moderation.
What Problem Does This Paper Address?
Telegram, a popular messaging app with over a billion users, has evolved into a powerful platform offering a wide range of services through programmable applications called "bots." While these bots facilitate everything from online shopping to group moderation, Telegram has also become a breeding ground for illicit activities like scams, data leaks, and non-consensual image sharing. Concerns are rising that bots are a key enabler for scaling these malicious operations. However, there's a significant knowledge gap regarding the true nature and extent of bot usage on Telegram, partly due to the lack of a centralized bot repository and prior research focusing primarily on messages and channels rather than bots themselves. This paper aims to fill this gap by providing a comprehensive, large-scale study of Telegram bots.
Methods
The researchers developed a multi-pronged approach to collect and analyze Telegram bots. Their data collection pipeline involved:
- Snowball Sampling: Starting with existing datasets (Pushshift and TG-Dataset), they iteratively discovered new channels and messages by extracting URLs from bot and channel mentions.
- Bot Discovery: Bots were identified through user links in messages (usernames ending in "bot") and by scraping third-party bot directories.
- Automated Bot Interaction: A novel system was built to automatically interact with bots by sending
/startand/helpcommands to extract their functionality, as bot descriptions are often insufficient. - Bot Classification: Bots were categorized into nine domains (e.g., Admin Tools, Finance, Underground) based on their descriptions, chat responses, and associated channels. This classification was initially done manually and then scaled using large language models (LLMs) like GPT-4o.
- Functionality Analysis: Key bot functionalities such as payment processing, referral systems, crowdsourcing, and AI usage were identified through keyword matching on command lists and bot descriptions.
- Community Analysis: Bot usage was analyzed by examining supported languages, bot duration (lifespan), reuse patterns, and channel network topology.
Datasets
The paper introduces a new, large-scale dataset of Telegram bots and associated data. Key components include:
- Largest General-Purpose Message Dataset: Approximately 809 million messages.
- First Bot Dataset: Over 32,000 unique bots identified.
- New Data Uncovered: Over 67,000 additional channels and 492 million messages not previously reported.
- Linkage Data: 9 million links between bots, channels, and users.
This dataset is made available to researchers under a CC BY-NC-ND 4.0 license, with access granted through a data usage policy to mitigate potential harms.
Key Insights and Ideas
- Bots as Critical Infrastructure: The study firmly establishes bots as a fundamental part of Telegram's ecosystem, serving as essential infrastructure for both legitimate services and illicit operations.
- The Dual Nature of Bots: While many bots offer benign services, a significant portion actively facilitates illegal activities, acting as payment gateways, referral systems, and even malicious AI endpoints.
- Telegram as a "Gray" Zone: Telegram appears to occupy a unique space between clearnet websites and darknet markets, offering a mobile-first, user-friendly platform that lowers the barrier for illicit activities.
- Language as a Clue: Bots described in English are often used in non-English speaking communities, underscoring the importance of analyzing message content to understand bot usage beyond descriptions.
- Finance Bots are Risky: Finance bots exhibit short lifespans, high reuse rates, and a disproportionately high number of scam warnings, suggesting a strong association with fraudulent activities.
- Underground and Utility Bots Leverage AI: The Underground and Utility domains show the most prominent use of AI, indicating its application in both illicit services and helpful tools.
Key Results
- Scale of the Dataset: The dataset comprises 105,970 channels and 809,481,087 messages, including 32,071 bots.
- Bot Domain Distribution: Finance bots are the most prevalent (34%), followed by Other (11%), Admin Tools (11%), Content & Media (10%), Utility (9%), Social & Gaming (9%), Ideology (6%), Underground (5%), and Shopping (5%).
- Malicious Bot Prevalence: Approximately 4% of bots are associated with fraud, and 5% with underground activities.
- Scam Warnings: Around 10% of Finance bots in the dataset had scam warnings, significantly higher than other domains.
- Functionality Prevalence: Payment processing is common, with 9.4% of bots having payment-related commands. Referral functionality is also widespread (10.4%).
- AI Usage: 607 bots were identified as providing AI-powered services.
- Community Interconnection: Ideology bots are found in highly interconnected communities, while Utility bots are more distributed across diverse channels.
Limitations and Future Work
- Dataset Representativeness: The dataset does not include private channels, potentially underestimating the prevalence of Underground domains. The bot sample might also be biased towards popular bots.
- Limited Interaction Depth: The study only sent basic commands (
/start,/help) and could not identify specific payment methods or AI models at scale due to the complexity of multi-stage interactions. - Future Work: The authors suggest that future research should consider bots as software infrastructure and explore more in-depth interactions, potentially using LLMs for longer conversations while adhering to ethical guidelines. They also propose that moderators and law enforcement focus on bots as a more effective intervention point for disabling monetization and expansion of illicit activities. Recommendations for moderation include analyzing command lists, focusing on specific domains and languages, clustering similar bots, and treating payment/referral bots with caution.
Summary updated