- [ ] The profile is incomplete or generic
- [ ] The username may look non-personal and sometimes include random numbers | - [ ] The account has typically been active for longer and has a post history
- [ ] The profile is complete and seems personal; it may present strong ideological or political self-description
- [ ] The username looks personal | - [ ] The profile looks plausible and personal, often with a profile picture and biographical details (sometimes stolen, AI-generated, or copied)
- [ ] Account history may be moderate and designed to look authentic over time |\n| **Posting Behaviour** | - [ ] The activity does not match normal human online behaviour
- [ ] The accounts post or repost content very frequently
- [ ] The accounts post or repost content at all hours, day and night | - [ ] The activity resembles normal human online behaviour
- [ ] The account posts or replies at irregular times
- [ ] The account becomes more active during controversial discussions | - [ ] Activity patterns resemble normal human use
- [ ] Multiple accounts run by the same operator may show similar active hours or rhythms
- [ ] Sockpuppets tend to start fewer discussions and write shorter posts than typical users |\n| **Interactions** | - [ ] The account does not have real conversations
- [ ] The accounts mostly like, share, or repost
- [ ] Replies are short and automated | - [ ] The account replies directly to other users
- [ ] The account engages in debates with the purpose of provoking reactions
- [ ] Conversations are extended to create or escalate conflict | - [ ] The account engages in real conversations, often supporting the operator's main account or other sockpuppets
- [ ] Replies are contextually appropriate and seem authentic
- [ ] Pairs of sockpuppets often interact in the same discussion at similar times |\n| **Content Features** | - [ ] The content is one-sided and repetitive
- [ ] The same narratives are posted many times | - [ ] The content is specifically tailored to harm or provoke a target
- [ ] The content targets individuals or social groups | - [ ] Content seems genuine and varied across accounts
- [ ] The underlying message or stance aligns suspiciously across the network
- [ ] More frequent use of personal pronouns such as \"I\" |\n| **Language** | - [ ] Generic expressions, repetitive phrasing with keywords | - [ ] Varied, emotional, often abusive or offensive language | - [ ] Natural and varied language
- [ ] Multiple accounts may share linguistic fingerprints (similar phrasing, vocabulary, punctuation, or error patterns) |\n| **Network & Technical Indicators** | - [ ] Social bots follow other social bots, but the relationship is typically one-way and not reciprocal
- [ ] Coordinated behaviour is observable across multiple bot accounts | - [ ] Trolls follow human accounts
- [ ] Connections are often reciprocal (they follow their followers and vice versa)
- [ ] Trolls typically act independently of each other | - [ ] Multiple accounts engaging with each other in mutually supportive ways
- [ ] Connections may be artificially reciprocal between sockpuppets in the same network, or deliberately absent to avoid detection
- [ ] Same IP address, device fingerprint, or login pattern \\\\\\*(platform-side detection)\\\\\\*
- [ ] More clustered ego-networks than ordinary users
- [ ] Correlated activity timing across accounts |\n\n\n\n:::info\n* Ferrara, E. (2023). Social bot detection in the age of ChatGPT: Challenges and opportunities. *First Monday, 28*(6). [https://doi.org/10.5210/fm.v28i6.13185 ](https://doi.org/10.5210/fm.v28i6.13185)![](data:image/png;base64,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 \"Add to Citavi project by DOI =16x16\")\n* Kumar, S., Cheng, J., Leskovec, J., & Subrahmanian, V. S. (2017). An army of me: Sockpuppets in online discussion communities. *Proceedings of the 26th International Conference on World Wide Web (WWW '17)*, 857–866. [https://doi.org/10.1145/3038912.3052677 ](https://doi.org/10.1145/3038912.3052677)![](data:image/png;base64,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 \"Add to Citavi project by DOI =16x16\")\n* Orabi, M., Mouheb, D., Al Aghbari, Z., & Kamel, I. (2020). Detection of bots in social media: A systematic review. *Information Processing & Management, 57*(4), 102250. [https://doi.org/10.1016/j.ipm.2020.102250 ](https://doi.org/10.1016/j.ipm.2020.102250)![](data:image/png;base64,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 \"Add to Citavi project by DOI =16x16\")\n* Solorio, T., Hasan, R., & Mizan, M. (2013). A case study of sockpuppet detection in Wikipedia. *Proceedings of the Workshop on Language Analysis in Social Media (LASM) at NAACL-HLT*, 59–68. Association for Computational Linguistics.
Information Visibility & Prominence
\nInformation Visibility refers to whether — and how prominently — specific information items, topics, sources, or accounts appear to users in digital environments. Information that exists in a platform's index or network does not automatically reach all users, nor reach them in the same way.
\nVisibility is the outcome of two interacting forces:
\n- \n
- how information is circulated (sharing, redistribution, spread) and\n \n
- how it is surfaced prominently, pushed down, or filtered out.\n \n
Distinguishing reliability and visibility is essential for information literacy.
\n- \n
- Reliability depends on who created an item (author) and how it was edited (editorial review) — not on whoever shared it, on the environment it appeared in, or on how prominently it was surfaced.\n \n
- Visibility depends on who shares an item, on the mechanisms that surface it, and on the practices through which it is amplified. Visibility is not a measure of reliability.\n
For example, a retweeted article: the account that shared it and the platform that surfaced it influenced its visibility — but the article has its own author and editorial history, whose reliability must be assessed separately.
\n \n
Information Circulation: Sharing, Forwarding, Linking, Reposting, Quoting
\nPeople, accounts (including bots), and organisations circulate existing information items by sharing, forwarding, linking, reposting, quoting, or otherwise redistributing them.
\n→ When an item is passed on with added commentary, interpretation, or reframing, a new information item is created with a new author.
\nSource-Driven Information Promotion & Visibility
\nPractices initiated by sources — those who create, publish, or promote information items. Sources include individual users, content creators, organisations, advertisers, and website operators. Unlike platform-driven mechanisms, these practices are driven by the sources themselves, who decide how to make their content visible. They operate in two modes: by adapting content to the platforms' ranking systems (Reach, Direct Addressing, SEO, SMO, Platform-specific Optimisation), or by paying directly for placement (Paid Placements).
\nAccount Reach
\nThe size and structure of a sharer's potential audience.
\nSeveral factors determine reach:
\n- \n
- #### follower or subscriber count\n \n
- #### verification status\n \n
- #### account standing: age, engagement history, platform reputation\n \n
Reach affects visibility in two ways:
\n- \n
- directly: items shared by high-reach accounts appear in more feeds at the moment of sharing\n \n
- indirectly: high-reach accounts generate more engagement signals, which platform algorithms then use to elevate items in ranking (→ Algorithmic Gatekeeping)\n \n
→ Reach varies widely: a private account with 100 followers and a public account with one million followers operate at fundamentally different scales of influence on visibility.
\nDirect Addressing
\nMechanisms by which sharers target specific recipients, producing immediate prominence for those users through platform notifications.
\nExamples:
\n- \n
- #### @mentions\n \n
- #### tagging (in photos, posts, or threads)\n \n
- #### quote-tweets & reply-mentions\n \n
- #### group direct messages\n \n
- #### mailing-list addressing (To, CC)\n \n
→ Direct addressing differs from regular sharing: the targeted user receives the item directly via notification, regardless of whether they would otherwise have encountered it through their feed.
\n→ Direct addressing is a hybrid mechanism. It operates user-side, but functions through platform infrastructure (notification systems). Its effect is immediate prominence for the addressed user.
\nSearch Engine Optimisation (SEO)
\nSource-side practices to adapt websites — their content, metadata, and link structure — so that they rank more prominently in general-purpose search engine results (Google, Bing, DuckDuckGo, etc.).
\nCommon practices include:
\n- \n
- keyword research and integration in titles, headings, body text\n \n
- metadata optimisation (title tags, meta descriptions, alt text)\n \n
- link building (acquiring inbound links from authoritative sites)\n \n
- site structure and internal linking\n \n
- page speed and mobile-friendliness\n \n
- producing content that matches search intent\n \n
→ SEO is the most formalised optimisation discipline because search engine ranking signals are relatively stable and well-documented (Lewandowski et al.). Specialised SEO professionals, agencies, and tools support its practice.
\n→ SEO does not change how search engines rank pages — it adapts the website to fit existing ranking criteria. Publishers can shape what the algorithm sees, not how it decides.
\nSocial Media Optimisation (SMO)
\nSource-side practices to maximise visibility, engagement, and shareability of content on social media platforms.
\nCommon practices include:
\n- \n
- hashtag strategies (trending or topic-specific tags)\n \n
- posting timing (when target audiences are active)\n \n
- content format choices (short video, carousels, reels)\n \n
- headline and hook design (catching attention quickly)\n \n
- encouraging engagement (questions, polls, calls to action)\n \n
- cross-platform repurposing of content\n \n
→ SMO is less formalised than SEO because social media ranking signals are more opaque and platform-specific. Practices shift as algorithms change.
\nPlatform-Specific Optimisation
\nOptimisation strategies tailored to the conventions and ranking logics of individual platforms — beyond general SEO or SMO principles.
\n- \n
- TikTok: hooking viewers in the first three seconds, using trending sounds, vertical short-form video\n \n
- Instagram: high-quality visuals, Reels-first strategy, hashtag mixing\n \n
- YouTube: thumbnail design, watch-time optimisation, keyword-rich titles and descriptions\n \n
- LinkedIn: long-form professional posts, native publishing, networked engagement\n \n
- X (Twitter): concise hooks, threads, replying to high-reach accounts\n \n
→ Platform-specific optimisation requires understanding each platform's ranking system, audience behaviour, and content format preferences. What works on TikTok rarely works on LinkedIn.
\nPaid Placements & Advertising
\nSource-side practice of paying for visibility — sponsored content placed alongside organic content, typically through advertising.
\nCommon forms:
\n- \n
- sponsored search results (search engine ads)\n \n
- sponsored posts and promoted content (social media)\n \n
- display ads (banners, videos)\n \n
- influencer partnerships (paid collaborations)\n \n
→ Paid placements bypass organic ranking systems: instead of optimising content to rank well, the source pays the platform directly for placement.
\n→ They are sometimes clearly labelled ("Sponsored", "Ad"), sometimes only weakly distinguishable from organic results. Labelling standards vary by jurisdiction and platform.
\nPlatform-Side Information Promotion & Gatekeeping
\nWhat users actually see is rarely the product of a single mechanism. In a search engine, an algorithmically ordered list of organic results is presented alongside paid placements, AI-generated summaries, and sometimes editorial highlights — and the underlying ranking signals can be deliberately influenced through Search Engine Optimisation. In a social media feed, algorithmically ranked posts appear next to sponsored content, recommended accounts, and trending overlays. Each component follows its own logic and contributes to a composite visibility outcome.
Editorial Curation
\nPlatform-side manual curation: information items deliberately featured by editorial teams or platform operators rather than surfaced through algorithmic ranking.
\nExamples:
\n- \n
- featured Snippets in search results\n \n
- curated trending sections\n \n
- editor-selected stories in news aggregators\n \n
- platform-promoted hashtags\n \n
- Editor's Picks in app stores\n \n
- featured creators or accounts\n \n
→ Editorial highlights sit alongside the algorithmic mechanisms and reflect the platform's own judgements about which content deserves prominent display.
\n→ Unlike algorithmic gatekeeping (curation and personalisation), editorial gatekeeping involve human editorial choices by the platform itself. Functionally, this is a form of Editorial Gatekeeping ) — performed by the platform rather than by traditional publishers.
\nAlgorithmic Gatekeeping
\nAlgorithmic Gatekeeping refers to the role of algorithms in deciding which information items reach which users — the digital counterpart to Editorial Gatekeeping (→ Information, Sources & Information Environments → Editorial Review). It involves both selection (what is surfaced and ranked highly) and exclusion (what is filtered, demoted, or hidden).
\nAlgorithmic gatekeeping operates across different platform types:
\n- \n
- in search engines, algorithms select and order results based on queries\n \n
- in social media feeds, algorithms decide which posts appear more prominently\n \n
- in video platforms, algorithms suggest what to watch next\n \n
- in AI-based answer systems, algorithms generate, summarise, or synthesise responses\n \n
Algorithmic gatekeeping operates in two modes that often work together: general operations applied across all users (Algorithmic Curation), and individual tailoring based on tracked user signals (Algorithmic Personalisation).
\nAlgorithmic Curation
\nGeneral algorithmic operations applied across users — they shape what information is available on the platform, regardless of who the user is.
\n- \n
- Crawling and Indexing — Which information items become available for display?\n
- \n
- search engines crawling the web\n \n
- content aggregators indexing news sources\n \n
- app stores cataloguing available apps\n \n
\n - Filtering and Moderation — Which items are blocked or down-ranked under platform rules?\n
- \n
- spam filters\n \n
- removal of policy-violating content (hate speech, illegal content, graphic material)\n \n
- down-ranking of low-quality or misleading material\n \n
\n - Quality Scoring — Which sources or items are evaluated as more credible or higher-quality?\n
- \n
- search engines penalising low-quality sites\n \n
- news aggregators ranking by source authority\n \n
- peer-review-influenced ranking on academic search engines\n \n
\n - Trending Detection — Which items are surfaced as currently popular?\n
- \n
- trending topics on social platforms\n \n
- top charts on streaming services\n \n
- "What's happening" and "Today's headlines" sections\n \n
- popular hashtags\n \n
\n
→ Algorithmic curation defines the pool of information available on the platform. It largely operates the same way for all users.
\nAlgorithmic Personalisation
\nAlgorithmic operations that adapt the selection, order, and presentation of information to individual users based on their tracked signals. These signals accumulate over time into user histories that algorithms draw on.
\n→ Two users on the same platform — even with the same query — typically see substantially different content.
\n- \n
- Personalised Ranking — Which items are ordered higher for this user?\n
- \n
- personalised search results (location, history, profile shape ordering)\n \n
- social media feed ordering ("For You" feeds, "Top posts")\n \n
- engagement-based ranking — optimisation for predicted interaction, dominant on social media\n \n
\n - Recommendations — Which items are suggested to this user beyond what they actively requested?\n
- \n
- "Recommended for you" video lists\n \n
- suggested accounts, groups, or topics to follow\n \n
- "People you may know"\n \n
- related articles, similar products, "Up next"\n \n
\n - Personalised Advertising — Which adverts are targeted to this user?\n
- \n
- search ads tailored to past queries\n \n
- social media sponsored posts based on profile and behaviour\n \n
- retargeted display ads on websites\n \n
- influencer partnerships matched to audience interests\n \n
\n
→ Two users on the same platform — even with the same query — typically see substantially different content.
\n→ Personalisation creates a feedback loop: what users do affects what they see next, and what they see next can influence what they do.
\n☑ User Signals Tracked by the Platform for Algorithmic Personalisation
\nActions a user performs — actively or passively — within an information channel that may be tracked and used by algorithms to personalise the selection and visibility of information items.
\n→ User actions are not limited to deliberate interactions such as clicking or liking. Many actions are passive or automatic, such as how long a user stays on a page, how far they scroll, or where they are located. Users are often unaware that these actions influence what they encounter next.
\n| Type | \nWhat it is | \nExamples | \n
|---|---|---|
| Explicit feedback | \nDeliberate interactions the user chooses to perform | \n- searches / search queries - clicks - likes / reactions - comments / replies - shares / reposts / forwards - follows / subscribes - saves / bookmarks - ratings / reviews - purchases / downloads | \n
| Implicit behaviour | \nPassive behavioural signals captured during use | \n- watch time / listen time / dwell time - scroll behaviour (how far, how fast) - hover behaviour - skip behaviour | \n
| Contextual data | \nInformation about the situation in which the user is accessing the platform | \n- user's location data while using the application - device type (e.g., phone or laptop) - time of access | \n
| Account and social data | \nInformation from the user's profile and social connections | \n- profile information (age, interests, profession, gender) - language settings - linked accounts - contact list / address book | \n
- \n
- Adomavicius, G., & Tuzhilin, A. (2011). Context-aware recommender systems. In F. Ricci, L. Rokach, B. Shapira, & P. B. Kantor (Eds.), Recommender Systems Handbook (pp. 217–253). Springer. 7\" class=\"use-hover-preview\" rel=\"noopener noreferrer nofollow\">https://doi.org/10.1007/978-0-387-85820-37
\n
\n - Kelly, D., & Teevan, J. (2003). Implicit feedback for inferring user preference: A bibliography. ACM SIGIR Forum, 37(2), 18–28. https://doi.org/10.1145/959258.959260 \n
- Li, W., Kuo, J.-C., Sheng, M., Zhang, P., & Wu, Q. (2025). Beyond explicit and implicit: How users provide feedback to shape personalized recommendation content. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI '25). Association for Computing Machinery. https://doi.org/10.1145/3706598.3713241 \n
- Narayanan, A. (2023). Understanding social media recommendation algorithms. Knight First Amendment Institute, Columbia University. https://knightcolumbia.org/content/understanding-social-media-recommendation-algorithms\n \n
Information Amplification
\nAmplification refers to the systematic boosting of an item's visibility beyond the individual-user level — to produce broad visibility across user accounts, and sometimes across information environments.
\nWhere Source-Driven Promotion (above) covers what a single source itself does to gain visibility, and Platform-Side Gatekeeping (above) describes the algorithmic operations through which platforms surface and rank content for individual users, Amplification refers to the resulting boost outcomes at scale — produced either as the aggregate effect of those platform operations (→ Algorithmic Amplification) or through coordinated activity by multiple actors (→ Coordinated Amplification).
\nAmplification Mechanisms
\nAmplification operates through two principal mechanisms.
\n- \n
- Algorithmic Amplification is platform-driven: it is the aggregate effect of Algorithmic Gatekeeping — the cumulative outcome of platform curation and personalisation on which items reach which users and how prominently.\n \n
- Coordinated Amplification is actor-driven: multiple accounts, groups, or campaigns deliberately act in concert to boost the visibility of an item, hashtag, or narrative beyond what individual user activity would produce. The literature classifies it along two dimensions — the coordination (transparent or concealed) and the accounts (real or fake) — and distinguishes accordingly (Rogers & Righetti, 2025):\n
- \n
- Coordinated Authentic Amplification: coordination is transparent and accounts are real (e.g. open civic campaigns, advocacy, marketing).\n \n
- Coordinated Inauthentic / Artificial Amplification: coordination is concealed, accounts are fake, or both — manufacturing an appearance of organic support (Meta's Coordinated Inauthentic Behaviour / CIB; Gleicher, 2018).\n \n
\n
The two mechanisms frequently combine. Coordinated networks exploit engagement-based ranking to trigger algorithmic boosts; algorithmic ranking, in turn, compounds whatever visibility coordination has already produced.
\n- \n
- Gleicher, N. (2018). Coordinated Inauthentic Behavior Explained. Meta Newsroom. https://about.fb.com/news/2018/12/inside-feed-coordinated-inauthentic-behavior/\n \n
- Rogers, R., & Righetti, N. (2025). Coordinated inauthentic behaviour on Facebook? A typology of manufactured attention. https://doi.org/10.1177/29768624251369784 \n
Algorithmic Amplification
\nAlgorithmic amplification is the cumulative effect of the gatekeeping mechanisms above (Curation and Personalisation): the systematic shaping of which items, topics, accounts, and formats appear prominently to users — and which are filtered, demoted, or pushed down.
\nEmpirical research shows that engagement-based ranking systematically amplifies emotionally charged and out-group hostile content, even when users themselves do not prefer such content (Milli et al., 2025). It also compounds existing reach: accounts and items with high prior engagement are rewarded with further visibility, producing highly skewed reach distributions (rich-get-richer effect).
\nAlgorithmic interventions can have nonlinear effects in the opposite direction as well. A reduction of around 20% in an item's feed prominence can cut its reach by an order of magnitude (Narayanan, 2023).
\nAlgorithmic amplification is not a neutral reflection of user activity. Its effects are emergent and visible primarily in the aggregate: individual recommendations are imprecise (engagement rates remain below 1% on most platforms), but ranking, recommendation, and demotion systematically shape what circulates across the platform.
\n- \n
- Milli, S., et al. (2025). Engagement, user satisfaction, and the amplification of divisive content on social media. PNAS Nexus.\n \n
- Narayanan, A. (2023). Understanding social media recommendation algorithms. Knight First Amendment Institute.\n \n
Coordinated Authentic Amplification
\nCoordinated Authentic Amplification is the deliberate boosting of an information item, topic, hashtag, account, or narrative through openly disclosed, organised activity by real accounts. The coordinated origin is not concealed: participants act under their real identities or under known group affiliations.
\nTypical contexts include
\n- \n
- civic campaigns (e.g. NGO petitions, advocacy hashtags),\n \n
- political mobilisation (e.g. party campaigning, get-out-the-vote efforts),\n \n
- marketing and brand campaigns,\n \n
- professional association communications, and\n \n
- cultural movements such as Fridays for Future or #MeToo.\n \n
Whether the underlying message is well-founded, balanced, or one-sided is a separate question — authenticity refers only to the transparency of the coordination, not to the truth-value or fairness of the content. An authentic campaign can amplify accurate information, misleading information, or a one-sided position.
Authentic and inauthentic coordination can produce visibility patterns that look identical from the outside — synchronised sharing, hashtag clustering, rapid uptake. The distinguishing feature is not the visible pattern but whether the coordinated origin is openly disclosed.
Coordinated Inauthentic / Artificial Amplification
\nCoordinated Inauthentic / Artificial Amplification is the deliberate boosting of an information item, topic, hashtag, account, or narrative through organised activity in which the coordinated origin is concealed, the participating accounts are fake, or both. The aim is to manufacture an appearance of organic, independent support. Meta's term Coordinated Inauthentic Behaviour (CIB) — now incorporated into the EU Digital Services Act — centres on this combination of false identities and adversarial methods to evade detection (Gleicher, 2018; Rogers & Righetti, 2025).
\nTypical contexts include political influence operations (state-sponsored or party-aligned), astroturfing campaigns (commercial or ideological), targeted disinformation around elections, public health, or geopolitical conflict, and reputation manipulation through fake reviews, ratings, or engagement. The operational means — bots, trolls, sockpuppets, and their coordinated networks (bot farms, troll farms, sockpuppet networks, click farms) — are described in detail below.
\nInauthenticity refers to the concealment of the coordinated origin or the use of fake accounts — not to the truth-value of the content being amplified. A coordinated network of fake accounts can amplify accurate information; a single authentic individual can spread fabricated information. Coordinated inauthentic amplification and the spread of false content are distinct phenomena that can occur independently or together.
The following account types described in this section apply across Digital Information Channels & Platforms where users can create accounts and post or interact publicly — particularly social media, discussion forums and community spaces, video and audio platforms, and review or comment sections. They are less prominent in private communication apps or in environments without user-generated content.They appear both independently and within coordinated networks. They are listed here because of their typical role in amplification dynamics; the explicitly coordinated formations are the Account Networks.
\n| Term | \nDefinition | \nControlled by | \nDefined by | \nTypical purpose | \n
|---|---|---|---|---|
| Social Bot | \nAn automated or partly automated account that posts, likes, follows, shares, or replies online. | \nSoftware | \nAutomation | \nTo amplify messages, create artificial popularity, spam, influence debate, or spread content at scale. | \n
| Cyborg | \nA hybrid account combining human operation with software automation. | \nMixed: human and software | \nSelective automation | \nTo combine the scale of automation with the contextual plausibility of human input — for legitimate scheduling/management or for harder-to-detect influence operations. | \n
| Troll | \nA person or account that deliberately provokes, disrupts, or inflames online discussion. | \nUsually a human user; sometimes coordinated groups | \nDisruptive / provocative / antagonistic behaviour | \nTo upset others, derail conversations, provoke reactions, spread hostility, or polarise debate. | \n
| Sockpuppet | \nA fake account used by someone to hide their real identity or pretend to be a different person. | \nA human user, though the account may also use automation | \nDeceptive identity | \nTo create false support, attack others anonymously, evade bans, manipulate debate, or give the impression of independent agreement. | \n
Social Bot
\nA social bot is a bot designed to operate on social media platforms, posting, commenting, sharing, or interacting in ways that simulate human users. Social bots are typically programmed to act at scale and at high speed, far beyond what a human user could manage. Their activity is often repetitive and coordinated across many accounts, which distinguishes it from normal human use.
\nSocial bots can be used for legitimate purposes — such as customer service, news distribution, or marketing — but they are also widely used to influence public opinion, amplify certain messages, manipulate discussions, or manufacture the appearance of widespread support for specific ideas, products, or causes. In the context of misinformation and disinformation, social bots play a particular role in spreading content rapidly and giving the false impression that many independent voices share the same view.
\nWhen social bots are deployed in coordinated networks, they form a Bot Farm.
\nBot
\nA bot is a computer programme that automatically performs tasks, often repetitive ones. Bots range from simple, harmless tools — such as web crawlers that index pages for search engines, automated testing systems, or chatbots that answer routine customer questions — to malicious programmes designed to spread spam, malware, or disinformation.
Cyborg
\nA cyborg is a hybrid account that is partly operated by a human and partly automated by software. A cyborg may have routine posts scheduled or generated by software while a person handles selected interactions, replies, or sensitive content. The balance between automated and human activity varies between accounts.
\nCyborgs can be used for legitimate purposes — such as content scheduling, brand or institutional account management, or hybrid customer service — but they are also used in influence operations to combine the scale and speed of automation with the contextual plausibility of human input.
\nCyborgs are more difficult to identify than purely automated bots because part of their behaviour is genuinely human, which means single detection indicators rarely suffice for reliable identification.
\nTroll
\nA troll is a real person who deliberately disrupts online discussions through provocative, aggressive, or hostile behaviour. Trolls typically use personal accounts and target controversial issues, public figures (such as politicians or journalists), or media organisations. Their aim is to upset others, trigger reactions, or escalate conflicts — sometimes in support of a particular agenda, sometimes for entertainment or attention.
\nWhile trolls often act independently, they may also operate in coordinated groups, sometimes paid by political or commercial actors (see Troll Farm under Mechanisms of Amplification).
\nTrolling is best understood as a pattern of online behaviour, not a specific kind of account. The same behaviour can be performed by automated accounts, and ordinary users can engage in trolling on occasion.
\nSockpuppet
\nA sockpuppet is a fake online identity created and operated by a real person who hides their true identity. Unlike trolls — who often act under a single openly hostile account — a sockpuppet operator typically runs multiple fake accounts in parallel to create the impression that several independent users hold the same opinion, support the same cause, or agree with the operator's own (often separate) main account.
\nSockpuppets are commonly used to manufacture artificial consensus, support one's own arguments under different names, attack opponents while appearing impartial, evade bans by creating new identities after suspension, or manipulate online reviews, votes, and polls.
\nSockpuppets differ from social bots in that they are manually operated by humans, which makes their content more contextually plausible and harder to detect through automated means. They differ from trolls in that their primary goal is deception about identity and the manufacturing of apparent consensus, not provocation — although sockpuppet operators can also engage in trolling behaviour through their fake identities.
\nWhen a person or small group operates a coordinated set of sockpuppets together, they form a Sockpuppet Network (see Mechanisms of Amplification).
\n☑ Differentiating Social Bots, Trolls, and Sockpuppets
\n| Detection Dimension | \nSocial Bots | \nTrolls | \nSockpuppets | \n
|---|---|---|---|
| Profile Characteristics | \n- [ ] The account looks newly created - [ ] The profile is incomplete or generic - [ ] The username may look non-personal and sometimes include random numbers | \n- [ ] The account has typically been active for longer and has a post history - [ ] The profile is complete and seems personal; it may present strong ideological or political self-description - [ ] The username looks personal | \n- [ ] The profile looks plausible and personal, often with a profile picture and biographical details (sometimes stolen, AI-generated, or copied) - [ ] Account history may be moderate and designed to look authentic over time | \n
| Posting Behaviour | \n- [ ] The activity does not match normal human online behaviour - [ ] The accounts post or repost content very frequently - [ ] The accounts post or repost content at all hours, day and night | \n- [ ] The activity resembles normal human online behaviour - [ ] The account posts or replies at irregular times - [ ] The account becomes more active during controversial discussions | \n- [ ] Activity patterns resemble normal human use - [ ] Multiple accounts run by the same operator may show similar active hours or rhythms - [ ] Sockpuppets tend to start fewer discussions and write shorter posts than typical users | \n
| Interactions | \n- [ ] The account does not have real conversations - [ ] The accounts mostly like, share, or repost - [ ] Replies are short and automated | \n- [ ] The account replies directly to other users - [ ] The account engages in debates with the purpose of provoking reactions - [ ] Conversations are extended to create or escalate conflict | \n- [ ] The account engages in real conversations, often supporting the operator's main account or other sockpuppets - [ ] Replies are contextually appropriate and seem authentic - [ ] Pairs of sockpuppets often interact in the same discussion at similar times | \n
| Content Features | \n- [ ] The content is one-sided and repetitive - [ ] The same narratives are posted many times | \n- [ ] The content is specifically tailored to harm or provoke a target - [ ] The content targets individuals or social groups | \n- [ ] Content seems genuine and varied across accounts - [ ] The underlying message or stance aligns suspiciously across the network - [ ] More frequent use of personal pronouns such as "I" | \n
| Language | \n- [ ] Generic expressions, repetitive phrasing with keywords | \n- [ ] Varied, emotional, often abusive or offensive language | \n- [ ] Natural and varied language - [ ] Multiple accounts may share linguistic fingerprints (similar phrasing, vocabulary, punctuation, or error patterns) | \n
| Network & Technical Indicators | \n- [ ] Social bots follow other social bots, but the relationship is typically one-way and not reciprocal - [ ] Coordinated behaviour is observable across multiple bot accounts | \n- [ ] Trolls follow human accounts - [ ] Connections are often reciprocal (they follow their followers and vice versa) - [ ] Trolls typically act independently of each other | \n- [ ] Multiple accounts engaging with each other in mutually supportive ways - [ ] Connections may be artificially reciprocal between sockpuppets in the same network, or deliberately absent to avoid detection - [ ] Same IP address, device fingerprint, or login pattern \\\\*(platform-side detection)\\\\* - [ ] More clustered ego-networks than ordinary users - [ ] Correlated activity timing across accounts | \n
- \n
- Ferrara, E. (2023). Social bot detection in the age of ChatGPT: Challenges and opportunities. First Monday, 28(6). https://doi.org/10.5210/fm.v28i6.13185 \n
- Kumar, S., Cheng, J., Leskovec, J., & Subrahmanian, V. S. (2017). An army of me: Sockpuppets in online discussion communities. Proceedings of the 26th International Conference on World Wide Web (WWW '17), 857–866. https://doi.org/10.1145/3038912.3052677 \n
- Orabi, M., Mouheb, D., Al Aghbari, Z., & Kamel, I. (2020). Detection of bots in social media: A systematic review. Information Processing & Management, 57(4), 102250. https://doi.org/10.1016/j.ipm.2020.102250 \n
- Solorio, T., Hasan, R., & Mizan, M. (2013). A case study of sockpuppet detection in Wikipedia. Proceedings of the Workshop on Language Analysis in Social Media (LASM) at NAACL-HLT, 59–68. Association for Computational Linguistics. https://aclanthology.org/W13-1107/\n \n
- Tomaiuolo, M., Lombardo, G., Mordonini, M., Cagnoni, S., & Poggi, A. (2020). A survey on troll detection. Future Internet, 12(2), https://doi.org/10.3390/fi12020031 \n
- Tsikerdekis, M., & Zeadally, S. (2014). Multiple account identity deception detection in social media using nonverbal behavior. IEEE Transactions on Information Forensics and Security, 9(8), 1311–1321. https://doi.org/10.1109/TIFS.2014.2332820 \n
- Uyheng, J., Moffitt, J. D., & Carley, K. M. (2022). The language and targets of online trolling: A psycholinguistic approach for social cybersecurity. Information Processing & Management, 59(5), 103012. https://doi.org/10.1016/j.ipm.2022.103012 \n
Account Networks
\nBot Farm
\nA bot farm is a network of bots operating simultaneously across multiple devices or servers, deployed by a single operator or organisation for a particular purpose.
\nBot farms have a range of legitimate uses, including web indexing, automated software testing, data aggregation, and website performance monitoring. However, they are also commonly used for malicious activities such as creating fake engagement, generating large volumes of content, distributing spam, or carrying out cybersecurity attacks. When used to manipulate online discourse, bot farms can create the false impression of widespread support, opposition, or interest in a topic, account, or campaign.
\nTroll Farm
\nA troll farm is an organised group of coordinated, often paid workers who post deliberately provocative, misleading, or false content online — typically through fake accounts. Their aim is usually to manipulate public opinion, spread disinformation, or create social and political unrest. Troll farms have been documented in connection with state-sponsored influence operations as well as commercial reputation manipulation.
\nSockpuppet Network
\nA sockpuppet network is a coordinated set of sockpuppet accounts operated by one person or a small group, used to simulate independent voices supporting a shared narrative, campaign, account, or cause. Sockpuppet networks are commonly used in political astroturfing, review and rating manipulation, and coordinated disinformation campaigns. Unlike bot farms, sockpuppet networks rely on manual operation by humans, which makes the content of individual accounts appear more authentic and harder to detect through automated means. Their coordination usually becomes detectable only when multiple accounts can be linked through behavioural patterns, shared technical signals, or mutual engagement.
\nClick Farm
\nA click farm is an operation where large numbers of low-paid workers, automated bots, or both are used to click on ads, follow social media accounts, like posts, leave reviews, or download apps. The goal is to artificially boost online engagement or traffic, making content, accounts, or products appear more popular than they actually are.
\nPhenomena
\n☑ Virality vs. Trending
\n| Feature | \nVirality | \nTrending | \n
|---|---|---|
| What is being spread | \nA single information item: a specific video, post, image, or other piece of content | \nA topic, hashtag, sound, format, or discussion cluster: not one specific item, but many posts referring to or using the same thing | \n
| Primary drivers | \nUsers share, repost, or forward the information item to others, who in turn pass it along; this cascading spread can be further amplified by recommendation algorithms | \nMany users post about, mention, or use the same topic, hashtag, or format within a short time; the platform detects this concentration of activity and highlights it in a dedicated "Trending" section (such as a trending topics list, trending hashtag overview, or trending sounds page) | \n
| Time pattern | \nOften short and explosive; may recur later | \nTime-bound; persists as long as activity stays high or the platform keeps surfacing it | \n
| How it can be manipulated | \nCoordinated sharing, bot amplification, artificial engagement directed at the specific information item | \nCoordinated posting campaigns, manufactured fake trends through bot networks, platform decisions to promote, filter, or suppress | \n
Both virality and trending can emerge organically or be artificially amplified through coordinated campaigns, bot activity, or platform decisions. Both can also give an advantage to emotionally arousing, morally charged, or divisive content, especially in political or conflict-oriented contexts.
\nVirality
\nThe pattern by which a specific information item spreads rapidly through sharing, recommendation, and re-circulation across networks, analogous to the way a virus propagates. Virality is shaped by content characteristics, social network structures, platform affordances, timing, and algorithmic amplification.
\nContent that evokes high-arousal emotions, moral reactions, or out-group animosity is often more likely to be shared, especially in political or conflict-oriented contexts. However, virality is not determined only by the size of the original source: smaller accounts or outlets can also produce highly viral items.
\nVirality can emerge organically, but it can also be artificially amplified through coordinated sharing, platform manipulation, or bot activity.
\nTrending
\nA platform-assigned status indicating that a topic, hashtag, sound, format, or discussion cluster has received unusually concentrated activity within a short period.
\nTrending is identified algorithmically and surfaced through platform features such as X / Twitter Trending Topics, trending hashtags, trending sounds, trending challenges, or other platform-specific trend features. Trending depends on platform-specific signals such as post volume, rate of increase, engagement, location, personalisation, and moderation filters.
\nTopics that generate high engagement — including divisive, emotionally arousing, or morally charged topics — may be more likely to trend, but this depends on the platform's ranking system and moderation rules.
\nTrending can emerge organically from many independent contributions, but it can also be influenced by coordinated campaigns, bot activity, or platform decisions about what to promote, filter, moderate, or suppress.
\n- \n
- Berger, J. (2013). Contagious: Why Things Catch On. New York: Simon & Schuster.\n \n
- Berger, J., & Milkman, K. L. (2012). What makes online content viral? Journal of Marketing Research, 49(2), 192–205. https://doi.org/10.1509/jmr.10.0353 \n
- Brady, W. J., McLoughlin, K., Doan, T. N., & Crockett, M. J. (2021). How social learning amplifies moral outrage expression in online social networks. Science Advances, 7(33), eabe5641. https://doi.org/10.1126/sciadv.abe5641 \n
- Jenkins, H., Ford, S., & Green, J. (2013). Spreadable Media: Creating Value and Meaning in a Networked Culture. New York: NYU Press.\n \n
- Lee, J., & Umback, J. (2026). The viral turn: rethinking virality in the creator economy on TikTok. Continuum, 1–26. https://doi.org/10.1080/10304312.2026.2625794 \n
- Maarouf, A., Pröllochs, N., & Feuerriegel, S. (2024). The virality of hate speech on social media. Proceedings of the ACM on Human-Computer Interaction, 8 (CSCW1), 1–22. https://doi.org/10.1145/3641025 \n
- Rathje, S., Van Bavel, J. J., & van der Linden, S. (2021). Out-group animosity drives engagement on social media. Proceedings of the National Academy of Sciences, 118(26), e2024292118. https://doi.org/10.1073/pnas.2024292118 \n
- Rathje, S., & Van Bavel, J. J. (2025). The psychology of virality. Trends in Cognitive Sciences, 29(10), 914–927. https://doi.org/10.1016/j.tics.2025.06.014 \n
- Rogers, E. M. (2003). Diffusion of Innovations (5th ed.). New York: Free Press.\n \n
- Sangiorgio, E., Cinelli, M., Cerqueti, R., & Quattrociocchi, W. (2024). Followers do not dictate the virality of news outlets on social media. PNAS Nexus, 3(7), pgae257. https://doi.org/10.1093/pnasnexus/pgae257 \n
- Schlessinger, J., Garimella, K., Jakesch, M., & Eckles, D. (2023). Effects of Algorithmic Trend Promotion: Evidence from Coordinated Campaigns in Twitter's Trending Topics. Proceedings of the International AAAI Conference on Web and Social Media (ICWSM), 17(1), 777–786. https://doi.org/10.1609/icwsm.v17i1.22187 \n
- Schöne, J. P., Parkinson, B., & Goldenberg, A. (2021). Negativity spreads more than positivity on Twitter after both positive and negative political situations. Affective Science, 2(4), 379–390. https://doi.org/10.1007/s42761-021-00057-7 \n
Spill-Over Effects & Epistemic Laundering
\nThe process by which information that gains visibility within one information environment — whether through artificial amplification, trending, or editorial selection — is picked up and further distributed in other information environments or information access sytsems, thereby reaching audiences beyond the original environment.
\nSpill-over can occur through journalistic reporting, cross-platform sharing, editorial curation, or user-driven redistribution.
\n→ A research finding shared on a scholarly forum may be discussed on social media and summarised by an AI assistant.
→ A topic artificially amplified by bots on a social media platform may be picked up by journalists.
→ Spill-over effects can increase the reach of both reliable and unreliable information, and can make information appear more widely established than it originally was.
\nEpistemic Laundering
\nSpill-over does not always preserve the apparent status of information. When the receiving channel carries stronger signals of authority or reliability than the channel of origin — academic format, peer review, formal publication — the information itself can be perceived as more reliable simply through having moved. This effect is known as Epistemic Laundering: information gains perceived reliability through its passage across channels, without any actual change to the underlying claims or evidence. It exploits the tendency of recipients to attribute the reliability of the channel in which they encounter information to the information itself.
\n→ A claim originating in an anonymous blog post may be cited in a preprint, reproduced in an AI-generated answer, and finally cited in a peer-reviewed paper — at each step gaining academic surface and apparent authority, while the underlying claim remains unchanged or unverified.
\nA team at the University of Gothenburg, led by a medical researcher, invented a fake skin condition called Bixonimania to test whether AI systems would absorb and repeat medical misinformation. They presented it as a supposed condition linked to blue-light exposure from screens, with symptoms such as sore, itchy eyes and a pinkish hue on the eyelids. They then created deliberately fake academic-looking preprints, planted with obvious warning signs — a fictional author with an AI-generated photo, a non-existent university, and references to Starfleet Academy and the USS Enterprise. Nature reported that the preprints have since been removed from Preprints.org. Within weeks, major AI chatbots began reproducing Bixonimania as a real medical condition, in some cases offering users explanatory or health-related advice. In parallel, the fake material was cited in at least one published paper, since retracted, in the Springer Nature journal Cureus. Spill-over: log posts → fake preprint → webcrawlers → AI chatbot answers → academic citation
Stokel-Walker, C. (2026). Scientists invented a fake disease. AI told people it was real. Nature, 652(8110), 559-561. https://doi.org/10.1038/d41586-026-01100-y
Information Narrowing
\nWhereas Information Amplification (above) describes how visibility is broadened across user accounts, Information Narrowing describes the inverse: how the range of perspectives reaching an individual user or social group becomes restricted. Two distinct mechanisms produce this narrowing — Filter Bubble (algorithmic personalisation) and Echo Chamber (user self-selection). The two are often conflated in popular discourse but operate differently.
\nFilter Bubble
\nA filter bubble is an isolated information environment created by Algorithmic Personalisation, in which a user is increasingly exposed to content that aligns with their inferred preferences and past behaviour, while content that diverges is filtered out — typically without the user's awareness. The term was coined by Eli Pariser (2011) to describe how personalisation algorithms on Google, Facebook, and similar platforms can produce systematic exposure asymmetries based on user signals such as click history, location, and profile data.
\nThe defining feature of a filter bubble is unintentionality from the user's side: the narrowing is generated by the platform's optimisation, not by the user's deliberate choice of sources.
\nEmpirical research has substantially qualified Pariser's original thesis. Studies have found that algorithmic personalisation does shape what users see, but most users still encounter ideologically diverse content — partly because their own social networks include varied views, and partly because algorithms do not isolate as completely as the popular discourse suggests (Bakshy et al., 2015; Flaxman et al., 2016; Bruns, 2019). The filter-bubble effect is real but typically weaker than commonly assumed; pre-internet selective exposure (e.g., choosing newspapers or TV channels) was in many cases stronger.
- \n
- Pariser, E. (2011). The Filter Bubble: What the Internet Is Hiding from You. Penguin Press.\n \n
- Bakshy, E., Messing, S., & Adamic, L. A. (2015). Exposure to ideologically diverse news and opinion on Facebook. Science, 348(6239), 1130–1132. https://doi.org/10.1126/science.aaa1160 \n
- Flaxman, S., Goel, S., & Rao, J. M. (2016). Filter bubbles, echo chambers, and online news consumption. Public Opinion Quarterly, 80(S1), 298–320. https://doi.org/10.1093/poq/nfw006 \n
- Bruns, A. (2019). Are Filter Bubbles Real? Polity Press.\n \n
Echo Chamber
\nAn echo chamber is a social information environment in which a user is primarily exposed to opinions, claims, or ideologies that reinforce their existing beliefs, while dissenting views are absent, dismissed, or actively discredited. Cass Sunstein (2017) describes the political consequences: when groups insulate themselves from outside perspectives, internal beliefs intensify and become more extreme over time (group polarisation).
\nUnlike Filter Bubble, which arises from algorithmic personalisation, an echo chamber results primarily from user self-selection: choices about whom to follow, which communities to join, which sources to trust, and which voices to dismiss. These choices are partly driven by Confirmation Bias — the cognitive tendency to seek out and trust information that aligns with existing beliefs. The reinforcing effect comes from the social structure itself, not from invisible algorithmic filtering.
\nC. Thi Nguyen (2020) draws a conceptual distinction that matters for intervention:
\n- \n
- An epistemic bubble is a social structure in which other relevant voices are simply absent. Its inhabitants do not hear opposing perspectives, but they do not actively reject them.\n \n
- An echo chamber in the strict sense is a social structure in which other relevant voices are actively discredited. Members may hear opposing perspectives but learn to distrust their sources.\n \n
An epistemic bubble can be opened by introducing new information; an echo chamber resists correction even when external evidence is presented, because the sources of that evidence have already been delegitimised.
\nEmpirical work suggests that strong, ideologically isolated echo chambers are less common than popular discourse implies (Cinelli et al., 2021; Guess et al., 2018), but where they exist, they can be highly resistant to correction. Mere agreement within a group is not in itself an echo chamber — the defining feature is the active exclusion or discrediting of outside perspectives.
- \n
- Sunstein, C. R. (2017). #Republic: Divided Democracy in the Age of Social Media. Princeton University Press.\n \n
- Nguyen, C. T. (2020). Echo chambers and epistemic bubbles. Episteme, 17(2), 141–161. https://doi.org/10.1017/epi.2018.32 \n
- Cinelli, M., De Francisci Morales, G., Galeazzi, A., Quattrociocchi, W., & Starnini, M. (2021). The echo chamber effect on social media. PNAS, 118(9), e2023301118. https://doi.org/10.1073/pnas.2023301118 \n
- Guess, A., Lyons, B., Nyhan, B., & Reifler, J. (2018). Avoiding the Echo Chamber about Echo Chambers: Why Selective Exposure to Like-minded Political News Is Less Prevalent Than You Think. Knight Foundation White Paper.\n \n