INDEPENDENT DEEP-RESEARCH ASSIGNMENT: Shadowbanning, Invisible Voices, and Algorithmic Accountability

Full report

On this page

The proliferation of algorithmic curation has fundamentally altered the mechanics of public discourse, transforming the internet from a decentralized network of independent nodes into a series of centralized, computationally managed public squares. Social media platforms rely on highly complex, often opaque recommender systems and automated content moderation pipelines to filter billions of daily interactions. While these systems are strictly necessary to combat spam, illicit material, and coordinated inauthentic behavior, they possess the unilateral capacity to diminish the reach of legitimate public expression. The covert restriction of visibility—commonly referred to in colloquial terms as shadowbanning—strikes at the heart of cognitive liberty, defined within the BuildToResist movement as the fundamental right to hear and be heard without unseen algorithmic interference. This comprehensive report investigates the wide spectrum of algorithmic and platform-driven suppression. It distinguishes verified mechanisms of visibility reduction from speculative claims, outlines the technical architecture of recommender systems utilizing mathematical network models, details five historical case studies of covert suppression, and establishes rigorous, peer-reviewed protocols for algorithmic auditing. Furthermore, it outlines regulatory policy frameworks and decentralized architectures designed to reduce reliance on centralized ranking systems, culminating in a blueprint for integration into independent, community-driven platforms like BuildToResist.com.

A Taxonomy of Visibility Restrictions

To empirically analyze algorithmic suppression, precise terminology is required. The colloquial use of "shadowbanning" often conflates disparate technical processes ranging from routine spam filtering to ideological quality demotion1. The following table establishes a formal taxonomy of visibility restrictions deployed by contemporary digital platforms, delineating their technical mechanisms and the platform rationales used to justify their deployment.

The Architecture of Visibility and Algorithmic Control

Understanding how voices are silenced requires a deep analysis of how they are amplified. Modern recommender systems typically utilize a sophisticated three-stage architecture: candidate generation, scoring, and re-ranking14. In the candidate generation phase, the algorithmic system sifts through a massive corpus of data—often billions of individual posts—to select a small subset of relevant items for a specific user based on broad heuristics and historical user interactions14. The scoring phase subsequently utilizes deep neural networks to predict the probability of user engagement, assigning a precise relevance score to each candidate item14. Finally, the re-ranking phase adjusts the final output to account for system-wide constraints such as diversity, freshness, and platform safety guidelines14. Visibility restrictions frequently occur as mathematical penalties applied during the re-ranking phase. Trust and Safety classifiers operate in parallel with engagement models. If an automated classifier assigns a high probability that a post contains borderline content—such as uncivil language, coordinated inauthentic behavior, or politically sensitive material—the re-ranking algorithm mathematically penalizes the post's score, pushing it down the feed or excluding it entirely from recommendation surfaces1. Emerging academic research models shadowbanning through differential equations and opinion dynamics, demonstrating how platforms can theoretically manipulate entire populations. For instance, in large topological networks, the opinion shift of a user can be modeled as a differential equation where shadowbanning applies a multiplicative control factor between zero and one to the baseline posting rate of specific network edges16. By dynamically adjusting this control variable—where zero represents no censorship and one represents total censorship—platforms can silently shift the mean opinion of a network or minimize ideological variance without ever removing a single user from the platform16. This optimization-based approach demonstrates the profound, mathematically provable power of shadowbanning to manipulate public discourse at scale, operating entirely outside the awareness of the user base16.

Legitimate Friction versus Covert Suppression

Digital platforms possess legitimate, competing interests that justify limiting content distribution in specific contexts. The primary rationale for utilizing covert restrictions—such as classic shadowbanning—is rooted in the adversarial "cat-and-mouse" dynamic with malicious actors19. If operators of spam bot networks or coordinated state-backed troll farms receive explicit notifications when their content is demoted, they can rapidly reverse-engineer the moderation algorithm, adapt their computational tactics, and deploy new evasion strategies1. Therefore, algorithmic opacity is deliberately utilized as a defensive mechanism to preserve the integrity of the service1. Furthermore, platforms utilize visibility reduction to manage borderline content that touches on user safety but does not strictly violate the Terms of Service. This includes mitigating the reach of self-harm ideation, borderline eating disorder content, or coordinated harassment campaigns that rely on implicit rather than explicit threats2. Moderating this content through overt bans often generates immense public backlash, whereas algorithmic demotion achieves the platform's safety goals with minimal public friction5. However, this opacity poses severe systemic risks to public expression and cognitive liberty. When automated classifiers lack cultural competency, rely on flawed training data, or operate under overly broad policy directives, legitimate public discourse becomes collateral damage8. Shadowbanning circumvents traditional procedural justice; users cannot appeal a decision they do not know has been made against them1. This lack of due process undermines algorithmic accountability and disproportionately impacts marginalized communities, political dissidents, independent journalists, and specialized creators whose content is frequently misclassified as dangerous or spam5. The legal landscape surrounding these practices remains complex. The Supreme Court's ruling in Moody v. NetChoice confirmed that platform content moderation, including automated algorithmic curation, is protected by the First Amendment as a form of editorial discretion23. Because platforms possess the constitutional right to curate the speech they host, the burden of ensuring fairness and protecting invisible voices shifts away from direct legal mandates toward public accountability, independent auditing, regulatory transparency frameworks, and decentralized technological governance26.

Distinguishing Fact from Fiction: A Guide to Avoiding False Accusations

While algorithmic suppression is an empirically verified reality, the human psychological response to declining engagement frequently results in false accusations of shadowbanning. Users naturally seek external explanations for poor content performance, a phenomenon researchers term "algorithmic folk theories," wherein individuals collaboratively develop unverified hypotheses to make sense of opaque machine behaviors5. To maintain credibility in the pursuit of algorithmic accountability, it is vital to distinguish between covert suppression and standard network dynamics. Declining engagement is often the result of several observable, non-punitive factors. First, modern recommender systems operate on a principle of algorithmic meritocracy and quality ranking. These systems constantly test new content against a small sample audience. If the content fails to generate high watch time, likes, or comments during this initial exposure—as extensively modeled in leaked ranking documents from major short-video platforms—the algorithm will simply cease propagating it to a wider audience28. This is not a punitive ban; it is a mathematical failure to meet predefined scoring thresholds. Second, content formats that previously succeeded may experience natural decay as audience preferences shift and platform saturation occurs5. Recommender systems are designed to dynamically penalize repetitive content to avoid user boredom and maintain overall session length30. Third, chronological displacement plays a significant role; in feeds where velocity and recency matter, posting during low-traffic periods inherently limits reach regardless of content quality. Finally, network effects dictate that changes in the behavior of highly connected nodes—such as a prominent follower going inactive or changing their interaction habits—can severely impact an account's secondary reach without any direct algorithmic intervention against the creator. To avoid false accusations based solely on declining engagement, users must refrain from assuming censorship based on fluctuating view counts alone. Claims of shadowbanning should be reserved for instances where there is observable evidence of systemic intervention. This includes complete, verifiable exclusion from search results, an inability to populate in chronological hashtag feeds when utilizing highly specific tags, or the sudden, inexplicable severing of reach to non-followers despite stable, high-quality engagement metrics from an existing follower base.

Five Thoroughly Documented Case Studies of Visibility Restrictions

To distinguish systemic shadowbanning from speculative paranoia, it is necessary to examine verified historical instances where platforms restricted visibility without providing adequate notice or explanation to their user base. The following five case studies illustrate the diverse mechanisms through which digital platforms exert covert control over public expression.

1. The Twitter Files: Manual Blacklists and Visibility Filtering

In late 2022, internal corporate documents—subsequently known as the "Twitter Files"—revealed that the platform maintained specialized internal portals and tags to explicitly restrict the visibility of specific accounts without user notification4. Independent journalists granted access to these systems documented the existence of "Trends Blacklists" and "Search Blacklists," alongside internal administrative tags such as "Do Not Amplify"7. These engineering tools allowed administrators to prevent specific users from auto-populating in search results or appearing in trending topics, regardless of their organic engagement metrics. This revelation fundamentally demonstrated that shadowbanning—internally justified by executives as "visibility filtering"—was actively deployed to manage controversial political figures, commentators, and ideological debates4.

2. Meta's Systemic Suppression of Palestine Content (2023)

Following the outbreak of the Gaza war in October 2023, Human Rights Watch and a coalition of global civil society groups documented the systemic censorship of pro-Palestinian content on Facebook and Instagram6. The suppression was characterized by heavy reliance on flawed automated translation tools, overly broad enforcement of the Dangerous Organizations and Individuals policy, and unjustified algorithmic downranking of neutral news reporting21. Human Rights Watch documented over 1,000 specific cases of censorship, including instances where the algorithm blocked the hashtag \#AlAqsa and penalized accounts for standard journalistic practices8. The covert nature of the algorithmic demotion left users experiencing dramatic, sudden drops in viewership—often visually documented via "before and after" analytics screenshots—without receiving formal strikes or explanations on their accounts34.

3. Instagram and Threads: The Political Content Opt-Out (2024)

In February 2024, Meta implemented a sweeping policy change across Instagram and Threads, altering default user settings to no longer "proactively recommend" political content to users from accounts they did not explicitly follow9. While presented via corporate communications as a user-experience enhancement designed to reduce polarizing fights, the policy functionally severed organic discovery for independent journalists, activists, and civic organizations. A rigorous statistical study by Accountable Tech auditing five prominent political Instagram accounts—with a combined 13.5 million followers—recorded a staggering 65% total decline in average weekly reach over an 11-week period following the update9. Because the definition of "political" was opaque and determined entirely by machine-learning classifiers without user input, civic expression was effectively corralled into hidden algorithmic silos.

4. Meta's Blanket Ban on the Word "Shaheed"

For years, Meta’s internal moderation algorithms enforced a strict blanket ban on the Arabic word "shaheed" (loosely translated as "martyr"), automatically removing content that used the term under the presumption it inherently violated the Dangerous Organizations and Individuals policy38. This rigid automated enforcement led to massive visibility reductions, account penalties, and the silencing of Arabic speakers attempting to report on localized conflicts, debate political events, or mourn civilian casualties38. Following intense public pressure, the independent Meta Oversight Board issued a comprehensive policy advisory opinion in 2024, concluding that the ban was unnecessary and disproportionately restricted free expression. This intervention prompted Meta to end the blanket algorithmic suppression of the word, resulting in a measurable 19.5% increase in highly viewed posts utilizing the term in the months following the policy reversal38.

5. TikTok's "Algo Refresh" Placebo Experiment

Internal corporate documents, unsealed during a landmark lawsuit by the New York Attorney General, revealed that TikTok tested its "Algo Refresh" safety tool using deceptive "placebo" control groups44. The tool was theoretically designed to allow users to reset their algorithmic feeds to escape harmful, repetitive filter bubbles. However, during testing, thousands of young users were falsely informed that their feeds had been successfully reset, while the core recommendation algorithm was intentionally left unchanged44. Internal communications showed that a TikTok well-being manager explicitly warned that this placebo test conflicted with user transparency, yet the experiment proceeded45. This incident highlighted how platform operators can deliberately mislead users regarding their visibility and data environments, prioritizing experimental engagement metrics over algorithmic transparency and user safety44.

Evidence-Based Visibility Audit Protocol

To objectively detect visibility restrictions and move beyond algorithmic folk theories, researchers, digital rights advocates, and independent users must rely on rigorous algorithmic auditing. Drawing upon the foundational methodological taxonomy proposed by Sandvig et al., an effective audit must systematically query the algorithmic system with controlled inputs and observe the corresponding outputs46. The Sandvig taxonomy identifies several core approaches, including code audits (which are typically impossible due to trade secrets), API scraping audits, sockpuppet audits, and crowdsourced audits19. The following protocol outlines a rigorous, evidence-based method for evaluating possible visibility restrictions using a combination of controlled comparisons and sockpuppet methodologies. Phase 1: Baseline Establishment and Account-State Checks The foundation of any algorithmic audit is establishing a clear mathematical baseline. The auditor must extract historical data for the account in question, calculating average impressions, non-follower reach, and engagement rates over a minimum of three months to establish a statistical norm37. Concurrently, the auditor must utilize any available in-app account status tools to check for overt, explicit flags regarding recommendation eligibility, ensuring the drop in visibility is not due to an acknowledged policy violation. Phase 2: The Sockpuppet Audit (Controlled Environment) A sockpuppet audit utilizes automated or manually operated accounts that impersonate baseline users to observe the platform from an outside, unauthenticated perspective50. To ensure validity, the auditor must isolate variables by creating newly registered, sterile sockpuppet accounts utilizing different IP addresses, distinct hardware device identifiers, and unique behavioral patterns to prevent the platform's security infrastructure from linking them to the target account50. Once isolated, the auditor conducts a Search Deindexing Test. From the sterile sockpuppet accounts, the auditor searches for the target account's exact handle and recent, highly unique text strings. If the account fails to auto-populate while similarly sized accounts with similar engagement metrics do, a search shadowban is highly probable1. Subsequently, the auditor conducts a Hashtag and Feed Audit by posting unique tracking strings from the target account and attempting to discover this content using the sockpuppet accounts via chronological sorting. Complete failure to appear in these raw feeds indicates active, severe visibility filtering53. Phase 3: The A/B Content Audit and Referral Analysis To detect content-specific downranking, the auditor must deploy controlled comparisons. This involves posting non-controversial "control" content alongside the suspected "flagged" content at similar times on different days. The auditor then meticulously monitors the impression distribution channels over a defined period37. A steep, sudden drop in algorithmic referrals (specifically, reach to non-followers via recommendation surfaces) on the flagged content, while direct follower engagement remains statistically constant compared to the control content, is the strongest statistical indicator of a targeted recommendation exclusion9. Auditors must acknowledge the inherent limitations of this methodology. Algorithmic auditing is limited by platform opacity, the lack of API access for independent researchers, and the highly dynamic nature of A/B testing conducted on the server side26. Furthermore, auditors must rigorously account for personalization bias, ensuring that their sockpuppet accounts are not inadvertently trained by the auditor's own browsing habits, which would skew the algorithmic output52.

Psychological and Social Effects: The Chilling Effect of Invisibility

The impact of shadowbanning extends far beyond depressed engagement metrics; it deeply affects the psychological well-being of creators and the sociological health of digital communities. The profound inability to ascertain whether one is organically being ignored by peers or artificially silenced by an algorithm creates a psychologically damaging phenomenon known as "black-box gaslighting"5. Users are left to constantly question their own reality, the quality of their work, and the value of their contributions, leading to documented instances of frustration, severe emotional distress, and social isolation5. This opaque environment breeds a pervasive chilling effect across the network. To avoid potential, unseen algorithmic penalties, users engage in preemptive self-silencing or drastically alter their behavior and language. Marginalized communities, sex educators, and political dissidents increasingly rely on "algospeak"—modifying spellings (e.g., using "seggs" for "sex" or "unalive" for "suicide") in an attempt to evade automated text classifiers5. While this linguistic evolution demonstrates remarkable user resilience, it fundamentally forces communities to degrade their own language and creates massive barriers to discovering vital information and support networks. This dynamic effectively institutionalizes an environment where cognitive liberty is subverted by the cold, mathematical thresholds of machine-learning models5.

Transparency, Accountability, and Appeals

Addressing the systemic challenges of algorithmic suppression requires structural shifts in platform governance, blending robust regulatory mandates with independent oversight mechanisms and internal policy reforms. The European Union’s Digital Services Act (DSA) represents the most significant global regulatory advancement in algorithmic transparency to date. Specifically, Article 17 of the DSA mandates that platforms provide a "clear and specific statement of reasons" whenever they remove content, suspend accounts, or deliberately restrict visibility (including shadowbanning)11. This legislation forces platforms to formally acknowledge the use of visibility filters, thereby re-establishing procedural justice and giving users a definitive, legally mandated answer regarding the status of their content57. In the United States, legislative progress is slower, though state-level initiatives like the Illinois Consumer Privacy Act focus on broad transparency regarding how consumer data is collected and utilized, even if they stop short of mandating granular algorithmic explanations for content moderation59. While self-regulation by technology conglomerates remains deeply flawed, models like the Meta Oversight Board demonstrate the immense potential value of quasi-independent review. The Board provides a formalized appellate mechanism capable of reviewing complex cases, overturning algorithmic moderation decisions, and issuing binding policy recommendations61. In recent operational years, the Board overturned Meta’s original decision in approximately 70% to 90% of the cases it selected for review, actively pushing the company to update penalty systems, provide better localization for automated classifiers, and increase transparency regarding government takedown requests62. However, true structural accountability requires that such independent boards be granted deeper, unfettered access to audit the underlying recommendation algorithms themselves, rather than merely reviewing the outcomes of individual moderation decisions post-facto27. To adequately protect public expression, platforms must adopt the following policy recommendations regarding notice, accountability, and appeals:

1. Mandatory Notice of Demotion: Platforms must provide immediate, transparent notifications when an account or post is mathematically downranked for Trust & Safety reasons, permanently moving away from the use of covert shadowbanning against legitimate users65.

2. Expanded Appeals for Visibility: The fundamental right to appeal must extend beyond outright content removal to include severe visibility restrictions, algorithmic deindexing, and recommendation exclusions56.

3. Algorithmic Choice and Default Settings: Platforms should offer users the explicit ability to default to chronological feeds or actively opt-in or opt-out of specific algorithmic filters (such as political content suppressors), rather than imposing sweeping restrictions globally by default10.

Alternative Distribution: Reducing Reliance on Ranked Feeds

The ultimate, most resilient defense against algorithmic suppression is establishing structural independence from centralized recommendation feeds. By diversifying distribution channels and embracing open web standards, independent creators and communities can maintain direct, unmediated relationships with their audiences66. The Fediverse, powered primarily by the W3C-standardized ActivityPub protocol, represents a paradigm shift in decentralized social networking. Rather than a central corporate authority dictating visibility through a monolithic algorithm, ActivityPub operates on a federated server-to-server and client-to-server architectural model68. In this architecture, an actor's "outbox" pushes content directly to the "inbox" of their followers across disparate servers69. Because this federation relies on direct cryptographic delivery rather than engagement-optimized ranking, the concept of a centralized shadowban is architecturally impossible; receiving servers either accept the incoming content based on their local rules or overtly block the sending instance entirely67. Similarly, the AT Protocol (which powers networks like Bluesky) approaches the issue of cognitive liberty through the lens of strict data portability and algorithmic choice. In this model, users entirely own their social graph and data within Personal Data Servers (PDS), while moderation and content curation are handled through composable, user-selected labeling services rather than a single, inescapable platform algorithm67. Finally, legacy technologies such as Really Simple Syndication (RSS) and direct email subscriptions remain the most robust forms of digital distribution. They guarantee strict chronological, unfiltered delivery of content. Transitioning communities away from algorithmic feeds toward static independent websites, dedicated community forums, and RSS syndication effectively neutralizes the overarching power of platform-driven visibility restrictions70.

BuildToResist.com Integration Proposals

To flawlessly align the operational infrastructure of BuildToResist.com with its core movement message of cognitive liberty, the platform must implement technical systems and community guidelines that explicitly reject black-box moderation and embrace profound structural transparency.

Proposed Transparency Standard and Contribution Policy

BuildToResist.com should adopt an open-source moderation log, making all administrative actions publicly auditable. The contribution policy must preserve room for vigorous disagreement while maintaining clear, enforceable safety standards that explicitly reject covert visibility restrictions:

• Enforceable Conduct: The platform will strictly prohibit doxxing, direct threats of violence, and automated spam. These rules will be enforced through human review augmented by transparent, open-source filtering tools.
• Transparent Enforcement: When content violates these specific standards, it is overtly removed, and the author is provided an immediate, specific citation of the broken rule alongside a mechanism for appeal.
• No Shadowbanning Guarantee: The platform formally and legally commits to zero use of algorithmic downranking for ideological disagreements or controversial opinions. All user-generated content will be presented chronologically or via purely user-directed sorting mechanisms54. Room for disagreement is preserved by allowing users to independently mute or block individual accounts at their own discretion, placing curation power squarely in the hands of the reader rather than the platform administrator.

Recommending Alternatives to Ranked Feeds

To ensure that supporters of the movement are not silently erased by third-party social media algorithms, BuildToResist.com should integrate and heavily promote the following decentralized distribution strategies:

• Native RSS Feeds: Provide highly granular, distinct RSS feeds for different categories of content on the site, explicitly educating users on how to aggregate these updates using open-source feed readers71.
• ActivityPub Integration: Implement ActivityPub endpoints directly into the website's architecture, allowing fediverse users (e.g., on Mastodon or standard WordPress deployments) to follow the website directly from their preferred platforms. Articles published on BuildToResist.com will appear natively and chronologically in followers' inboxes without ever being subjected to Meta's or X's algorithmic scoring matrices69.
• Direct Subscriber Models: Encourage supporters to bypass social media entirely by utilizing direct email newsletters or encrypted messaging channels (e.g., Signal or Telegram broadcast channels) to guarantee a 1:1 delivery rate unencumbered by corporate recommender systems.

Public-Facing Guide: Don't Shadowban Yourself—or Let Others Silently Erase Your Reach

Cognitive liberty is the fundamental right to hear and be heard. In an era where centralized algorithms govern the flow of global information, maintaining your digital voice requires proactive defense and a deep understanding of the systems mediating your speech. 1\. Understand the Machine, but Do Not Internalize It Declining engagement is not always a shadowban. Algorithms are mathematically designed to prioritize highly stimulating, novel content that keeps users on the platform. If your reach drops, it may be due to audience fatigue, natural network decay, or algorithmic meritocracy. Do not fall victim to "black-box gaslighting," where you question the value of your own voice simply because a machine-learning model failed to amplify it to strangers. 2\. Refuse to Self-Censor The most insidious, damaging effect of algorithmic suppression is self-silencing. When you rely on "algospeak" or actively dilute your message to appease a platform's invisible Trust & Safety classifier, you are doing the algorithm's job for it. Speak clearly and authentically. If a platform deems your legitimate expression "borderline," force them to take overt, documented action. Do not shadowban yourself. 3\. Build Direct Connections Never build your community entirely on rented land. If your only link to your audience is governed by a Silicon Valley recommender system, your reach can be severed instantly and silently without any recourse.

• Use RSS: Encourage your audience to use RSS readers to follow your work chronologically, bypassing algorithmic sorting entirely.
• Collect Emails: A direct email list remains the most valuable asset on the internet; it is entirely immune to social media visibility filters.
• Explore the Open Web: Transition your community to decentralized protocols like the Fediverse (ActivityPub) or AT Protocol, where you cryptographically own your social graph and distribution is based on chronological delivery, not behavioral scoring.

4\. Audit Your Visibility If you suspect severe suppression, test it objectively rather than relying on anxiety. Log out of your accounts, utilize a VPN to mask your IP address, and search for your exact username or a highly specific, unique phrase you recently posted. If you do not appear in chronological search results, you may be facing a platform-level restriction. Document this evidence, inform your audience via alternative channels, and direct them immediately to your independent platforms. 5\. Demand Algorithmic Accountability Support legislative and civil society efforts that demand structural transparency, such as the European Union's Digital Services Act, which requires a formal "statement of reasons" when content is restricted. Visibility filtering should not be a corporate trade secret; when it controls the public square, it is a matter of profound public interest.

Conclusion

The architecture of modern digital platforms relies on incredibly complex algorithmic curation to manage the vast, unprecedented scale of human interaction. However, the weaponization of this architecture through shadowbanning, deindexing, and visibility filtering poses a systemic, existential threat to cognitive liberty. As demonstrated by verified historical interventions—from the manual blacklists of the Twitter Files to the automated algorithmic suppression of political and regional conflict content—platforms frequently deploy these covert tools to bypass procedural justice, leaving users voiceless and entirely unaware of their suppression. Combating this phenomenon requires a comprehensive, dual-track approach. First, society must demand strict algorithmic accountability through robust regulatory frameworks that mandate transparency, coupled with independent auditing protocols that drag black-box suppression into the light. Second, users and communities must actively decentralize their digital presence. By migrating toward chronological feeds, federated protocols like ActivityPub, and direct-distribution channels such as BuildToResist.com, we can successfully circumvent opaque recommender systems. True cognitive liberty demands a decentralized, open digital ecosystem where the fundamental right to hear and be heard is governed by the user, not the algorithm.

Works cited

1. (PDF) Shadowbanning \- ResearchGate, https\://www\.researchgate.net/publication/385317977\_Shadowbanning

2. Shadowbanning, https\://d-nb.info/1358292213/34

3. Shadowbanning \- EconStor, https\://www\.econstor.eu/bitstream/10419/315771/1/12599\_2024\_Article\_905.pdf

4. The Twitter Files, Part Two, Explained \- Gizmodo, https\://gizmodo.com/elon-musk-twitter-files-bari-weiss-part-two-shadow-ban-1849876635

5. (PDF) "What are you doing, TikTok?" : How Marginalized Social, https\://www\.researchgate.net/publication/380188336\_What\_are\_you\_doing\_TikTok\_How\_Marginalized\_Social\_Media\_Users\_Perceive\_Theorize\_and\_Prove\_Shadowbanning

6. Meta's censorship of Palestine content is 'systemic,' Human Rights, https\://mashable.com/article/human-rights-watch-meta-palestine-censorship-report

7. Twitter had 'secret blacklists' to limit users, journalist claims, https\://www\.aljazeera.com/economy/2022/12/9/twitter-had-secret-blacklists-to-limit-users-journalist-claims

8. Human Rights Due Diligence of Meta's Impacts in Israel and, https\://www\.bsr.org/reports/BSR\_Meta\_Human\_Rights\_Israel\_Palestine\_English.pdf

9. Report Finds Meta's Approach to Political Content Reduces Reach, https\://www\.techpolicy.press/report-finds-metas-approach-to-political-content-reduces-reach-for-activist-accounts/

10. Instagram users to see less of what Meta deems 'political' content, https\://www\.theguardian.com/technology/2024/mar/26/instagram-meta-political-content-opt-in-rules-threads

11. A Guide to Transparency under the EU Digital Services Act, https\://technologycoalition.org/wp-content/uploads/Tech-Coalition-Guide-to-Transparency-under-the-EU-DSA-Final.pdf

12. Meta censors pro-Palestinian views on a global scale, report claims, https\://www\.theguardian.com/technology/2023/dec/21/meta-facebook-instagram-pro-palestine-censorship-human-rights-watch-report

13. What is shadowbanning? How do I know if it has happened to me, https\://inforrm.org/2022/11/23/what-is-shadowbanning-how-do-i-know-if-it-has-happened-to-me-and-what-can-i-do-about-it-marten-risius-and-kevin-marc-blasiak/

14. Recommendation systems overview | Machine Learning, https\://developers.google.com/machine-learning/recommendation/overview/types

15. An introduction to Recommendation Systems \- AI Summer, https\://theaisummer.com/recommendation-systems/

16. arXiv:2310.20192v4 \[cs.SI\] 6 Feb 2024 \- Yale Insights, https\://insights.som.yale.edu/sites/default/files/2025-01/shadow%20banning.pdf

17. How Shadow Banning Can Silently Shift Opinion Online \- Yale Insights, https\://insights.som.yale.edu/insights/how-shadow-banning-can-silently-shift-opinion-online

18. Shaping Opinions in Social Networks with Shadow Banning \- arXiv, https\://arxiv.org/abs/2310.20192

19. Auditing Algorithms: Research Methods for Detecting Discrimination, https\://ai.equineteurope.org/system/files/2022-02/ICA2014-Sandvig.pdf

20. GitHub \- erwanlemerrer/awesome-audit-algorithms: A curated list of, https\://github.com/erwanlemerrer/awesome-audit-algorithms

21. How Meta censors Palestinian voices \- Access Now, https\://www\.accessnow.org/publication/how-meta-censors-palestinian-voices/

22. Shadowbanning: Some marginalized social media users believe, https\://www\.si.umich.edu/about-umsi/news/shadowbanning-some-marginalized-social-media-users-believe-their-content-suppressed

23. Moody Decision Confirms First Amendment Protects Online Platforms, https\://www\.dwt.com/insights/2024/07/scotus-moody-ruling-a-win-for-online-platforms

24. MOODY v. NETCHOICE, LLC | Supreme Court \- Law.Cornell.Edu, https\://www\.law.cornell.edu/supremecourt/text/22-277

25. Moody v. NetChoice: The Supreme Court Meets the Free Speech, https\://www\.journals.uchicago.edu/doi/10.1086/735007

26. platform-supported auditing of social media algorithms \- OpenReview, https\://openreview.net/pdf?id=qj6KORbx5o

27. Meta's Oversight Board Needs Access to Facebook's Algorithms to, https\://www\.brennancenter.org/our-work/analysis-opinion/metas-oversight-board-needs-access-facebooks-algorithms-do-its-job

28. Leaked TikTok Doc Reveals Its Obvious Secret to an Addictive Feed, https\://gizmodo.com/leaked-tiktok-doc-reveals-its-obvious-secret-to-an-addi-1848166901

29. How TikTok's Algorithm Really Works: A Deep Dive Into ... \- Kingy AI, https\://kingy.ai/news/how-tiktoks-algorithm-really-works-a-deep-dive-into-the-unsupervised-learning-engine-behind-the-for-you-page/

30. Leaked Info Reveals How TikTok's Algorithm Works \- DeepLearning.AI, https\://www\.deeplearning.ai/the-batch/what-makes-tiktok-tick

31. Bari Weiss Twitter Files Reveal 'Blacklisting' of Disfavored Content, https\://reason.com/2022/12/09/bari-weiss-twitter-files-elon-musk-blacklist-shadow-banning/

32. Twitter Files \- Wikipedia, https\://en.wikipedia.org/wiki/Twitter\_Files

33. Twitter Files of internal company documents attract extensive media, https\://www\.aa.com.tr/en/world/twitter-files-of-internal-company-documents-attract-extensive-media-coverage/2779677

34. Meta's Broken Promises: Systemic Censorship of Palestine Content, https\://www\.hrw.org/report/2023/12/21/metas-broken-promises/systemic-censorship-palestine-content-instagram-and

35. Meta Will Stop Recommending Political Content On Instagram, https\://www\.mediapost.com/publications/article/393411/meta-will-stop-recommending-political-content-on-i.html?edition=133287

36. Instagram, Threads to stop recommending some political content, https\://mashable.com/article/instagram-threads-political-content

37. Meta's Political Content Limit Causes Steep Drop in Reach for, https\://accountabletech.org/research/metas-political-content-limit-causes-steep-drop-in-reach-for-accounts/

38. Meta's oversight board urges Facebook, Instagram to lift ban on, https\://www\.aljazeera.com/news/2024/3/30/metas-oversight-board-urges-review-of-shaheed-censorship-policy

39. Referring to Designated Dangerous Individuals as “Shaheed”, https\://www\.oversightboard.com/decision/pao-lopp03uk/

40. Meta ends ban on the word 'shaheed' amid censorship concerns, https\://www\.business-humanrights.org/en/latest-news/meta-ends-ban-on-the-word-shaheed-amid-censorship-concerns/

41. Meta Accepts Key Oversight Board Recommendations to End, https\://www\.oversightboard.com/news/meta-accepts-key-oversight-board-recommendations-to-end-blanket-ban-on-shaheed/

42. Oversight Board Publishes Policy Advisory Opinion on Referring to, https\://www\.oversightboard.com/news/oversight-board-publishes-policy-advisory-opinion-on-referring-to-designated-dangerous-individuals-as-shaheed/

43. 2024 Annual Report | The Oversight Board, https\://www\.oversightboard.com/wp-content/uploads/2025/08/2024-Annual-Report.pdf

44. TikTok Accused of Misleading Users Over Algorithm Reset and, https\://the420.in/tiktok-algo-refresh-safety-lawsuit/

45. TikTok faces scrutiny over placebo test of algorithm-reset feature, https\://www\.cryptopolitan.com/tiktok-faces-scrutiny-over-placebo-test-of-algorithm-reset-feature/

46. Auditing Algorithms \- Stanford HCI Group, https\://hci.stanford.edu/publications/2021/FnT\_AuditingAlgorithms.pdf

47. Auditing Algorithms : Research Methods for Detecting Discrimination, https\://www\.semanticscholar.org/paper/Auditing-Algorithms-%3A-Research-Methods-for-on-Sandvig-Hamilton/b7227cbd34766655dea10d0437ab10df3a127396

48. Understanding the Power of Everyday Users in Surfacing Harmful, https\://www\.andrew.cmu.edu/user/hongs/files/Everyday.Algorithm.Auditing\_CSCW2021.pdf

49. Examining Algorithmic Curation on Social Media \- arXiv, https\://arxiv.org/pdf/2502.20491

50. A Systematic Literature Review of Algorithm Audits \- arXiv, https\://arxiv.org/pdf/2102.04256

51. Auditing Algorithms \- Morgan Klaus Scheuerman, https\://www\.morgan-klaus.com/readings/auditing-algorithms.html

52. Auditing Algorithmic Explanations of Social Media Feeds, https\://ojs.aaai.org/index.php/ICWSM/article/download/31376/33536/35432

53. How Algorithms Can Influence Content Visibility on Social Media, https\://arxiv.org/html/2410.17390v1

54. Can We Fix Social Media? Testing Prosocial Interventions using, https\://arxiv.org/html/2508.03385v1

55. \[2212.11366\] Statistical Challenges in Online Controlled Experiments, https\://arxiv.org/abs/2212.11366

56. Full article: 'Dysfunctional' appeals and failures of algorithmic justice, https\://www\.tandfonline.com/doi/full/10.1080/1369118X.2024.2396621

57. EU Should Improve Transparency in the Digital Services Act, https\://itif.org/publications/2025/10/20/eu-should-improve-transparency-in-the-digital-services-act/

58. Overview Documentation \- DSA Transparency Database, https\://transparency.dsa.ec.europa.eu/page/documentation

59. Privacy Bill Essentials: Illinois | Hinshaw & Culbertson LLP, https\://www\.hinshawlaw.com/en/insights/privacy-cyber-and-ai-decoded-alert/privacy-bill-essentials-illinois

60. Data Privacy | Consumer Law | CTM Legal Group, https\://www\.ctmlegalgroup.com/practice-areas/consumer-law/data-privacy

61. Oversight Board | Improving how Meta treats people and, https\://www\.oversightboard.com/

62. Meta's Oversight Board: A Review and Critical Assessment, https\://www\.researchgate.net/publication/364674429\_Meta's\_Oversight\_Board\_A\_Review\_and\_Critical\_Assessment

63. 2023 Annual Report | The Oversight Board, https\://www\.oversightboard.com/wp-content/uploads/2024/06/Oversight-Board-2023-Annual-Report.pdf

64. 2021 Annual Report \- The Oversight Board, https\://www\.oversightboard.com/wp-content/uploads/2023/11/Annual-Report-English.pdf

65. Shedding Light on Shadowbanning, https\://cdt.org/insights/shedding-light-on-shadowbanning/

66. FAQ \- AT Protocol, https\://atproto.com/guides/faq

67. Understanding the Protocols Behind Mastodon and Bluesky | Fediview, https\://fediview.com/articles/activitypub-vs-atproto-understanding-protocols/

68. Guide for new ActivityPub implementers \- SocialHub, https\://socialhub.activitypub.rocks/t/guide-for-new-activitypub-implementers/479

69. ActivityPub \- W3C, https\://www\.w3.org/TR/activitypub/

70. Static ActivityPub Publishing \- Social Web Foundation, https\://socialwebfoundation.org/2026/09/04/static-activitypub-publishing/

71. A Guide to Implementing ActivityPub in a Static Site (or Any Website), https\://maho.dev/2024/02/a-guide-to-implementing-activitypub-in-a-static-site-or-any-website-part-4/