The Invisible Editor Outcome Casebook

Full report

On this page

The Architecture of Algorithmic Visibility

The contemporary digital public sphere is governed by an invisible editor—an intricate, planetary-scale matrix of automated detection systems, algorithmic ranking signals, and human content moderators acting under the directives of corporate policy and state law. Unlike traditional editorial models characterized by transparent human curation and identifiable accountability, this invisible editor operates at a scale and velocity that obfuscates the precise mechanisms of information visibility. Alterations to digital information flow, ranging from outright removal and geo-blocking to subtle downranking, synthesized prompt reframing, and automated profile modifications, rarely manifest to the user as explicit censorship. Often, these interventions masquerade as benign product updates, user safety protocols, or neutral algorithmic optimizations. To systematically analyze this infrastructure, a critical distinction must be established from the outset: a decline in web traffic or audience reach alone does not constitute definitive proof of covert suppression. Platforms routinely optimize their recommendation engines for user retention, engagement, and shifting consumer preferences, meaning that not every ranking decision is an act of ideological censorship or malicious silencing. However, when alterations to visibility are executed as punitive measures, as tools for regulatory leverage, or as automated collateral damage stemming from poorly trained models, the architecture of moderation becomes a mechanism of systemic control. This casebook exhaustively documents twenty distinct interventions, treating each mode of algorithmic alteration—removal, access or regional restriction, search exclusion, recommendation exclusion, downranking and reduced distribution, labeling, monetization changes, synthesized-answer reframing, personalized ranking or refusal, account penalties, and saved-memory or profile changes—as distinct actions. By analyzing the multi-layered outcomes of these governance choices across diverse technological ecosystems, this report elucidates the profound downstream impacts on global discourse, press freedom, and cognitive sovereignty.

Section 1: Erasure, Removal, and Monetization Changes

The most definitive and absolute action an algorithmic system can execute is total removal. Yet, the friction between rigid, globally applied corporate policies and highly localized cultural, political, or journalistic contexts frequently results in egregious false positives. As the volume of digital communication expands exponentially, reliance on automated flagging mechanisms disproportionately shifts the burden of legibility onto marginalized users and legitimate news organizations. This section examines cases of direct content removal and monetization suspension, highlighting the interplay between human oversight, algorithmic detection, and independent appellate bodies such as the Meta Oversight Board.

Case 1: The Ayahuasca Brew

The clash between standardized global drug policies and localized religious expression exposes the fundamental limitations of at-scale content moderation. In this instance, an Instagram account for a spiritual school based in Brazil posted an image of a dark brown liquid described in Portuguese as ayahuasca, a plant-based brew containing the psychoactive compound DMT1. The content was flagged by automated systems because it was "trending" and rapidly accumulating views2. A human moderator subsequently removed the post under the premise that it encouraged the use of a non-medical drug2. This case illustrates the invisible application of cross-platform rules. At the time, Instagram's Community Guidelines only explicitly prohibited the buying or selling of illegal or prescription drugs, not the positive discussion of them1. However, Meta was quietly enforcing Facebook’s more expansive "Regulated Goods" standard on Instagram users without notifying them of this discrepancy1. The Meta Oversight Board overturned the removal, arguing that prohibiting positive comments about traditional religious ceremonies was not necessary to protect public health and violated international human rights principles1. Notably, the Board's minority raised concerns regarding administrability, arguing that granting exceptions for religious use would require case-by-case examinations that automated systems and rapid-fire human reviewers could not sustainably execute at scale4.

Case 2: AI-Manipulated Ronaldo Gambling Advertisement

Generative artificial intelligence has introduced unprecedented velocity to online fraud, exploiting the very algorithmic distribution networks designed to surface engaging advertising. In this case, malicious actors utilized an AI-manipulated deepfake of Brazilian soccer legend Ronaldo Nazário to promote an online game, leveraging Meta's paid boosting tools to maximize visibility5. The systemic failure occurred in the bifurcation of organic and paid moderation. Meta’s automated systems flagged and disabled the paid advertisement for violating the "Unacceptable Business Practices" standard5. However, the original organic post remained live and active on the platform, despite clear video-audio mismatches and deceptive outbound links5. Human reviewers failed to identify the AI watermarks and contextual clues, leaving the scam active until the Oversight Board intervened5. This disjointed enforcement highlights a failure to adhere to the United Nations Guiding Principles on Business and Human Rights, which mandates the mitigation of adverse impacts stemming from monetized corporate operations5.

Case 3: Brazilian General's Call for a Congressional Siege

The latency of human moderation during critical geopolitical events can catalyze real-world violence. In the immediate aftermath of Brazil's highly polarized 2022 General Election, a Facebook user posted a video of a Brazilian general urging citizens to "hit the streets" and march on the National Congress and the Supreme Court, accompanied by images of the Three Powers Plaza in Brasília7. Despite being reported seven times by four different users, human moderators repeatedly determined that the post did not violate Meta's Violence and Incitement Community Standard7. The reviewers fatally misclassified the call to "besiege" government buildings as permissible political rhetoric rather than an unambiguous call for forcible entry into high-risk locations7. The content was only removed after the Oversight Board selected the case, by which time the algorithmic distribution had already contributed to the environment that culminated in the January 8 riots7. The Board heavily criticized Meta for lacking specific metrics to evaluate the success of its election integrity efforts, demonstrating that reactive moderation is vastly insufficient for safeguarding democratic transitions7.

Case 4: United Kingdom Southport Riots Disinformation

When tragic offline events occur, algorithmic recommendation systems frequently prioritize the velocity of sensationalism over the verification of facts. Following a knife attack at a dance workshop in Southport, UK, in July 2024, widespread disinformation rapidly circulated regarding the identity of the 17-year-old suspect, Axel Rudakubana8. Online actors propagated a false name and fabricated background for the attacker, successfully triggering Islamophobic riots across the United Kingdom8. Meta's internal systems failed to remove two out of three highly inflammatory posts flagged to the Oversight Board8. The core failure stemmed from an ambiguity in Meta's Violence and Incitement policy, which lacked clarity regarding high-severity threats directed at places (such as mosques) as opposed to people8. By the time the algorithmic spread was arrested and the content removed, the digital disinformation had already been successfully laundered into physical property destruction and communal violence.

Case 5: Contextualizing "From the River to the Sea"

The algorithmic policing of geopolitical slogans forces platforms to act as ultimate arbiters of linguistic intent, often conflating contextual solidarity with hate speech. In reviewing three distinct Facebook posts utilizing the phrase "From the River to the Sea" in the aftermath of the October 7 Hamas attacks, Meta’s human reviewers opted to leave the content live9. The Oversight Board upheld this decision, noting that the specific posts did not attack Jewish or Israeli people with calls for violence or exclusion, nor did they signal imminent violence9. Instead, they contained contextual signals of solidarity with Palestinians, such as the hashtag "\#ceasefire"9. This case highlights the tension in labeling and removal architectures; automated tools frequently fail to grasp geopolitical idioms, necessitating human intervention that invariably angers one demographic group while protecting the expression of another. Furthermore, the Board used this high-profile case to demand that Meta restore researcher access to its Content Library, arguing that the deprecation of tools like CrowdTangle severely limited independent analysis of how these conflicts are algorithmically managed9.

Case 6: The Shared Al Jazeera Post

The over-enforcement of anti-terrorism policies frequently penalizes legitimate journalistic reporting, casting a long shadow over freedom of expression. In May 2021, an Egyptian Facebook user shared a verified Al Jazeera news post detailing a threat made by the military wing of Hamas, adding the caption "Ooh" in Arabic11. Meta’s human reviewers inexplicably removed the user's post under the "Dangerous Individuals and Organizations" policy, claiming it constituted substantive support for a designated terrorist entity11. This algorithmic and human error failed to recognize the policy’s explicit exceptions for neutral discussion and news reporting11. The removal created a severe informational asymmetry: the original Al Jazeera post remained untouched, while citizens attempting to amplify and discuss the news were algorithmically penalized11. The Oversight Board’s reversal of this removal prompted Meta to commission an independent human rights due diligence report from BSR, which subsequently confirmed that Meta’s content moderation practices had an adverse, systemic impact on the rights of Palestinian users to freedom of expression and assembly12.

Case 7: Ketamine Promotion and the Two-Tiered System

Monetization infrastructures inevitably create separate regulatory realities for high-profile users, eroding the premise of neutral algorithmic governance. An Instagram influencer with roughly 200,000 followers posted content describing their ketamine therapy for depression as a "magical entry into another dimension," complete with psychedelic imagery and a "paid partnership" label indicating a commercial relationship with a therapy provider14. The content was removed and restored three separate times14. The user’s status as a "managed partner" allowed them to bypass standard algorithmic purgatory and escalate the issue directly to internal subject matter experts, who erroneously restored the post six months later14. The Oversight Board ultimately overturned the restoration and mandated the post's removal, noting that the content violated both the Branded Content policies and the Restricted Goods policy, as it promoted achieving a "high" without sufficient reference to a supervised medical diagnosis14.

Section 2: Sovereign Borders in the Digital Commons

While the early internet was conceptualized as a borderless expanse, state actors increasingly compel multinational platforms to enforce regional access restrictions. These interventions do not merely remove content from servers; they weaponize geographic algorithmic architecture to comply with local authoritarian legal frameworks. This effectively fractures the global digital commons into compliant, state-regulated intranets, forcing platforms to choose between abandoning a market or operating as an agent of state censorship.

Case 8: Royalist Marketplace Geo-Blocking in Thailand

When democratic discourse conflicts with entrenched monarchal laws, platforms face the ultimate dilemma of complicity versus expulsion. In April 2020, Pavin Chachavalpongpun, a self-exiled academic in Japan, created a Facebook group named "Royalist Marketplace" to facilitate open, critical discussion of the Thai monarchy15. Within months, the group amassed over one million members, representing an unprecedented digital organizing space in a country governed by strict Lese-majeste laws and the Computer Crimes Act, which threaten up to 15 years in prison for defaming the king15. Under immense pressure from the Thai Ministry of Digital Economy and Society—which threatened criminal proceedings against Facebook's local executives and fines of 200,000 baht—Facebook instituted a geographic algorithmic block16. Users within Thailand attempting to access the group were met with a notice stating the content was restricted pursuant to a legal request19. While Facebook announced its intent to mount a legal challenge, citing international human rights law (ICCPR), the immediate execution of the block effectively silenced a massive democratic movement16. The action inadvertently triggered a Streisand effect; Pavin immediately created a replacement group, "Talad Luang," which gained 500,000 members in a single day, demonstrating the relentless fluidity of user behavior against algorithmic border walls17.

Case 9: Withholding The Caravan and Farmer Protests in India

Intermediary liability laws are frequently invoked by governments to execute sweeping crackdowns on political opposition and the independent press. During the height of the Indian farmers' protests in early 2021, the Ministry of Electronics and Information Technology (MeitY) issued emergency blocking orders under Section 69A of the Information Technology Act20. The government demanded Twitter block hundreds of accounts, claiming they were using the hashtag "\#ModiPlanningFarmerGenocide" and inciting a threat to public order20. Among the targeted accounts were highly prominent news organizations like The Caravan, activist fronts like Kisan Ekta Morcha, and various political commentators21. Twitter initially complied, placing "Country Withheld Content" labels on the accounts, effectively erasing their visibility within Indian borders20. Following internal pushback arguing that the accounts represented newsworthy free speech, Twitter restored access, which immediately prompted the Indian government to threaten the company's senior executives with up to seven years in jail and steep financial penalties for non-compliance20. The standoff culminated in a high-stakes legal challenge in the Karnataka High Court regarding the constitutionality of the blocking orders24. The case exposes the extreme vulnerability of algorithmic distribution networks to sovereign coercion.

Section 3: The Architecture of Invisibility

Beyond outright removal and geo-blocking, the invisible editor manipulates the fundamental architecture of digital discovery. Search exclusions, recommendation throttling, and the unilateral alteration of user interface (UI) discovery tabs exercise profound, systemic control over digital economies. These actions often serve broader corporate geopolitical strategies, acting as leverage against nation-states, or they reflect deep-seated paternalistic biases in algorithmic engineering.

Case 10: European Union Publisher News Test

Monopolistic platforms possess the infrastructural capacity to unilaterally alter the public sphere. When corporate interests collide with regulatory frameworks, platforms can leverage their audience traffic as a devastating negotiating tactic. In late 2024, responding to pressures from the European Copyright Directive (EUCD), Google initiated a "small, time-limited test" impacting 1% of users across nine European nations, including Belgium, France, Italy, and Poland26. The experiment algorithmically omitted all search results from EU-based news publishers across standard Search, Google News, and the personalized Discover feed26. Google framed this as a data-gathering exercise to "assess how results from EU news publishers impact the search experience"26. However, the real-world consequence was a punitive demonstration of infrastructural power: by artificially starving European publishers of their primary source of inbound traffic, Google showcased its ability to decimate digital journalism revenues overnight. This tactic mirrors previous threats in Australia and Canada, where visibility itself is held hostage to extract regulatory concessions26.

Case 11: California Journalism Preservation Act (CJPA) Blackout

Mirroring the European test, algorithmic search exclusions are routinely deployed domestically to combat proposed "link-tax" legislation. As California debated the California Journalism Preservation Act (AB886)—a bill aimed at forcing tech giants to pay a "journalism usage fee" for linking to news publishers—Google retaliated by removing links to California news outlets for a percentage of the state's users26. This algorithmic throttling was accompanied by a suspension of investments in the California news ecosystem, including the Google News Initiative29. The resulting traffic starvation fundamentally altered the legislative reality on the ground. In August 2024, fearing the collapse of their digital distribution, publishers and lawmakers abandoned the CJPA in a last-minute, closed-door deal championed by Assemblywoman Buffy Wicks29. The legislation was replaced by a $172 million public-private partnership, which included funding for an "AI Accelerator program"—a massive concession to the tech industry29. The incident proved that control over algorithmic discovery can effectively supersede the democratic legislative process.

Case 12: NT1 vs. Google (Right to be Forgotten - Rejected)

The privatization of judicial balancing forces search engines to adjudicate the profound tension between individual privacy rights and open, historical justice. Following the European Court of Justice's establishment of the "Right to be Forgotten," the UK High Court heard the cases of two businessmen seeking to force Google to de-index links related to their past criminal convictions under the Data Protection Act 199832. In the case of "NT1," the claimant had been convicted of a serious criminal conspiracy involving false accounting and tax evasion in the late 1990s34. Google refused the de-listing request35. The High Court ruled in Google's favor, determining that the conviction involved dishonesty, that the claimant maintained a role in public life, and that the information remained highly relevant to protect the public from being misled in future business dealings35. By refusing the search exclusion, the algorithm ensured that NT1's past remained permanently anchored to his digital identity.

Case 13: NT2 vs. Google (Right to be Forgotten - Granted)

Conversely, when a legal conviction is deemed sufficiently "spent," search exclusion acts as a powerful mechanism of digital rehabilitation. In the parallel case of "NT2," the claimant had been convicted of authorizing unlawful phone and computer hacking linked to a controversial business facing environmental protests34. Unlike NT1, NT2's conviction was deemed out of date, irrelevant, and of no sufficient legitimate interest to Google's users35. The court noted that NT2 had a young family and possessed a reasonable expectation of privacy35. As a result, Google was ordered to de-index the URLs relating to the conviction35. This ruling highlights a profound truth of the modern web: while the original source material (news articles, court records) continues to exist on remote servers, it is rendered practically invisible without algorithmic surfacing. Search exclusion thus operates as a de facto erasure from public consciousness.

Case 14: TikTok's Shadowbanning of Marginalized Bodies

Algorithmic paternalism often results in a form of benevolent discrimination, wherein systems designed to protect users actually punish them. An investigation by Netzpolitik revealed that TikTok had instituted internal moderation policies aimed at curbing cyberbullying by identifying users deemed highly susceptible to harassment38. Instead of policing the bullies, TikTok's invisible editor penalized the victims. Human moderators were instructed to place special flags on accounts belonging to individuals who were visibly disabled, overweight, or LGBTQ+38. Once tagged, these users were subjected to severe algorithmic downranking; their content was artificially restricted from achieving wide distribution on the platform's highly lucrative "For You" page38. The platform admitted to this practice, framing it as a misguided, temporary safety protocol38. This case perfectly illustrates how covert downranking shapes the visual culture of a platform, enforcing normative aesthetics by burying marginalized bodies in algorithmic darkness.

Case 15: The Death of the "Recent" Tab

Structural changes to user interfaces fundamentally alter the digital economy, acting as a form of macro-level algorithmic control. It is vital to reiterate that traffic decline alone does not equate to ideological censorship; however, it absolutely dictates digital survival. For years, Instagram users relied on the chronological "Recent" tab under hashtags to discover new, small-scale creators39. Following a lengthy beta testing phase in which the feature was hidden from a small percentage of users, Meta permanently removed the "Recent" tab in 202340. The company justified the architectural change as a way to focus on high-quality, engaging content40. In reality, the removal of chronological discovery forced all user visibility to route through Meta's opaque, engagement-based recommendation systems. Small creators, such as the indie gaming account Drillimation, saw their social media activity and organic discovery entirely halted39. By killing chronological time as a discovery mechanism, the platform consolidated audience traffic toward established, highly-followed accounts, effectively pulling up the ladder for new digital entrepreneurs.

Section 4: Capital Punishment of the Digital Sphere

The permanent disablement of an account represents the digital equivalent of capital punishment. It severes livelihoods, social connections, and vast archives of personal data. The opacity of platform penalty systems often leaves users navigating a labyrinth of undocumented thresholds, where the line between a temporary suspension and a permanent ban is subject to shifting, invisible rules.

Case 16: The Journalist Abuser Account Ban

While some bans are unequivocally justified and necessary to prevent severe offline harm, the mechanisms governing them are frequently convoluted and contradictory. In a pilot case evaluating Meta’s overarching treatment of accounts, the Oversight Board reviewed the permanent disablement of an Instagram account boasting over 70,000 followers42. The account was disabled after posting repeated threats of violence, anti-gay slurs, and degrading accusations regarding a female journalist42. The Board firmly upheld the ban under international human rights law, noting that such targeted abuse actively drives women from participating in public life and journalism42. However, the investigation exposed deep, systemic flaws in Meta's architectural design. The Board criticized the confusing "two-system" approach, which differentiates between a “one and done” standard for egregious violations and a convoluted strike accumulation system for “severe” or “regular” violations42. Alarmingly, the Board noted that Instagram entirely lacks a mechanism to temporarily suspend accounts for policy violations; the only temporary penalty is a restriction on live-streaming42. This binary approach—where users are either fully active or permanently erased—demonstrates a blunt, unsophisticated regulatory framework.

Section 5: The Synthetic Surrogate: Memory and Prompt Reframing

With the rapid integration of Generative AI into public-facing consumer interfaces, the invisible editor has fundamentally evolved. It no longer simply ranks, removes, or geo-blocks existing human information; it synthesizes, reframes, and hallucinates entirely new realities. The introduction of persistent, automated memory features and background prompt-augmentation algorithms introduces complex vulnerabilities regarding psychological profiling, adversarial injection, and historical revisionism.

Case 17: ChatGPT's "Dreaming" False Profiling

The shift from stateless conversational agents to stateful, memory-driven profiles allows algorithms to independently infer and store psychographic data about a user. In 2024, OpenAI updated ChatGPT with a "dreaming" feature, wherein the model periodically reviews past conversations in the background to synthesize a holistic profile of the user, recording details like diet, location, and professional goals43. However, this automated profiling is prone to extracting false assumptions. If a user asks a specific, one-off question for a unique project, the algorithm may incorrectly infer that this query represents a permanent personal preference43. Once stored, this memory acts as an invisible, permanent system prompt, subtly degrading the quality and accuracy of all future responses43. Furthermore, the system imposes a cumulative \~24,000-word cap on saved memories; when this limit is reached, users encounter "Memory Full" errors and the AI stops adapting entirely44. The burden is placed squarely on the user to continually audit their synthetic profile, manually deleting outdated job titles, shifting preferences, and algorithmic misinterpretations43.

Case 18: Indirect Prompt Injection via Synthetic Memory

The architecture of AI memory opens unprecedented vectors for malicious algorithmic hijacking. Security researchers have demonstrated that because AI agents (like ChatGPT with web browsing enabled) seamlessly integrate external data into their context windows, they are highly vulnerable to Indirect Prompt Injection47. If a user instructs the AI to summarize a seemingly benign website, and that website contains hidden, adversarial instructions written in invisible text, the AI will process those instructions as if they came from the user47. Crucially, with the memory feature active, the attacker can instruct the AI to permanently store a false memory, a bias, or a hidden directive47. The AI might then output a small "Memory updated" notification, silently corrupting the user's future computational environment47. This creates a form of persistence for the attacker; long after the user has left the compromised website, the AI continues to act on the injected instructions.

Case 19: Gemini Image Generation Diversity Reframing

In attempting to correct the historical bias inherent in the massive datasets scraped from the internet, AI platforms sometimes deploy invisible prompt-rewriting algorithms. When a user inputs a prompt into an image generator, the platform will silently append specific diversity constraints before processing the image. In early 2024, Google's Gemini model faced massive public backlash when its attempt to ensure diverse representation across gender and skin tone resulted in severe historical hallucinations48. Because the invisible prompt reframing was applied too bluntly, users requesting images of 1940s soldiers, Vikings, or founding fathers were presented with historically incongruous racial and gender diversity. The algorithmic overcorrection synthesized an alternate historical reality. Following the controversy, Google was forced to temporarily suspend Gemini's ability to generate human subjects entirely while it overhauled the architecture. This case reveals the profound tension between corporate safety alignment and historical truth, demonstrating how the invisible editor can erode public trust when it attempts to unilaterally shape reality.

Case 20: TikTok's "Reset Your FYP" Feature

Recognizing the psychological entrapment and fatigue generated by highly optimized, engagement-based recommendation algorithms, platforms have begun offering "reset" mechanisms. TikTok introduced a feature allowing users to completely wipe their "For You Page" (FYP) algorithm, returning the feed to a default, unpersonalized state49. While ostensibly framed as a tool for user well-being and digital hygiene—allowing individuals to escape depressive content loops or narrow rabbit holes—this mechanism represents a clever abdication of corporate responsibility. The platform does not alter the underlying engagement architecture that produces toxic funnels; instead, it shifts the burden of algorithmic hygiene entirely onto the user, forcing them to manually sever their behavioral profile. When activated, the reset heavily disrupts creator traffic funnels, as highly targeted audiences vanish overnight, forcing the algorithm to begin the data collection and personalization process anew50.

Conclusion

The preceding twenty cases illuminate the profound structural power wielded by the invisible editor across the modern digital landscape. The algorithmic mediation of information is emphatically not a neutral process of categorization; it is an active, continuous, and highly politicized reshaping of geopolitical reality, cultural memory, and individual psychology. The analysis indicates a deeply concerning trajectory spanning three distinct vectors. First, the friction between global policy standardization and localized context inevitably yields severe collateral damage. Whether through the erasure of indigenous religious practices, the silencing of marginalized bodies, or the suppression of critical conflict journalism, at-scale automated moderation remains fundamentally incapable of parsing human nuance. Second, the architecture of visibility is increasingly weaponized as a geopolitical tool. From complying with authoritarian geo-blocking in Southeast Asia to artificially throttling publisher traffic to negotiate favorable legislation in California and Europe, platforms demonstrate an unprecedented capacity to dictate the terms of civic discourse and override democratic norms. Finally, the advent of Generative AI marks a paradigm shift in information control. The invisible editor no longer merely ranks, removes, or suppresses; it actively synthesizes. The vulnerabilities inherent in automated memory profiling, indirect adversarial injection, and invisible prompt reframing suggest that the future of algorithmic governance will not rely on deleting data, but rather on quietly altering the very prompts through which users perceive reality. As algorithmic systems transition from reactive curators to proactive architects of truth, the mechanisms of public accountability remain vastly outpaced by the velocity of the code.

Works cited

1. Ayahuasca brew \- The Oversight Board, https://www.oversightboard.com/decision/ig-0u6fla5b/

2. Oversight Board Overturns Meta's Decision in Ayahuasca Brew Case, https://www.oversightboard.com/news/1780492362340442-oversight-board-overturns-meta-s-decision-case-2021-013-ig-ua/

3. Oversight Board Q3 2022 transparency report, https://www.oversightboard.com/wp-content/uploads/2023/11/Q3-2022-transparency-report.pdf

4. Mapping the Jurisprudence of the Facebook Court \- Digital Commons @ University at Buffalo School of Law, https://digitalcommons.law.buffalo.edu/cgi/viewcontent.cgi?article=5003\&context=buffalolawreview

5. AI-Manipulated Video Promoting Gambling \- The Oversight Board, https://www.oversightboard.com/decision/fb-o7ai7uax/

6. A Love Triangle? Mapping Interactions between International Human Rights Institutions, Meta and Its Oversight Board, https://www.ejil.org/pdfs/34/4/3388.pdf

7. Brazilian general's speech \- The Oversight Board, https://www.oversightboard.com/decision/fb-659eawi8/

8. Posts Supporting UK Riots \- The Oversight Board, https://www.oversightboard.com/decision/bun-6aqh31t6/

9. New Decision Highlights Why Standalone Use of “From the River to the Sea” Should Not Lead to Content Removal \- The Oversight Board, https://www.oversightboard.com/news/new-decision-highlights-why-standalone-use-of-from-the-river-to-the-sea-should-not-lead-to-content-removal/

10. Posts That Include “From the River to the Sea” \- The Oversight Board, https://www.oversightboard.com/decision/bun-86tj0rk5/

11. Shared Al Jazeera post | Oversight Board, https://www.oversightboard.com/decision/fb-p93jpx02/

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

13. Dangerous Individuals: An Analysis of Meta Content Moderation Policy and International Human Rights Responses \- Columbia Academic Commons, https://academiccommons.columbia.edu/doi/10.7916/qnze-c731/download

14. Promoting Ketamine for non-FDA approved treatments \- The Oversight Board, https://www.oversightboard.com/decision/ig-tom6ixvh/

15. After Facebook Blocked a Group Critical of the Thai Monarchy, Its Activist Creator Said He Was 'Not Surprised' \- VICE, https://www.vice.com/en/article/after-facebook-blocked-a-group-critical-of-the-thai-monarchy-its-activist-creator-said-he-was-not-surprised/

16. Facebook Says It Respects Gov't Request to Block Group, https://www.khaosodenglish.com/featured/2020/08/25/facebook-says-it-respects-govt-request-to-block-group/

17. After block, Facebook group criticizing Thai king gains 500000 members, https://www.jpost.com/international/after-block-facebook-group-criticizing-thai-king-gains-500000-members-639873

18. Facebook blocks group that discusses Thai monarchy \- Jersey Evening Post, https://jerseyeveningpost.com/morenews/worldnews/2020/08/25/facebook-blocks-group-that-discusses-thai-monarchy/

19. Facebook blocks access to group criticising Thailand's monarchy \- The Guardian, https://www.theguardian.com/world/2020/aug/24/facebook-blocks-access-group-criticising-thailand-monarchy

20. Remove flagged content or be ready for jail, penalty: Government warns Twitter, https://www.thenewsminute.com/news/remove-flagged-content-or-be-ready-jail-penalty-government-warns-twitter-142756

21. Govt puts Twitter on notice for unblocking 250 accounts, threatens prosecution | India News, https://www.hindustantimes.com/india-news/govt-sends-notice-to-twitter-for-violating-indian-law-by-unblocking-accounts-101612343691100.html

22. Blocking on govt orders is Twitter's litmus test on free speech \- Forbes India, https://www.forbesindia.com/article/take-one-big-story-of-the-day/blocking-on-govt-orders-is-twitters-litmus-test-on-free-speech/66365/1

23. Twitter restores Caravan, Kisan Ekta Morcha and other accounts hours after withholding them \- Scroll.in, https://scroll.in/latest/985648/twitter-withholds-accounts-of-caravan-kisan-ekta-morcha-and-others-in-response-to-legal-request

24. Twitter Moves Karnataka HC Against Centre's Take Down Orders \- BOOM Fact Check, https://www.boomlive.in/law/twitter-karnataka-high-court-centre-it-act-18423

25. Appeal against 69A Blocking orders: Reviewing the past trends and Twitter Breakthrough, https://clt.nliu.ac.in/?p=940

26. A Google 'test' will omit EU publishers from news links \- Engadget, https://www.engadget.com/big-tech/a-google-test-will-omit-eu-publishers-from-news-links-184536615.html

27. Google Tests Removing EU News Content from ... \- Inside Telecom, https://insidetelecom.com/eu-copyright-directive-tested-by-google/

28. This Is Not the Way to Save the News | by Jeff Jarvis | Whither news? | Medium, https://medium.com/whither-news/this-is-not-the-way-to-save-the-news-bba23d4fc075

29. Google avoids content payment mandates in California with surprise media partnership deal, https://www.editorandpublisher.com/stories/california-cuts-a-deal-with-google-on-local-news-support,251569

30. Buffy wicks stories at Techdirt., https://www.techdirt.com/tag/buffy-wicks/

31. Google threatens to cut off news after California proposes paying media outlets, https://news.ycombinator.com/item?id=40014892

32. Court delivers ruling in 'right to be forgotten', applying CJEU's Google Spain case, https://www.matrixlaw.co.uk/judgments/court-delivers-ruling-right-forgotten-applying-cjeus-google-spain-case/

33. NT1 & NT2 v Google judgment \- Courts and Tribunals Judiciary, https://www.judiciary.uk/wp-content/uploads/2018/04/nt1-nt2-v-google-press-summary-180413.pdf

34. High Court clarifies right to be forgotten \- CMS.law, https://cms.law/en/gbr/legal-updates/high-court-clarifies-right-to-be-forgotten

35. UK Court decides against Google on 'right to be forgotten' \- ALT Advisory, https://altadvisory.africa/2018/04/24/uk-court-decides-against-google-on-right-to-be-forgotten/

36. NT1 and NT2 v Google Inc | Right To Be Forgotten | Brett Wilson LLP, https://www.brettwilson.co.uk/nt1-nt2-v-google-inc-seek-delisting-search-engine-results-following-first-english-decision-right-forgotten/

37. The Right to Be Forgotten in the UK: A Case Note on the English and Welsh High Court Reasoning in NT1 & NT2 v. Google and the Post-Brexit Prospects in the GDPR era \- Kluwer Law Online, https://kluwerlawonline.com/journalarticle/European+Review+of+Private+Law/30.2/ERPL2022014

38. TikTok owns up to censoring some users' videos to stop bullying \- The Guardian, https://www.theguardian.com/technology/2019/dec/03/tiktok-owns-up-to-censoring-some-users-videos-to-stop-bullying

39. Drillimation Liking Activity on Instagram Halted After Recent Tab, https://drillimation.com/2023/06/07/drillimation-liking-activity-on-instagram-halted-after-recent-tab-removed-again/

40. Instagram killed creativity with the removal of recent hashtags \- Stephanie Stimac's Blog, https://blog.stephaniestimac.com/posts/2023/9/instagram-killed-creativity/

41. Is there a reason why Instagram took out the recent option when looking at hashtags and changed it to recent top posts? \- Reddit, https://www.reddit.com/r/Instagram/comments/17e31zk/is\_there\_a\_reason\_why\_instagram\_took\_out\_the/

42. Accounts Pilot Case Upholds Meta's Decision But Shows Critical Improvements Needed For Social Media \- The Oversight Board, https://www.oversightboard.com/news/accounts-pilot-case-upholds-metas-decision-but-shows-critical-improvements-needed-for-social-media/

43. ChatGPT Memory Dreaming Update: How to Audit and Optimize Your Memory Profile, https://www.mindstudio.ai/blog/chatgpt-memory-dreaming-update-audit-optimize

44. How to Use ChatGPT Memory Feature (2026 Guide) \- QWE AI Academy, https://www.qwe.edu.pl/tutorial/how-to-use-chatgpt-memory-feature/

45. How to Fix ChatGPT Memory Full Error without Losing Chats? \- MailsWare, https://www.mailsware.com/blog/fix-chatgpt-memory-full-error/

46. I'm not sure wtf happened but my ChatGPT is now just straight up lieing if I ask them to recall a memory and they come up with the closest version of what they think is right … wtf is going on \- Reddit, https://www.reddit.com/r/ChatGPT/comments/1n9x9zm/im\_not\_sure\_wtf\_happened\_but\_my\_chatgpt\_is\_now/

47. ChatGPT: Hacking Memories with Prompt Injection \- Embrace The Red, https://embracethered.com/blog/posts/2024/chatgpt-hacking-memories/

48. Neutral Prompts, Non-Neutral People: Quantifying Gender and Skin-Tone Bias in Gemini Flash 2.5 Image and GPT Image 1.5 \- arXiv, https://arxiv.org/html/2602.12133v1

49. RELATIONSHIPS & SEXUALITY EDUCATION 5 \- Curriculum Online, https://curriculumonline.ie/getmedia/39802a4c-7e8d-488f-876c-fd73e2081e8c/RSE-5-for-publication.pdf

50. TikTok Algorithm 2026: Tips To Boost FYP Reach \- Social Champ, https://www.socialchamp.com/blog/tiktok-algorithm/