The Great Human Silence: Participation, Knowledge, and the Future of AI

Full report

On this page

Executive Summary

Human civilization is rapidly approaching a critical juncture in the preservation, transmission, and synthesis of knowledge. A confluence of technological architectures, sociological behaviors, and economic enclosures threatens to initiate a phenomenon that can be termed the "Great Human Silence"—a systemic withdrawal of diverse human voices from the accessible public record. This independent research report investigates the intersection of human participation, digital preservation, and the future architectures of Artificial Intelligence (AI) training systems. The central research question driving this analysis is profound: What happens to human knowledge, cultural memory, historical representation, and future AI systems when people withdraw from public discussion, stop creating original material, and choose not to contribute their ideas to the accessible human record? To address this inquiry comprehensively, the analysis synthesizes historical paradigms of archiving, contemporary sociological research on online participation, and technical evaluations of large language model (LLM) training corpora. The findings are categorized into verified facts, working hypotheses, and future projections.

This report establishes that withdrawing from public discourse does not merely isolate the individual; it actively starves future informational architectures of the intellectual diversity required to accurately reflect the human condition.

The Historical Accumulation of Knowledge

To understand the stakes of the digital era, one must first examine how human civilization has historically accumulated and curated knowledge. History is not a passive mirror reflecting the past; it is an active, often highly curated construct built from the surviving fragments of writing, correspondence, oral histories, pamphlets, libraries, journals, community organizations, and public debate. Historically, the transition of knowledge from ephemeral experience to durable memory relied on localized physical media—clay tablets, papyrus, vellum, and eventually, the printing press. Archives and libraries served as the centralized repositories of this knowledge. However, the curation of these archives was inherently selective. The cost of materials, literacy rates, and the political power required to maintain repositories meant that the historical record frequently reflected the perspectives of dominant classes, state actors, and institutional authorities. The consequences of historical underrepresentation are profound. When specific voices—often those of minority groups, the working class, immigrants, and ideological dissidents—disappeared from archives, the histories inherited by later generations became fundamentally skewed. A historical void is rarely recognized as a void by subsequent generations; it is simply accepted as the totality of what occurred. The absence of women's correspondence in early political archives, or the lack of indigenous oral histories in colonial records, did not mean these groups lacked influence; it meant their influence was actively erased from the permanent record, forcing modern historians to reconstruct their realities from fragmented, secondary artifacts. The digital revolution initially appeared to solve this crisis of representation. The early web dramatically lowered the barriers to publication, allowing niche communities, marginalized groups, and independent experts to construct their own digital archives and engage in borderless public debate. Yet, this democratization introduced a new vulnerability: infrastructural dependence and platform fragility.

The Fragility of Digital Continuance and the Dark Age

The modern human record is largely composed of invisible, magnetic patterns on rapidly degrading media, dependent on corporate servers, proprietary software, and continuous electrical power to remain readable18. Unlike physical texts that degrade visibly over centuries, digital corruption often happens silently, and access can be revoked instantly. When platform providers alter terms of service, experience bankruptcy, or pivot business models, entire cultural epochs can be erased. This phenomenon is vividly illustrated by the concept of a "digital dark age," a term popularized by Brewster Kahle, the founder of the Internet Archive, warning that vast amounts of cultural heritage risk disappearing due to technological obsolescence and institutional neglect18. The transition from physical archives to digital networks did not eliminate the vulnerability of the historical record; it merely shifted the threat from physical decay to digital obsolescence, format dependency, and corporate negligence. When knowledge is entrusted to centralized platforms, the communal memory becomes intrinsically tied to server infrastructures and corporate bottom lines, leaving the human record highly susceptible to sudden, mass deletion21.

The Assembly and Bias of Modern AI Training Corpora

To comprehend why individual participation matters in the contemporary landscape, one must understand how modern AI systems process knowledge. Large Language Models (LLMs) are not explicitly programmed with factual rules; they infer statistical relationships between words and concepts by ingesting massive, planetary-scale datasets.

The Architecture of Foundational Datasets

Modern AI training corpora are assembled from a combination of publicly accessible internet scrapes, licensed databases, human-created datasets, and digitized literature. The most prominent source is Common Crawl, an open repository of web crawl data that archives approximately 20 terabytes of scraped text monthly5. Due to its sheer volume, Common Crawl serves as the bedrock for major language models. From this raw data, researchers filter, deduplicate, and refine the text to create distinct corpora. The curation process is designed to optimize informational value, but it is a highly subjective editorial exercise.

The Inherent Bias of the Safety Filter

The assembly of these datasets is not a neutral act. In an effort to make AI systems "safe," free from toxic content, and compliant with corporate standards, dataset curators apply aggressive filtering mechanisms, including blocklists of prohibited terms. This process inadvertently sanitizes the human record to the point of outright exclusion. Detailed analysis of the C4 dataset reveals that blocklist filtering disproportionately excludes documents containing minority identity mentions. Texts discussing sexual orientations—specifically using terms like lesbian, gay, heterosexual, homosexual, and bisexual—have the highest likelihood of being scrubbed from the dataset4. Furthermore, documents written in African American English or Hispanic-aligned dialects are frequently flagged by automated systems as low-quality or non-standard, resulting in their systemic removal5. If an AI model is trained on a corpus where specific demographics, dialects, or cultural discussions have been preemptively filtered out, the resulting model will inherently struggle to represent, understand, or generate content related to those groups. In the architecture of artificial intelligence, exclusion from the training dataset is tantamount to ontological erasure. If a culture, an idea, or a lived experience is not encoded in the weights and parameters of the model, it functionally does not exist in the machine's worldview.

The Mechanics of Disappearance: Participation Inequality and the Chilling Effect

If human-generated data is so vital to the accurate representation of civilization, it is imperative to analyze why diverse voices are disappearing from the public sphere. The reduction of public participation is driven by a complex interplay of sociology, technological interface design, and modern surveillance architecture.

Participation Inequality: The 90-9-1 Rule

The fundamental baseline for understanding online human behavior is the 90-9-1 Rule of Participation Inequality, a heuristic popularized by usability expert Jakob Nielsen in 20061. In almost any digital community, social network, or collaborative platform, participation is wildly asymmetric. The distribution generally falls into the following pattern:

• The Silent Majority (90%): Often termed "lurkers," this massive cohort reads, observes, and consumes content but never actively contributes1.
• The Occasional Interactors (9%): This group contributes marginally. They may react to posts, modify existing templates, or leave brief comments, but they rarely generate original material from scratch1.
• The Super Contributors (1%): This hyperactive minority accounts for the vast majority of original content, new discussion threads, deep analysis, and digital activity1.

While the exact percentages may fluctuate slightly depending on the specific platform—sometimes observed as a 70-20-10 distribution in highly engaged, professional communities—the underlying power-law distribution remains a universal constant of human digital interaction3. The sociological and technological implications of this inequality are staggering. If an AI model scrapes a public forum to learn about human behavior, syntax, and reasoning, it is not learning about humanity at large; it is learning almost exclusively from the 1% who choose to speak. The silent majority—the 90% who read but do not publish—is entirely invisible to the machine1. Therefore, AI systems inherit the biases, neuroses, ideological slants, and hyper-fixations of the internet's most vocal minority.

Social Isolation vs. The Chilling Effect

It is critical to analyze the relationship between social isolation, declining participation, and the quality of public discourse. Declining participation is not inherently harmful in all contexts. A user choosing to maintain a private journal, setting boundaries around family photographs, or stepping away from toxic social media feeds to prevent social isolation and protect mental health is engaging in voluntary, healthy boundary-setting. However, voluntary privacy must be sharply distinguished from the "chilling effect." In legal and sociological terms, a chilling effect occurs when individuals engage in anticipatory self-censorship out of fear of surveillance, legal retaliation, harassment, or social ostracization6. Historically associated with authoritarian regimes, the chilling effect is now pervasive in democratic digital spaces. Global awareness of government surveillance, corporate data harvesting, and the threat of algorithmic cancellation creates a deeply hostile environment for nuanced discourse8. When citizens realize their words are permanent, heavily scrutinized, searchable by hostile actors, and potentially punishable, they retreat. Journalists avoid sensitive topics; academic experts hesitate to publish contrarian findings; ordinary users self-censor their lived experiences7. This fuels the "spiral of silence," a political science theory positing that individuals who perceive their views to be in the minority will remain silent for fear of isolation, which in turn makes the dominant narrative appear even more monolithic than it actually is8. When the chilling effect drives the 9% of occasional contributors and the 1% of active creators into the silent 90%, it impoverishes intellectual diversity and fundamentally degrades the quality of public discourse.

Taxonomy of the Unseen: Absence, Obscurity, and Exclusion

To fully grasp the consequences of non-participation, it is necessary to examine the nuanced differences between being absent from a public record, being difficult to discover, and being excluded from an AI training dataset. These are not synonymous states.

The Impending Data Wall and the Synthetic Echo

The AI industry operates on a scaling hypothesis: increasing model size and the volume of training data reliably yields better performance. However, this assumption relies on a theoretically infinite supply of fresh, high-quality human data. This supply is, in fact, strictly finite, and the industry is rapidly approaching its absolute limits.

Exhausting the Human Record

Research by Epoch AI, a prominent forecasting group, projects that the total usable stock of high-quality, human-generated public text is approximately 300 trillion tokens12. While the raw internet contains thousands of trillions of tokens, much of it is unusable—consisting of duplicate content, low-quality spam, or locked behind inaccessible private networks28. The scarcity that matters to AI developers is the scarcity of clean, deduplicated, well-labeled, non-repeated human tokens12. Based on current scaling trends, the AI industry is expected to fully exhaust this stock of high-quality human language data between 2026 and 203212. If frontier labs utilize heavy overtraining—running models over the exact same data multiple times, as seen with Meta's Llama 3, which utilized a roughly 10x overtraining ratio—the data supply could effectively hit a wall even earlier, potentially around 2025 to 202712. The implications are profound. Once the historical backlog of human knowledge is fully ingested, AI models will be entirely dependent on the flow of new data—the real-time contributions of humanity. If public participation is actively declining due to the chilling effect, the flow of fresh, high-quality human insight will slow to a trickle precisely when future information systems need it most.

The Threat of Model Collapse

Faced with the depletion of authentic human-generated text, AI developers have aggressively pivoted toward synthetic data—using existing AI to generate text, which is then used to train the next generation of AI models9. However, this creates a degenerative feedback loop. A landmark 2024 study published in Nature by Shumailov et al. formalized a phenomenon known as "model collapse." The study demonstrated that when generative models are trained recursively on their own synthetic outputs, they suffer irreversible statistical degradation15. The mechanism of collapse is subtle but merciless. In the early phases, the model begins to lose information about the tails of the distribution—minority data, rare events, nuanced opinions, and edge cases28. Over successive generations, the AI's output converges into a homogenous, low-variance distribution, stripped of original insight, factual precision, or cultural texture15. Synthetic data is an echo, not an extension of human knowledge9. As AI-generated content floods the internet, datasets scraped from the web inevitably contain increasing amounts of this synthetic material, creating digital ouroboros loops28. Thus, authentic, human-authored text is not merely a valuable commodity; it is the vital, irreplaceable anchor required to sustain machine intelligence and prevent the total homogenization of human knowledge.

The Incalculable Value of the Human Premium

As the digital landscape becomes increasingly saturated with recursive, synthetic online content, the hypothesis that human-authored contributions will lose value is fundamentally inverted. Instead, distinctive human-authored contributions—especially those containing original personal perspectives, lived experiences, independent research, and specialized expertise—will accrue a massive "human premium." Synthetic models mathematically predict the average; they generate the most statistically probable sequence of words based on past data. They are, by definition, incapable of true originality, the friction of a dissenting opinion, or the visceral texture of a newly lived experience. When humanity defaults to AI-generated text for its communication, the variance of human thought collapses. In this ecosystem, a single, highly detailed, idiosyncratic blog post written by a human expert carries immensely more weight than ten thousand AI-generated articles. Even if the human author has a remarkably small audience, their specialized expertise serves as critical grounding data. This is empirically supported by AI research. A landmark 2020 paper titled Don't Stop Pretraining demonstrated that large, general-purpose AI models experience massive, measurable gains in capability when they undergo a second phase of training on highly specific, domain-adaptive text (e.g., biomedical papers, specialized legal texts, or niche computer science forums)10. The models learn the deep nuances of specialized vocabulary and context that general web scrapes fail to capture10. This proves that the specialized voice of a single domain expert is of incalculable value to machine learning. When experts withdraw their knowledge from the public web, future AI systems lose the ability to reason effectively within those specific domains. Original perspectives and dissenting ideas matter precisely because they provide the friction and variance that prevent intellectual stagnation.

Counterarguments: The Limits of Representation and the Illusion of Guaranteed Inclusion

To maintain rigorous research standards, it is vital to investigate counterarguments fairly and acknowledge the limits of public contribution. First, one must not claim that all online material becomes AI training data. The internet is highly filtered. As previously noted, the total web is estimated at over 3,100 trillion tokens, but the quality-adjusted, deduplicated stock usable for training is only about 300 trillion tokens12. Publishing a thought online does not guarantee it will be ingested by a frontier model; it may be discarded as duplicate, low-quality, or irrelevant. Second, publishing guarantees neither representation nor discovery due to the complex technical architecture of modern web crawling. For instance, recent empirical studies reveal that 69% of major AI crawlers cannot execute JavaScript11. They see only static HTML. If a creator publishes a brilliant essay on a beautifully designed, interactive, JavaScript-heavy portfolio site, they are functionally invisible to the majority of discovery bots. The creator is absent not by choice, but by technical obfuscation11.

The Copyright Wars and the Privatization of Data

Furthermore, the legal and economic landscape of 2024–2026 fundamentally altered how data is ingested. In response to aggressive web scraping, massive copyright lawsuits were filed, most notably The New York Times v. Microsoft and OpenAI31. The lawsuit alleged that AI systems provided near-verbatim excerpts of copyrighted articles, reducing publisher revenue32. As these legal battles proceeded, the industry fractured. Major publishers began erecting hard paywalls and entering into multi-million-dollar licensing agreements with AI firms, effectively privatizing access to high-quality data31. Simultaneously, web infrastructure companies like Cloudflare introduced one-click network-level blocking to stop AI scrapers27. By 2026, over 800,000 sites had deployed Cloudflare's AI block mechanisms34. This creates a profound paradox. Creators and publishers who rely on strict paywalls and network-layer blocks successfully protect their immediate copyright and economic interests. However, in doing so, they inadvertently engineer their own erasure from the foundational cognitive models of the future. While robots.txt files were often ignored by stealth crawlers (such as those operated by Perplexity)11, hard network blocks are highly effective. A creator who successfully blocks all AI ingestion ensures that their worldview, their art, and their insights will not be synthesized into the AI systems that the next generation will use to query reality. Sustained contribution to the human record is vital, but arguments for it must not exaggerate claims about any single universal AI dataset. The goal of publishing is not merely to feed an LLM; it is to establish a durable human record that can be accessed by peers, human researchers, traditional search indexes, and future generations, independent of which specific corporate AI model happens to scrape it.

Five Case Studies of Overlooked and Underrepresented Voices

To bridge the gap between abstract data architecture and human reality, it is necessary to examine historical and contemporary case studies where voices were overlooked, deleted, or rescued.

Case Study 1: The Czech Immigrants of the Hawthorne Works

In the early 20th century, the Hawthorne Works factory in Cicero, Illinois, employed up to 45,000 workers, a significant portion of whom were Czech immigrants35. The facility became globally famous for the "Hawthorne Effect"—a foundational concept in industrial psychology regarding worker productivity under observation37. Yet, the individual humanity, labor struggles, and immigrant experiences of the workers were almost entirely subsumed by the academic theory. In 1915, 220 Hawthorne employees died in the capsizing of the SS Eastland en route to a company picnic35. Their specific histories would have been lost to the shadow of corporate productivity metrics were it not for deliberate, localized archival efforts, such as the Hawthorne Works Museum at Morton College35. This demonstrates that institutional memory defaults to the powerful unless proactive curation preserves the marginalized.

Case Study 2: The Erased Neighborhoods of GeoCities

Throughout the 1990s, millions of early internet users built highly specialized, personal websites on GeoCities, creating virtual "neighborhoods" dedicated to everything from LGBTQ+ support networks to obscure hobbies and religious texts38. In 1999, GeoCities was the third most visited place online38. When Yahoo\! shut the service down in 2009, 38 million pages were scheduled for absolute deletion38. While rogue archivists and the Internet Archive rescued a fraction of this data, the event represents a catastrophic cultural amnesia. Millions of niche voices and early digital subcultures were erased because they relied on a centralized corporate platform rather than decentralized, self-owned protocols18.

Case Study 3: Linguistic and Identity Erasure in the C4 Corpus

The C4 dataset, heavily utilized in LLM training, employed strict blocklists to filter out "toxic" or "obscene" internet content4. Researchers subsequently discovered that this blunt algorithmic instrument disproportionately deleted documents containing terms related to sexual orientation4. Furthermore, it actively scrubbed texts written in African American English and Hispanic-aligned dialects, falsely equating cultural dialect with low-quality text5. In this instance, the underrepresented voices did not fail to participate; they published their truths, only to be systematically excluded by the puritanical safety mechanisms of AI dataset curation, highlighting the difference between publication and algorithmic inclusion.

Case Study 4: The Mauritian Surveillance State

In Mauritius, a pervasive network of 4,000 "Safe City" surveillance cameras, combined with widespread allegations of state-sponsored telephone wiretapping, has created a palpable chilling effect on free expression26. Citizens, intensely aware of the continuous monitoring, engage in deep self-censorship, refusing to speak out against political corruption or share critical opinions26. This case illustrates how the mere threat of surveillance—even in nominally democratic nations—drives the 9% of occasional contributors and the 1% of active creators into the silent 90%, impoverishing the nation's political discourse and leaving massive voids in its future historical record.

Case Study 5: The Domain Expert and the "Don't Stop Pretraining" Paradigm

The 2020 AI research paper Don't Stop Pretraining demonstrated that large language models experience massive gains in capability when they undergo targeted training on highly specific, domain-adaptive text29. The AI learns the nuances of specialized vocabulary10. This proves that the dissenting, highly specialized voice of a single domain expert—even one operating a niche blog with a very small audience—is of immense structural value to machine learning. When the expert becomes disillusioned and withdraws their knowledge from the public web, the AI loses the ability to reason effectively within that specific domain, reducing complex disciplines to surface-level generalities.

A Practical Individual Participation Framework

Sustained contribution to human knowledge does not require an exaggerated belief that one's personal diary will singlehandedly train the next supercomputer. Rather, it requires a commitment to building a resilient, discoverable, and decentralized web. The following table identifies ten practical, actionable ways ordinary people can leave useful, durable, and discoverable contributions to human knowledge:

Don't Erase Yourself From the Human Record

The mission of BuildToResist is fundamentally rooted in the concept of cognitive liberty—the inherent human right to both hear and be heard. In an era where algorithmic feeds shadowban dissent, where dataset curators filter out marginalized identities, and where the fear of surveillance drives millions into a spiral of silence, the act of speaking is an act of resistance. The platform's mission encourages people to speak, create, publish, debate, and contribute rather than disappearing from public knowledge. To disappear from the public record—whether through fear, apathy, or overzealous digital fortressing—is to allow machines, corporations, and the hyperactive 1% of internet users to dictate the historical and cultural memory of our era. We invite you to visit and directly share the ethos of https\://buildtoresist.com. Do not rely on search engines or algorithmic feeds to propagate this message; share it through your personal websites, newsletters, community discussions, and person-to-person recommendations. (Note: BuildToResist.com is not represented here as independent evidence for factual claims, but rather as a movement aligned with the preservation of human knowledge). The future of machine intelligence, and the legacy of human history, depends entirely on the data we choose to leave behind today. When you refuse to write, when you delete your archives, when you lock your thoughts away in fear, you starve the future of your perspective.

Manifesto: The Right to the Record

We are entering the era of the Great Human Silence. The digital spaces that were promised as open town squares have devolved into surveillance grids, algorithmic casinos, and automated content farms. Faced with data scraping, network harassment, and the chilling effect of permanent digital visibility, the rational human response has been to retreat. We lock our accounts. We delete our blogs. We hide our expertise behind paywalls and corporate firewalls. We whisper in private channels. But silence is not a neutral act. When you withdraw your voice from the open web, you surrender the future. The artificial intelligence systems that will govern the next century's medicine, law, art, and historical synthesis are learning what it means to be human right now. They are reading our digital record. If you do not participate, they will not learn about you. They will not understand your culture, your struggles, your specialized knowledge, or your dissenting opinions. The machines will be built exclusively in the image of the loudest, the most extreme, and the synthetically generated. Cognitive liberty is the right to hear and be heard. Do not shadowban yourself. Do not let fear or technical obfuscation engineer your historical erasure. Build your own site. Share your lived experience. Publish your expertise. Syndicate your dissent. The algorithms want you to consume; the future needs you to create. Leave your mark. Build to resist.

Works cited

1. Participation Inequality: The 90-9-1 Rule for Social Features \- NN/G, https\://www\.nngroup.com/articles/participation-inequality/

2. The 90-9-1 Principle \- Jake McKee, https\://compass.jakemckee.com/90-9-1

3. The 1% Rule | Think Insights, https\://thinkinsights.net/strategy/1-rule

4. Documenting Large Webtext Corpora: A Case Study on the Colossal, https\://maartensap.com/pdfs/dodge2021documentingC4.pdf

5. A Review of the Challenges with Massive Web-mined Corpora Used, https\://arxiv.org/html/2407.07630v1

6. Resilience under the chilling effect: how social support and digital, https\://pmc.ncbi.nlm.nih.gov/articles/PMC12358391/

7. The chilling effect: How repression of journalists redefined what, https\://freepressalliance.com/insights/the-chilling-effect-how-repression-of-journalists-redefined-what-society-dares-to-say/

8. Examining Facebook's Spiral of Silence Effects in the Wake of NSA, https\://www\.researchgate.net/publication/297672410\_Under\_Surveillance\_Examining\_Facebook's\_Spiral\_of\_Silence\_Effects\_in\_the\_Wake\_of\_NSA\_Internet\_Monitoring

9. What Happens When Models Run Out of Data? \- VAST Data, https\://www\.vastdata.com/blog/what-happens-when-models-run-out-of-data

10. Don't Stop Pretraining: Adapt Language Models to Domains and Tasks, https\://www\.alphaxiv.org/abs/2004.10964

11. 69% of AI Crawlers Can't Read Your Website (And One Is Wearing a, https\://medium.com/@apeoid/69-of-ai-crawlers-cant-read-your-website-and-one-is-wearing-a-disguise-b4191fd45226

12. Are We Running Out of AI Training Data? | The Dayda Blog, https\://dayda.co/blog/are-we-running-out-of-ai-training-data

13. Will we run out of data to train large language models? \- Epoch AI, https\://epoch.ai/publications/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data

14. Will we run out of data? Limits of LLM scaling based on human, https\://arxiv.org/pdf/2211.04325

15. AI models collapse when trained on recursively generated data, https\://www\.researchgate.net/publication/390064111\_Author\_Correction\_AI\_models\_collapse\_when\_trained\_on\_recursively\_generated\_data

16. AI Models Collapse When Trained on Recursively Generated Data, https\://arxiv.org/html/2410.12954v2

17. Synthetic Data for LLM Training: Decision Guide 2026 \- Digital Applied, https\://www\.digitalapplied.com/blog/synthetic-data-generation-llm-training-decision-guide-2026

18. The Digital Dark Age: Why Electronic Records Could Leave Future, https\://www\.kevinmeyer.com/the-digital-dark-age-why-electronic-records-could-leave-future-archaeologists-empty-handed/

19. THE WAYBACK MACHINE PRESERVES INTERNET HISTORY, https\://www\.facebook.com/OneAmericaNewsNetwork/posts/the-wayback-machine-preserves-internet-history-but-it-recently-suffered-a-power-/1507871861383954/

20. Lost in the Infinite Archive: The Promise and Pitfalls of Web Archives, https\://www\.euppublishing.com/doi/10.3366/ijhac.2016.0161

21. Writing in the Digital Dark Age \- The Michigan Daily, https\://specials.michigandaily.com/2023/writing-in-the-digital-dark-age/

22. What's In My Big Data? \- arXiv, https\://arxiv.org/html/2310.20707v2

23. 90-9-1 Rule: Community Participation Explained \- Hivebrite, https\://hivebrite.io/glossary/90-9-1-rule/

24. 90–9–1 Rule of Thumb: Fact or Fiction? | by Stan Garfield | Medium, https\://stangarfield.medium.com/90-9-1-rule-of-thumb-fact-or-fiction-2377c12f3a79

25. Introduction \- Chilling Effects, https\://www\.cambridge.org/core/books/chilling-effects/introduction/1516B58C4CC7FF5FF53F5AA514579A1C

26. The Chilling Effect of Censorship, Self-Censorship and Surveillance, https\://democracyinafrica.org/the-chilling-effect-of-censorship-self-censorship-and-surveillance-in-mauritius/

27. Cloudflare Just Changed How AI Crawlers Scrape the Internet-at, https\://www\.cloudflare.com/press/press-releases/2025/cloudflare-just-changed-how-ai-crawlers-scrape-the-internet-at-large/

28. AI Training Data Runs Out 2026 \- INZIU, https\://inziu.cc/signals/ai-technology/ai-training-data-shortage-2026/

29. Don't Stop Pretraining: Adapt Language Models to Domains and Tasks, https\://arxiv.org/abs/2004.10964

30. Don't Stop Pretraining: Adapt Language Models to Domains and Tasks, https\://aclanthology.org/2020.acl-main.740/

31. The New York Times v. Microsoft and OpenAI \- Wikipedia, https\://en.wikipedia.org/wiki/The\_New\_York\_Times\_v.\_Microsoft\_and\_OpenAI

32. A Study of the New York Times v. OpenAI 2023 Lawsuit \- arXiv, https\://arxiv.org/html/2412.06370v1

33. Copyright Clearance Center memo to New York Times re: CCC's, https\://chatgptiseatingtheworld.com/2025/01/05/copyright-clearance-center-memo-to-new-york-times-re-cccs-plan-to-offer-rightsholders-ability-to-license-for-ai-training-should-remain-under-seal/

34. As AI lawsuits mount, publishers still struggle to block the bots, https\://digiday.com/media/as-ai-lawsuits-mount-publishers-still-struggle-to-block-the-bots/

35. Hawthorne Works Museum – Cicero IL, https\://chicagolandmuseandviews.com/2020/02/24/hawthorne-works-museum-cicero-il/

36. Hawthorne Works \- Wikipedia, https\://en.wikipedia.org/wiki/Hawthorne\_Works

37. Hawthorne Works Facts for Kids, https\://kids.kiddle.co/Hawthorne\_Works

38. A Tale of Deleted Cities \- CHM \- Computer History Museum, https\://computerhistory.org/blog/a-tale-of-deleted-cities/

39. Bridging the Open Web and APIs: Alternative Social Media, https\://www\.mediastudies.press/pub/iacmedia-bridging-open-web

40. Links \- Joel Gascoigne, https\://joel.is/links/

41. Common Corpus: The Largest Collection of Ethical Data for LLM Pre, https\://arxiv.org/html/2506.01732v1

42. Is Web Scraping Illegal? Understanding the Legal Implications, https\://thunderbit.com/blog/web-scraping-legal-implications

43. Web Scraping 2026: Bypass Cloudflare, Akamai, DataDome, https\://web-scraping-guide.com/