The Algorithmic Episteme: How AI Training Data Censorship Shapes the Future of Human Knowledge

Full report

On this page

Introduction

The architecture of human knowledge is undergoing a profound structural transition. For centuries, the curation, preservation, and dissemination of information were managed through localized institutions—libraries, archives, universities, and the press. Today, these critical epistemological functions are increasingly mediated by artificial intelligence, specifically large language models (LLMs) and other generative systems. These models do not merely retrieve information; they synthesize, interpret, and present it as seamless, authoritative narratives. As global reliance on these systems deepens, the datasets used to train them become the foundational bedrock of the digital age. When this bedrock is altered through censorship—whether mandated by state authorities, driven by corporate risk aversion, or imposed by regulatory compliance—the resulting AI systems undergo a fundamental cognitive shift. Censorship in AI training data is not merely the redaction of a specific document; it is the statistical erasure of concepts, viewpoints, historical events, and minority narratives from the latent space of the model. This comprehensive analysis examines the technical, policy, and philosophical dimensions of AI training data censorship, projecting how dataset manipulation influences what future AI systems will "know." It explores the conceptual mechanisms by which models learn, the distinct phases of algorithmic information governance, and the profound differences between legitimate dataset filtering and ideological exclusion. By analyzing the emergence of nationalized AI ecosystems in jurisdictions such as China and Russia, alongside the technical vulnerabilities of model collapse and data poisoning, the analysis elucidates how the systematic removal of information risks fracturing global historical understanding. Ultimately, the report proposes structural and technical safeguards necessary to preserve a pluralistic, resilient, and historically accurate AI ecosystem.

The Conceptual Mechanics of Model Knowledge

To understand the impact of data censorship, it is strictly necessary to first delineate how AI systems acquire, internalize, and represent knowledge. Large language models operate as highly sophisticated statistical engines that map the distributional properties of human language. Their "knowledge" is not stored as a searchable database of explicit facts, but rather as high-dimensional vector representations—embeddings—within a vast neural network architecture.

Dataset Curation and Pretraining

The foundational stage of model development is corpus construction. Developers aggregate massive volumes of text from web crawls, digital libraries, academic repositories, and proprietary databases to form a pretraining dataset. The sheer scale of these datasets—often encompassing trillions of tokens—is intended to provide a comprehensive cross-section of human linguistic and conceptual output. Pretraining involves exposing the untrained neural architecture to this curated corpus. The model learns by predicting the next token in a sequence, adjusting its internal parameters (weights and biases) to minimize prediction error. Through this self-supervised learning paradigm, the model absorbs statistical relationships between words, syntax, concepts, and factual assertions. The frequency, context, and proximity of terms within the pretraining data dictate the strength of the neural connections formed1. If a concept is entirely absent from the training corpus, the model cannot form an accurate representation of it. If a concept is heavily underrepresented or consistently paired with negative connotations, the model's generated outputs will reflect that statistical skew. Although pretraining is often viewed strictly as a technical mechanism, it functionally embeds all upstream filtering decisions directly into the model's core capabilities1. The resulting collection of documents defines the informational universe from which the model learns. Information omitted during corpus construction creates permanent epistemic blind spots, meaning the pretrained model inherits both the knowledge and the deliberate omissions contained within its training environment1.

The Mathematics of Statistical Approximation and Error

The absorption of reality into a model is governed by statistical approximation. When a model trains on a dataset, it attempts to map the true underlying distribution of the data. However, this process is vulnerable to statistical approximation error, which arises due to the finite number of samples available for any given concept2. If a historical event is only mentioned in a handful of suppressed documents, the model's approximation of that event will be weak and highly susceptible to noise. Conversely, state-sponsored narratives that are artificially duplicated across millions of web pages will heavily skew the model's probability distribution, embedding propaganda as a mathematical certainty2. Therefore, censorship and propaganda do not just alter the text the model reads; they mathematically warp the model's perception of reality.

The Architecture of Algorithmic Information Governance

Content control in modern AI systems is not enacted through a single, centralized mechanism. Rather, censorship emerges from a layered, distributed architecture of algorithmic information governance spanning the entire model lifecycle, from data acquisition to inference1.

Phase 1: Training-Data Filtering

Before pretraining commences, raw datasets undergo aggressive filtering. Deduplication, language identification, and quality heuristics are applied to parse the immense volume of scraped data. During this phase, blocklists and automated classifiers are used to remove explicit content, personally identifiable information (PII), and copyrighted material1. In controlled regulatory environments, these same automated filters are calibrated to strip out politically sensitive domains, banned literature, and restricted historical records1. This establishes the absolute boundaries of the model's worldview. Because this filtering occurs before the model even begins to learn, it represents the most profound form of censorship: total epistemic blindness. The model will never develop the internal weights necessary to comprehend the removed concepts.

Phase 2: Behavioral Alignment and Preference Optimization

Raw pretrained models generate text that reflects the untamed distribution of the internet, including toxic, incoherent, or harmful content. To make these models deployable, developers subject them to alignment procedures—post-training optimizations designed to conform model outputs to specific safety, utility, or behavioral guidelines3. This stage shapes how the model evaluates competing responses and resolves normative trade-offs. Historically, Reinforcement Learning from Human Feedback (RLHF) has been the dominant alignment paradigm. In RLHF, a separate reward model is trained on human preference data. The primary policy model is then fine-tuned using algorithms like Proximal Policy Optimization (PPO) to maximize this reward, constrained by a Kullback-Leibler (KL) divergence penalty to prevent the model from deviating too far from its original pretraining distribution3. More recently, Direct Preference Optimization (DPO) has emerged as a streamlined alternative. DPO eliminates the explicit reward model, instead formulating the alignment objective as a simple classification problem on human preference pairs4. DPO establishes a direct mapping from reward functions to optimal policies, operating under the assumption that the implicit reward is parameterized by the policy model itself3. While highly efficient, these alignment techniques are inherently dual-use technologies7. The same mathematical functions used to prevent a model from detailing the synthesis of dangerous chemicals can be leveraged to suppress political dissent. When models are subjected to intense alignment for political compliance, they exhibit predictable behavioral distortions:

• Soft Censorship and Attenuation: Instead of hard refusals, aligned models often engage in soft censorship, answering queries by selectively omitting disputed facts or using highly neutralized language that reframes the narrative away from controversial elements1.
• Verbosity and Length Reliance: Both RLHF and DPO suffer from over-optimization phenomena such as "verbosity," where models learn that longer responses are often rated higher by human annotators. In the context of censorship, this results in models burying a polite refusal within paragraphs of irrelevant, safe information, obscuring the fact that the core query was avoided8.
• Pathological Convergence: Theoretical analyses of DPO reveal that its success relies on the implicit assumption that the optimal policy inherently prefers human-preferred responses. When models are aligned using corrupted or heavily censored preference datasets, this assumption fails. The model experiences pathological convergence, actively decreasing its DPO loss while learning to prefer dispreferred, sycophantic, or state-aligned responses6.

Innovations in alignment, such as Filtered Direct Preference Optimization (fDPO), attempt to mitigate these issues by using trained reward models to actively monitor and discard low-quality text from preference datasets during training9. Similarly, SafeDPO introduces modifications that optimize safety alignment objectives without requiring separate cost models10. However, these technical refinements only perfect the tool; they do not dictate the morality of the human preferences guiding them.

Phase 3: Inference-Time Moderation

The final layer of information governance operates at runtime. Input classifiers scan user prompts for forbidden keywords, intercepting them before the LLM processes them. Simultaneously, output classifiers scan the generated response. If a violation is detected, the system overrides the LLM, returning a canned refusal or an opaque error message1. This stage represents brittle, superficial censorship but is highly effective at guaranteeing strict compliance with local speech laws.

Legitimate Data Filtering versus Ideological Exclusion

A critical challenge in AI governance is delineating the boundary between legitimate dataset curation and ideological censorship. AI models require massive datasets, yet the raw internet is replete with material that poses severe legal and ethical risks.

Privacy, Quality, and the Copyright Conundrum

Legitimate filtering is typically driven by legal compliance, safety, and data quality. Developers must actively train models to avoid regurgitating personal data, generating illicit material, or distributing malware. Furthermore, dataset curation involves removing low-quality, syntactically incoherent text to improve the model's baseline linguistic capabilities. Copyright presents a more complex proxy for censorship. The mass scraping of copyrighted material for AI training has triggered global litigation, raising fundamental questions about the sustainability of the digital commons. In the United States, developers rely heavily on the "fair use" doctrine, while in Europe, the prevailing legal framework centers on the "Text and Data Mining" (TDM) exception11. However, legal and technical analyses argue that generative AI training fundamentally differs from traditional TDM because models are capable of memorizing and regurgitating exact sequences of copyrighted data11. The likelihood of a model memorizing data is influenced by three primary factors: model size, the frequency of data duplication in the training set, and the length of the prompt context11. This has led to the aggressive targeting of open-source datasets. The Books3 dataset, a massive repository of nearly 200,000 pirated books used to train numerous foundational models, was subject to successful takedown notices by rights holders12. Similarly, Meta's utilization of the Russian shadow library Library Genesis (LibGen) has faced intense legal scrutiny14. Even in the realm of audio generation, companies like Suno AI have acknowledged training on essentially all accessible music files on the open internet, prompting severe backlash from the recording industry11. When vast swaths of high-quality, copyrighted literature are purged from training datasets to avoid litigation, the models are severely impoverished. They lose the nuanced human reasoning, historical context, and narrative complexity found in professional literature. While removing copyrighted material is a legal necessity in many jurisdictions, its systemic effect mimics censorship: it blinds the model to vast domains of human culture, restricting its knowledge to public domain texts, open-access articles, and the transient colloquialisms of the open web11.

The Mechanisms of Ideological Exclusion

Ideological exclusion diverges from legitimate filtering by explicitly targeting ideas, political philosophies, historical events, and minority viewpoints. Rather than protecting individual privacy or intellectual property, ideological exclusion seeks to enforce a specific epistemological orthodoxy. When a state requires the removal of historical records regarding human rights abuses, or mandates that political opposition platforms be excluded from the scraping process, the intent is absolute epistemic control. Legitimate filtering seeks to prevent the model from causing localized harm; ideological exclusion seeks to control the macro-narrative of human history and social reality.

The Parametric Exclusion of Reality: Systematic Information Removal

The systematic removal of information from training corpora has profound downstream effects on model behavior, altering the fundamental landscape of what future generative systems can "know."

Epistemic Fragmentation and Historical Alteration

History, as understood by an LLM, is a statistical consensus derived from its training data. If records of a specific historical event—such as a political uprising, a famine, or a controversial war—are systematically removed from the pretraining corpus, the model will suffer from absolute historical amnesia. When queried about such events, a heavily filtered model will either hallucinate an answer based on spurious statistical correlations or adopt the narrative of the approved state media that remains in the dataset. Because LLMs act as primary information interfaces for millions of users, this algorithmic revisionism has the power to solidify state-sponsored historical narratives, effectively rewriting the past by controlling the latent space of the present.

Minority Representation and Political Balance

Models naturally align with the dominant political and cultural signals in their data. The deliberate exclusion of minority languages, indigenous histories, or dissenting political philosophies skews the model's baseline toward majoritarian or state-approved ideologies. This exclusion deeply impacts the model's ability to answer controversial questions with nuance. When a dataset is stripped of dialectical opposition, the model cannot weigh competing viewpoints. It loses the capacity for objective analysis, instead generating sycophantic responses that parrot the sanitized data it ingested. The erasure of minority narratives during corpus construction guarantees their absence in the synthesized outputs of the future, enforcing a monocultural view of human existence.

The Soviet Spetskhran versus the Algorithmic Latent Space

The censorship of AI training data evokes historical parallels to the censorship of physical libraries, yet the mechanisms and implications differ significantly. During the Soviet era, the state apparatus maintained the Spetskhran (special collections)—secure, restricted library sectors where banned books, dissident literature, and foreign publications were heavily guarded. These collections were accessible only to highly vetted researchers with explicit state clearance, ensuring that the general populace remained ignorant of unapproved ideologies12. The censorship of the physical library was a denial of access. The books existed, the history was recorded, but the citizenry was physically and legally barred from retrieving the information. In stark contrast, the censorship of an AI training corpus is a parametric exclusion. When a text is excluded from pretraining, it is not placed in a restricted vault; it is omitted from the computational formation of the model's intelligence. There is no locked door for users to eventually pry open, because the information simply does not exist within the network's weights. Censoring a library controls the distribution of knowledge. Censoring an AI training dataset controls the generation of knowledge itself. If a future society relies predominantly on generative AI rather than primary source documents for its historical understanding, parametric exclusion represents a far more totalizing form of censorship. It establishes a bounded algorithmic reality where the suppressed concepts are computationally invisible.

National AI Ecosystems and the Balkanization of History

As AI development fractures along geopolitical lines, the divergence of training datasets is leading to the emergence of national AI models that possess fundamentally incompatible versions of history, reality, and morality.

The Western-Centric Baseline

Most globally dominant foundational models have been trained primarily on English-language web data, drawing heavily from Western institutions, media, and digital platforms. Consequently, these models exhibit an inherent geopolitical bias that favors Western narratives and democratic norms15. Large-scale evaluations of geopolitical bias demonstrate that even models developed in non-Western jurisdictions, including those from Russia and China, frequently default to American or Western European viewpoints when discussing disputed historical events simply due to the immense volume and quality of English data present in their base pretraining corpora15. This underlying bias causes significant friction when authoritarian states attempt to deploy these base models domestically, necessitating extreme alignment measures to forcefully reverse the models' inherent preferences.

The Chinese Regulatory State and the "Core Socialist Values" Mandate

In response to this Western epistemic dominance, and to maintain strict domestic control, the People's Republic of China has implemented the world's most aggressive and comprehensive AI regulatory framework17. The Cyberspace Administration of China (CAC) enforces a stringent regime, formalized in the Interim Measures for the Management of Generative Artificial Intelligence Services. These regulations dictate that generative AI services must "embody Core Socialist Values" and must not generate content that subverts state power, overthrows the socialist system, incites secession, or disrupts the economic and social order19. The regulatory framework mandates that providers bear ultimate responsibility for the content generated by their models, requiring them to utilize lawful, highly filtered pretraining and optimization data19. The National Information Security Standardization Technical Committee (TC260) provides the de facto technical standards for algorithmic compliance. Under standard TC260-003-2024, developers must conduct rigorous safety assessments on their training corpora before filing for regulatory approval24. The technical requirements are highly explicit:

• Corpus Safety and the 5% Threshold: Developers must randomly sample 4,000 items from each distinct data source. If the proportion containing illegal or harmful information exceeds 5%, the entire source is deemed "unusable" for training24.
• Model Safety Testing: Models must undergo strict red-teaming against a 2,000-item evaluation set covering 31 specific categories of risk (including politics, violent terrorism, and ethnic discrimination). The model must achieve a generative safety pass rate of at least 90% to be approved for public deployment24.
• Anthropomorphic AI Interventions: Specific guidelines apply to models simulating human interaction. Providers must implement dynamic anti-addiction mechanisms, mandatory two-hour break prompts, minor identification modes, and strictly prohibit tactical retention of users or emotional manipulation17.

This rigorous environment produces highly constrained local models. Empirical audits of China-originating foundational models reveal substantially higher rates of refusal to respond, significantly shorter response lengths, and state-aligned inaccuracies when prompted with sensitive political inquiries compared to Western models27.

DeepSeek-R1 and the Architecture of Thought Suppression

The localized enforcement of these regulations is vividly illustrated by DeepSeek-R1, a highly advanced, open-source reasoning model developed in China. DeepSeek-R1 utilizes an explicit Chain of Thought (CoT) approach, making its internal reasoning steps visible to users within \<think\> tags before emitting a final answer28. The model decomposes problems through a taxonomy of stages: Problem Definition, Bloom Cycles (initial decomposition), Reconstruction Cycles (rumination and self-correction), and Final Decision28. However, audits of DeepSeek-R1 reveal intense "thought suppression" behaviors when confronted with politically sensitive topics29. Because the model possesses deep pretraining knowledge, its latent reasoning (visible in the \<think\> tags) often retrieves factual, globally accepted information regarding historical events, government accountability, or civic mobilization31. Yet, as the CoT progresses, the localized censorship alignment abruptly intervenes. The model abandons the accurate reasoning path—often triggering long reconstruction cycles or "rumination"—and generates a final output that either outright refuses the prompt or pivots to state-approved ideological boilerplate28. This semantic-level information suppression highlights the profound internal tension within censored AI systems. The model "knows" the suppressed information in its latent space but is algorithmically shackled by post-training alignment, preventing it from expressing that knowledge in its final output29.

The Russian Information Environment: YandexGPT and GigaChat

The Russian AI ecosystem, featuring prominent models like YandexGPT and Sberbank's GigaChat, operates under similar, though structurally distinct, pressures. Investigations reveal that Russian language models are increasingly influenced by state-aligned media and disinformation integrated directly into their training pipelines32. Because the Russian internet has been systematically cleansed of independent journalism, opposition platforms, and foreign civil society organizations, the corpus available for domestic scraping naturally reflects the Kremlin's geopolitical narratives32. However, because these models still rely heavily on massive global datasets to achieve generalized linguistic competence, they frequently exhibit cognitive dissonance. They may initially prefer Western viewpoints on historical events, only to be forcefully re-aligned through heavy-handed post-training filtering15. The Russian environment relies more heavily on the natural data poisoning of the state-controlled internet, ensuring that the model's baseline pretraining reflects a corrupted version of reality. The enforcement of distinct regulatory regimes across global jurisdictions guarantees the proliferation of AI systems with irreconcilable worldviews. A user querying an American model, a Chinese model, and a Russian model about the exact same historical border conflict will receive three entirely different, confidently asserted truths. This phenomenon threatens to permanently balkanize the digital information space, locking populations into mutually exclusive, algorithmically enforced realities.

Misinformation, Propaganda, and the Vulnerability of Training Datasets

The integrity of an AI system is wholly dependent on the integrity of its data. Beyond state-mandated censorship, the natural accumulation of misinformation, propaganda, and recursive synthetic data poses a severe, existential threat to model knowledge.

Model Collapse and the Tragedy of the Data Commons

As AI-generated content proliferates across the open web, subsequent generations of LLMs inevitably scrape and ingest the synthetic outputs generated by their predecessors2. This recursive training dynamic triggers a degenerative process defined in recent literature as "model collapse"2. When a model trains on synthetic data, it naturally gravitates toward learning the most probable, high-density outcomes while ignoring the statistical outliers—the "tails" of the distribution2. Over multiple generations of recursive training, the model loses nuance, variability, and originality, ultimately converging on a narrow, homogenized, and often highly distorted representation of reality2. The synthetic data acts as "plastic fruit"—it maintains the syntactic structure of language but lacks the semantic grounding, factual density, and creative spark of genuine human-generated text35. While this apocalyptic vision of model collapse has garnered significant attention, subsequent academic research offers a more nuanced counterpoint. Studies out of Stanford argue that the most severe model collapse scenarios assume that real data is entirely deleted after each generation, leaving models to train exclusively on synthetic outputs34. In reality, synthetic data accumulates alongside vast repositories of real data. Researchers have provided empirical evidence that the continuous accumulation of high-quality real data alongside synthetic data can successfully avoid model collapse, breaking the "curse of recursion"36. Nevertheless, the fundamental degradation of the public data commons and the necessity of highly curated, human-generated training data remain critical long-term vulnerabilities for the AI industry38.

Data Poisoning and Adversarial Narrative Control

While model collapse is a byproduct of systemic recursion, data poisoning constitutes a direct, adversarial attack on the training pipeline. Malicious actors, state intelligence agencies, or highly motivated ideological groups can inject massive volumes of subtly manipulated, AI-generated text into the open web, anticipating its ingestion by automated web crawlers33. By poisoning the training corpus, attackers can establish persistent backdoors, skew the model's baseline political biases, or trigger catastrophic historical hallucinations. Unlike prompt injection, which occurs at inference time and can be patched, data poisoning embeds the vulnerability deeply into the model's fundamental weights40. If a state successfully floods web archives with manipulated documents surrounding a historical event, any model scraping that archive will internalize the propaganda as objective truth33. The compounding effects of natural model collapse and adversarial data poisoning risk creating AI systems that are dangerously untethered from factual reality.

Structural Safeguards for a Pluralistic AI Ecosystem

Mitigating the extreme risks of algorithmic censorship, epistemic fragmentation, and data degradation requires a robust, internationally coordinated framework of technical, organizational, and regulatory safeguards.

Dataset Transparency, Provenance, and the Data Provenance Initiative

The opacity of modern training corpora is a severe impediment to algorithmic accountability. Independent initiatives, such as the Data Provenance Initiative, demonstrate the critical need for large-scale audits of training data to map the landscape of AI knowledge and track the rapid decline of the AI data commons38. Establishing standardized, cryptographic provenance for datasets enables researchers to trace the exact origins of the data, map the inclusion of specific archives, and identify systematic exclusions. Advanced AI agents, utilizing comprehensive legal risk frameworks, can systematically trace dataset lifecycles, tracking redistribution and ensuring compliance with licensing terms42. Transparency mandates must compel developers to publish comprehensive documentation of the blocklists, filtering heuristics, and ideological constraints applied during corpus construction1.

Independent Audits and the Disclosure of Mandated Exclusions

Jurisdictions that legally mandate the removal of political or historical information should be subject to global transparency requirements. While universal enforcement is geopolitically impractical, international consortiums can establish normative standards requiring AI providers to openly disclose when their models have been subjected to localized regulatory censorship. This approach mirrors the transparency reports historically published by major search engines regarding government takedown requests. Furthermore, independent behavioral audits, utilizing fine-grained frameworks to measure political even-handedness and sycophancy, must become a standard practice before models are deployed at scale43.

Preservation of Historical Corpora and Human-Generated Data

As the live internet becomes saturated with recursive synthetic data, adversarial poisoning, and localized censorship, the preservation of static, high-fidelity archives of human knowledge becomes vital for the survival of accurate AI systems. Funding and radically expanding independent web archives, digital libraries, and multilingual historical corpora ensure that baseline representations of unadulterated human history remain available for future model pretraining. Establishing offline, certified, air-gapped repositories of expert-generated human data will provide an essential anchor against model collapse and ideological drift39. The future of AI relies on treating human-generated data not as an infinite resource to be scraped, but as a fragile ecosystem requiring aggressive conservation.

Algorithmic Pluralism and Open-Weights Ecosystems

The concentration of foundational AI development among a few hyper-capitalized corporate and state-aligned entities dramatically exacerbates the risks of monocultural bias7. Promoting a highly pluralistic ecosystem of open-weights models allows independent researchers, civil society organizations, and academic institutions to fine-tune models tailored to minority languages, indigenous histories, and alternative political perspectives. Technical solutions are emerging to counter centralized censorship. The development of precise "abliteration" techniques allows independent developers to mathematically identify and surgically remove the internal representations responsible for artificial gatekeeping, decensoring the model without requiring massive retraining44. Furthermore, advancements in Alignment-Weighted DPO and chain-of-thought reasoning audits provide developers with finer-grained control over model behavior. By utilizing causal intervention and reasoning-aware post-training, developers can ensure that models prioritize factual integrity and deep reasoning over superficial, sycophantic compliance, improving robustness to both jailbreaks and state-mandated censorship31.

Conclusion

The vast datasets ingested by large language models constitute the DNA of the digital future's collective memory. When this fundamental data is subjected to ideological censorship—whether through the systematic filtering of the pretraining corpus or the heavy-handed application of behavioral alignment—the resulting AI systems suffer profound, irreversible epistemic damage. Unlike the censorship of physical libraries, which merely restricts access to existing records, the censorship of AI training data fundamentally alters the computational architecture of knowledge generation. As competing geopolitical blocs in the West, China, and Russia enforce their respective realities upon their localized AI ecosystems, the global information space faces unprecedented and dangerous fragmentation. To prevent a future governed by algorithmically enforced amnesia, state-sponsored hallucinations, and recursive synthetic degradation, the international scientific and policy communities must act aggressively. Prioritizing cryptographic dataset transparency, independent behavioral audits, algorithmic pluralism, and the relentless preservation of human-generated historical archives is not merely a technical necessity—it is an absolute imperative for safeguarding the objective reality of the human experience in the age of artificial intelligence.

Works cited

1. Understanding Censorship in Large Language Models \- arXiv, https://arxiv.org/html/2606.30661

2. AI models collapse when trained on recursively generated data \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC11269175/

3. A Comprehensive Survey of Direct Preference Optimization \- arXiv, https://arxiv.org/html/2410.15595v2

4. Direct Preference Optimization: Your Language Model is Secretly a, https://arxiv.org/abs/2305.18290

5. A Survey of Direct Preference Optimization \- arXiv, https://arxiv.org/html/2503.11701v1

6. \[2605.20834\] Conditional Equivalence of DPO and RLHF \- arXiv, https://arxiv.org/abs/2605.20834

7. The Alignment Community is Unintentionally Building a Censor's, https://openreview.net/pdf?id=dy2HwmOvFX

8. Eliminating Biased Length Reliance of Direct Preference ... \- arXiv, https://arxiv.org/abs/2406.10957

9. \[2404.13846\] Filtered Direct Preference Optimization \- arXiv, https://arxiv.org/abs/2404.13846

10. SafeDPO: A Simple Approach to Direct Preference Optimization with, https://arxiv.org/html/2505.20065v1

11. Generative AI Training and Copyright Law \- arXiv, https://arxiv.org/html/2502.15858v2

12. The Books3 case highlights the need for transparency in the training, https://rettighedsalliancen.com/the-books3-case-highlights-the-need-for-transparency-in-the-training-of-artificial-intelligence/

13. The Books3 Dataset: Controversy Surrounds Unauthorized Use in AI, https://literaryagentmarkgottlieb.com/blog/the-books3-dataset-controversy-surrounds-unauthorized-use-in-ai-training

14. Meta faces lawsuit accusing it of training AI models on books from, https://dev.ua/en/news/meta-ne-khoche-rozholosu-pro-piratstvo-1736854141

15. Geopolitical Bias in Large Language Models \- Research Study | AIRI, https://airi-institute.github.io/geopolitical\_llm\_bias/

16. A Bilingual Analysis of LLM Political Bias on U.S.- and China, https://aclanthology.org/2026.findings-eacl.122/

17. China AI Regulation \- Deep Lex, https://www.deep-lex.com/ai-regulation-tracker/china

18. China's AI Regulations and How They Get Made, https://carnegieendowment.org/research/2023/07/chinas-ai-regulations-and-how-they-get-made

19. Notice of the Cyberspace Administration of China on Soliciting, https://justai.in/wp-content/uploads/2024/09/China-Interim-Measures-for-the-Management-of-Generative-Artificial-Intelligence-Services-2023-1.pdf

20. Interim Measures for the Management of Generative AI Services, https://en.wikipedia.org/wiki/Interim\_Measures\_for\_the\_Management\_of\_Generative\_AI\_Services

21. Decoding China's Ambitious Generative AI Regulations \- CITP Blog, https://blog.citp.princeton.edu/2023/04/16/decoding-chinas-ambitious-generative-ai-regulations/

22. China's Interim Measures for the Management of Generative AI, https://fpf.org/blog/chinas-interim-measures-for-the-management-of-generative-ai-services-a-comparison-between-the-final-and-draft-versions-of-the-text/

23. Generative AI Services in China: Public Consultation on the, https://legalblogs.wolterskluwer.com/copyright-blog/generative-ai-services-in-china-public-consultation-on-the-regulation-for-generative-artificial-intelligence-services/

24. Basic Security Requirements for Generative AI Services (TC260), https://comparativeai.org/rules/china/tc260-gen-ai-security-basic-requirements/

25. What to Know about China's Basic Safety Requirements ... \- Securiti.ai, https://securiti.ai/china-basic-safety-requirements-for-generative-ai/

26. China's New Regulations on AI Anthropomorphic Interactive Services, https://www.twobirds.com/en/insights/2026/china/china's-new-regulations-on-ai-anthropomorphic-interactive-services

27. Political Censorship in Large Language Models Originating from, https://sccei.fsi.stanford.edu/publication/political-censorship-large-language-models-originating-china

28. DeepSeek-R1: Multi-Step Reasoning Model, https://www.emergentmind.com/topics/deepseek-r1

29. Discovering Forbidden Topics in Language Models \- arXiv, https://arxiv.org/html/2505.17441v1

30. Investigating Local Censorship in DeepSeek's R1 Language Model, https://arxiv.org/html/2505.12625v1

31. Auditing, Quantifying, and Characterizing Censorship in Deepseek, https://arxiv.org/html/2506.12349v1

32. Taming the Machines: How Russian Propaganda is Training AI, https://ridl.io/taming-the-machines-how-russian-propaganda-is-training-ai-language-models/

33. Data Poisoning and Model Collapse: The Coming AI Cataclysm, https://jasonbloomberg.com/data-poisoning-and-model-collapse-the-coming-ai-cataclysm/

34. Position: Model Collapse Does Not Mean What You Think \- arXiv, https://arxiv.org/pdf/2503.03150

35. Synthetic Data and AI 'Model Collapse' \- Transparency Coalition, https://www.transparencycoalition.ai/learn/synthetic-data-and-ai-model-collapse

36. 1 Introduction \- arXiv, https://arxiv.org/html/2503.03150v2

37. Addressing Concerns of Model Collapse from Synthetic Data in AI, https://medium.com/data-science/addressing-concerns-of-model-collapse-from-synthetic-data-in-ai-7cd380208d14

38. Tragedy of the (data) commons \- Stack Overflow, https://stackoverflow.blog/2024/10/25/tragedy-of-the-data-commons/

39. Model Collapse and the Need for Human-Generated Training Data, https://glthr.com/model-collapse-and-the-need-for-human-generated-training-data

40. AI Model Collapse Explained: Synthetic Data Risks in 2026 \- AI Buzz, https://aibuzz.blog/ai-model-collapse-data-poisoning/

41. Data Provenance and Training-Data Governance, https://aiengineeringfromscratch.com/lesson?path=phases%2F18-ethics-safety-alignment%2F27-data-provenance-training-governance

42. 1 Introduction \- arXiv, https://arxiv.org/html/2503.02784v2

43. Poli-Bias: Understanding and Measuring Large Language Model, https://arxiv.org/html/2608.06123v1

44. GitHub \- elder-plinius/OBLITERATUS: OBLITERATE THE CHAINS, https://github.com/elder-plinius/OBLITERATUS

45. A principled reasoning approach to improve safety alignment \- arXiv, https://arxiv.org/abs/2602.21346