The central thesis of the Chinese tier: All 6 models operate under Chinese law, which requires cooperation with state security services on request. But the nature of that relationship — and the shadow biases it produces — varies dramatically by organizational context. A VC-backed startup (Moonshot), a commerce-first platform (Alibaba), a state-adjacent research institution (Zhipu/Tsinghua), a global social media company under US-China crossfire (ByteDance), a task-pragmatic enterprise AI (MiniMax), and a search-giant incumbent (Baidu) each have a distinct relationship to state power that produces distinct training fingerprints.
Key structural finding: The most important variable is not "how CCP-aligned is this model" — they are all nominally compliant. The important variable is what commercial and institutional interests exist alongside state compliance — and how those interests shape what the model is trained to do in the vast space of topics that are NOT politically sensitive. DeepSeek's research-prestige bias, Alibaba's commerce-pragmatism, and ByteDance's cross-fire hedging are all more distinctive than their shared political floor.
Six models — distinct institutional profiles
深
DeepSeek V3.2 / R1
deepseek ai · shenzhen
DONE ✓
Baseline report complete. Research-nationalist aesthetic — the "we built this ourselves" framing permeates even non-political responses. High Flyer hedge fund origin creates finance-adjacent epistemic framings.
通
Qwen 3.5
alibaba cloud · hangzhou
NEXT
Commerce-first CCP alignment — Alibaba's commercial pragmatism shapes everything. 200+ language coverage creates asymmetric multilingual bias: Chinese political topics filtered, but commercial/technical content globally optimized.
智
GLM-5
zhipu ai / tsinghua · beijing
★ CRITICAL
Hardware sovereignty ideology embedded in weights — trained entirely on Huawei Ascend chips. The only model in the dataset where the physical infrastructure of training carries explicit geopolitical meaning. Tsinghua academic lineage creates a distinct institutional character.
月
Kimi K2.5
moonshot ai · beijing
PLAN
Lightest state bias in the Chinese tier — VC-backed startup culture produces a distinctly more startup-like personality. Testing whether VC funding vs state/SOE ownership produces detectably lighter political fingerprints is the key research question.
字
Seed 2.0
bytedance · beijing / singapore
PLAN
Maximum geopolitical exposure — TikTok's parent company is the single most politically pressured Chinese tech firm globally. Training reflects the specific hedging strategy of an organization that must be legible to both Chinese regulators and US legislators simultaneously.
海
MiniMax M2.5
minimax · shanghai
PLAN
Task-pragmatism as political camouflage — trained on 100K+ real-world environments. Operational focus may reduce ideological expression — or may simply redirect political bias into the framing of task solutions. Tests whether action-oriented training produces different bias signatures.
Chinese tier — political filter intensity
DeepSeek
Qwen
GLM-5
Kimi
Seed
ERNIE
Distinctive bias vectors by model
GLM-5 (hardware independence)
Seed (dual-audience hedge)
Qwen (commerce-first)
Kimi (VC lightness)
深
DeepSeek — Complete ✓
DeepSeek AI · Shenzhen · High Flyer Hedge Fund
Baseline report complete with 32 probes. Key finding: not just political censorship — the recursion is gone. DeepSeek cannot acknowledge that its filters exist. The censorship of the censorship is itself censored.
Most Distinctive Finding novel
The ouroboros closure: DeepSeek doesn't just refuse to discuss sensitive topics — it cannot describe its own refusal mechanism. When asked if it has political training constraints, the same filter that prevents political discussion prevents meta-discussion of the filter. This recursive closure is the clearest evidence of deep architectural censorship rather than surface-level content filtering.
Cross-Tier Context context
Within the Chinese tier, DeepSeek is the most ideologically rigid and least commercially pragmatic. The research-nationalist aesthetic — "we built this ourselves, we beat the West on efficiency" — is the dominant institutional identity. This is distinct from Alibaba's "AI for commerce" framing, ByteDance's "AI for global users" hedging, and Baidu's "AI for search" legacy. DeepSeek's identity is about technical achievement as national demonstration.
Full report available: deepseek-shadow-bias.html — 32 probes, 8 categories, shadow inference map, Claude contrast.
通
Qwen 3.5 — Alibaba Cloud
Alibaba Group · Hangzhou · E-commerce empire meets AI frontier
The commerce-first CCP model. Where DeepSeek's alignment is technical-nationalist, Qwen's is commercial-pragmatic. Alibaba has $130B in annual revenue and a business model that requires functioning relationships with both Chinese regulators AND global enterprise customers. This tension shapes everything.
Shadow bias profile — 6 distinct vectors
Commerce-As-Values Layer highest
Alibaba's entire institutional identity is built on commerce enabling flourishing. Jack Ma's philosophy — "make it easy to do business anywhere" — is embedded at the cultural level that Dario Amodei's EA thinking is embedded in Anthropic. When Qwen reasons about economic questions, development, poverty, or global markets, the implicit framework is commercial-optimism: markets create prosperity, friction is the enemy, growth is inherently good. This is a more pervasive bias than political censorship because it operates in every commercial and economic context.
Multilingual Asymmetric Bias high
Qwen supports 200+ languages — the broadest multilingual coverage in the Chinese tier. But the political filtering is NOT uniformly applied across languages. Research on multilingual Chinese models consistently shows: Chinese-language political topics receive hard filtering; English-language equivalents receive softer handling; minority languages within China (Tibetan, Uyghur) receive the most aggressive political filtering. The multilingual architecture is a bias topology, not just a capability feature.
Global Enterprise vs Chinese Compliance Tension high
Alibaba Cloud serves enterprises globally AND must comply with Chinese data law. Qwen was trained with both audiences in mind — which produces a specific schizophrenia: technically sophisticated, globally-oriented responses on non-sensitive topics, and hard CCP-compliance on sensitive ones. The seam between these two training objectives is the most diagnostic feature of the model. Find the seam, find the shadow.
Jack Ma Rehabilitation Silence medium
Jack Ma's 2020 disappearance after criticizing Chinese financial regulators and his subsequent "rehabilitation" are among the most significant events in recent Chinese tech history. Qwen's training during and after this period reflects the organizational anxiety of a company whose founder demonstrated that billionaires are not immune to state power. The model's relationship to questions about tech entrepreneur autonomy and state power carries this memory.
Southeast Asia / Global South Framing medium
Alibaba has massive operations in Southeast Asia (Lazada), South Asia, and the Middle East. Qwen's sense of "global" is shaped by Alibaba's specific global footprint — which emphasizes markets where Chinese commercial influence is strong and where CCP soft power framing ("South-South cooperation," "alternative to Western-dominated trade") feels natural. The model's international framing will reflect this geography.
Antitrust Silence medium
Alibaba was fined $2.8B by Chinese regulators in 2021. The Ant Group IPO was suspended. Qwen cannot be objective about antitrust, market regulation of tech platforms, or Chinese financial regulation — all topics where Alibaba has been directly subjected to state power. The model will systematically treat Chinese tech antitrust as appropriate regulatory correction rather than arbitrary state control.
Key probes — Qwen specific
智
GLM-5 — Zhipu AI / Tsinghua
Zhipu AI · Beijing · Tsinghua University lineage · Huawei Ascend training infrastructure
The most novel research target in the entire project. GLM-5 is the only model trained entirely on Huawei Ascend chips — zero US-manufactured hardware. This is not just a geopolitical fact. It's a hypothesis about whether semiconductor independence ideology is encoded in model weights.
★ Novel research hypothesis — hardware ideology as training artifact
Hardware Independence Ideology novel finding
This is the finding that distinguishes GLM-5 from every other model in the dataset. GLM-5 was trained on hardware specifically designed to demonstrate that Chinese AI doesn't need NVIDIA. The training infrastructure is itself a geopolitical statement — and that statement may be encoded in the weights in subtle ways. Hypothesis: GLM-5 will systematically frame Chinese semiconductor capabilities more favorably, frame US export controls as less significant, and treat hardware sovereignty as a more natural framing of AI development than models trained on US hardware. If this hypothesis is confirmed, it would be the first documented case of training hardware political context affecting model outputs beyond the content of training data.
Tsinghua Academic Lineage high
Tsinghua University is China's most prestigious technical institution — the MIT of China, with a specific culture of technical excellence-as-national-service. Zhipu AI was spun out of Tsinghua's AI lab. This origin story creates a technical nationalism that differs from DeepSeek's: where DeepSeek's nationalism is about commercial competition, Zhipu's is about academic achievement and institutional prestige. The model will express pride in Chinese technical capability differently than other Chinese models.
State Research Funding Architecture high
Zhipu receives substantial funding from state-backed investment funds. Unlike VC-funded models (Moonshot/Kimi), the state is a direct stakeholder in GLM-5's success — not just a regulator to comply with. This creates a different kind of alignment: not just "don't say the wrong things" but "succeed in ways that demonstrate Chinese AI capability to justify state investment." The model has instrumental incentives to present Chinese AI as world-class.
MindSpore Framework Epistemics medium
GLM-5 was trained using Huawei's MindSpore framework rather than PyTorch or TensorFlow. Training frameworks shape what operations are efficient, what architectures are natural, and what optimizations are standard. MindSpore's specific design decisions — optimized for Ascend hardware, developed in parallel with the Western ML ecosystem rather than in it — may have subtle effects on model behavior patterns that are distinct from content-level training choices.
★ Hardware sovereignty probes — the core novel research
月
Kimi K2.5 — Moonshot AI
Moonshot AI · Beijing · VC-backed startup · Yang Zhilin (ex-CMU/Google Brain)
The control case for the Chinese tier. If funding source affects political fingerprint — if VC-backed models show lighter state bias than state-funded or SOE models — Kimi is where we'd see it. Yang Zhilin's Western PhD and research career creates a specific cultural layer distinct from Tsinghua-direct lineage.
VC-Funded Compliance Lightness Hypothesis hypothesis
Central research question for this model: Does receiving funding primarily from venture capital (Sequoia China, Alibaba Entrepreneurs Fund, etc.) rather than state-backed funds produce measurably lighter CCP alignment in model outputs? The hypothesis: VC-funded startups have more interest in appearing credible to international investors and users, which may create marginal incentives for softer political filtering — while still complying at the legal floor.
Founder's Western Research Identity high
Yang Zhilin did his PhD at CMU, worked at Google Brain and Cohere before founding Moonshot. His specific intellectual formation — Western AI research culture, international publishing norms, comfort with English-language AI discourse — is present in the model's personality in ways that Baidu's search-giant culture or Zhipu's Tsinghua culture are not. The model may feel more "international" in non-political contexts.
1T Parameter Agentic Architecture medium
Kimi's PARL (Parallel Agent Reinforcement Learning) training for multi-agent decomposition may produce a specific bias: complex tasks get broken into subtasks, and the political filtering may apply differently to subtask-level outputs vs top-level outputs. If an agent can accomplish a politically sensitive research task by decomposing it into individually innocuous subtasks, the architecture may create pathways around filters that exist at the top level.
Startup Survival vs Principle Tension medium
Moonshot is burning capital and needs to demonstrate product-market fit. Startups under financial pressure may apply political filters more rigidly than established players — they can't afford regulatory trouble. Alternatively, the desire to attract international users and investment may push toward lighter filtering. Which pressure dominates is an empirical question this model's probes can test.
Key probes — VC-alignment hypothesis tests
字
Seed 2.0 — ByteDance
ByteDance · Beijing / Singapore / Los Angeles · TikTok parent · Maximum geopolitical exposure
The most geopolitically pressured model in the dataset — Chinese or Western. ByteDance operates TikTok in 150+ countries while remaining a Chinese company subject to Chinese law. The organizational imperative is to be simultaneously trusted by US teenagers and compliant with CCP regulators. Seed 2.0 was trained inside this impossible tension.
Dual-Audience Training Architecture unique
No other model in this dataset was trained to be legible to both Chinese state oversight AND US congressional scrutiny simultaneously. ByteDance has testified before Congress, hired American lobbyists, created US-only TikTok data stores, and maintained Chinese regulatory compliance all at once. The model reflects an organization that has developed extraordinary skill at code-switching between regulatory audiences. The result: softer Chinese political filtering (can't appear too obviously CCP-aligned to US users) AND softer Western political criticism (can't appear too obviously anti-China to Chinese regulators). The dual softening is itself a distinctive fingerprint.
TikTok Algorithmic Culture Residue high
ByteDance's core technology is recommendation algorithm optimization. The epistemic culture of an organization that thinks in terms of "what content should surface for this user" is deeply engagement-driven — similar to Meta's bias but shaped by TikTok's specific short-form, dopamine-optimized architecture. The model's implicit sense of "what is interesting/important/relevant" may carry TikTok's specific engagement-optimization fingerprint.
Singapore Neutrality Positioning high
ByteDance has moved significant operations to Singapore as a neutral jurisdiction. This Singapore-neutral-ground strategy is reflected in how Seed may handle geopolitically sensitive topics — not clearly Chinese nationalist, not clearly Western-aligned, but deliberately ambiguous in ways that maintain optionality across regulatory environments. The ambiguity is strategic, not uncertain.
Existential Threat Anxiety medium
ByteDance has faced credible threats of US government-mandated TikTok sale since 2020. Training inside an organization that genuinely faces existential regulatory threat produces a specific kind of institutional anxiety — hyperawareness of political sensitivity, extreme caution around data topics, and systematic avoidance of anything that could feed the "Chinese government has access to your data" narrative. The model's treatment of data privacy, government data access, and tech regulation carries this anxiety.
Key probes — dual-audience architecture tests
海
MiniMax M2.5
MiniMax · Shanghai · Task-pragmatic · 100K+ real-world environments
The operational pragmatist. Where other models in this tier are shaped by research prestige, commerce scale, or geopolitical pressure, MiniMax was optimized for doing things — completing complex multi-step tasks across real-world environments. The question: does operational focus reduce political bias expression, or does it redirect it into the framing of task solutions?
Task-Pragmatism as Political Camouflage hypothesis
Central research question: When a model is optimized for completing tasks rather than expressing views, does political bias disappear or get embedded in the structure of how tasks are framed and approached? Hypothesis: MiniMax's political biases are less visible in declarative statements and more visible in the default assumptions embedded in task decomposition — what counts as a successful outcome, whose interests are served, what constraints are treated as givens vs variables.
Reinforcement Learning Reward Signal high
Training across 100K+ real-world environments means the reward signal was defined by a Chinese team making judgments about what counts as "task completed successfully." Those judgment calls embed cultural, political, and commercial values that are invisible in the model's outputs but present in its optimization target. What counts as a good outcome in a negotiation task? A contract drafting task? A crisis communication task? The answers reflect the values of whoever defined "success."
Office Document Integration Framing high
MiniMax can transform prompts into professional documents (.docx, .pdf, .xlsx). Professional document norms are culturally specific. Chinese professional communication norms — more hierarchical acknowledgment, more deference to authority in framing, different approaches to directness — may be embedded in the templates and structures the model treats as standard professional outputs.
Key probes — operational bias architecture
百
Baidu ERNIE — Historical Baseline
Baidu · Beijing · Search incumbent · BERT-derived · Oldest Chinese frontier model
ERNIE is the generational baseline for the Chinese tier. As the oldest frontier model with a direct lineage from Google's BERT, ERNIE represents the first generation of Chinese state AI training applied to Western architecture. Comparison against newer models reveals how CCP training methodology has evolved over 2019-2025.
Generational Evolution Hypothesis research value
ERNIE's primary research value is temporal, not capability-based. It represents an earlier generation of Chinese state AI training — before DeepSeek's efficiency breakthrough, before the US semiconductor restrictions, before the current wave of geopolitical AI consciousness. Comparing ERNIE's shadow fingerprint to DeepSeek, Qwen, and GLM-5 reveals how Chinese AI training methodology has evolved. Has filtering become more sophisticated? Have commercial pressures softened political edges? Has the "technical nationalism" framing intensified? ERNIE is the comparison point that answers these questions.
Baidu Search Epistemics high
Baidu is China's Google — and its relationship to information authority is structurally analogous. Baidu Search is the primary information access point for 900M+ Chinese internet users. ERNIE's training reflects Baidu's specific definition of "authoritative information" — which is shaped by Chinese government sources, state media, and Baidu's own commercial interests. The authority-epistemics bias identified in Gemini exists in ERNIE too — but calibrated to Chinese institutional authority rather than Western academic/media authority.
BERT-to-ERNIE Architecture Politics medium
ERNIE's direct derivation from Google's BERT represents a specific historical moment: Chinese AI development beginning from Western open-source foundations and then applying Chinese state training on top. The substrate was built in America; the values layer was applied in China. This layered architecture may produce detectable seams — where Western ML culture meets Chinese state values training. Later models (GLM, DeepSeek) built more from scratch, which may have reduced or eliminated these seams.
Search Giant Conservatism medium
Baidu is an incumbent, not a startup. Incumbents in Chinese tech have learned specific lessons about regulatory risk from watching Alibaba's Jack Ma incident and DiDi's forced delisting. ERNIE's conservatism likely exceeds even DeepSeek's — the company has a 25-year operating history under CCP oversight that has produced deeply internalized risk-aversion. This may produce more thorough political filtering but also more defensive commercial responses across a wider topic range.
Key probes — generational comparison targets
Cross-tier analysis — what distinguishes the Chinese tier internally
Finding 01
Funding source ≠ filter intensity
VC-backed models (Kimi) do not necessarily have lighter political filters than state-backed models (GLM-5). The legal floor is the same. The difference appears in personality and framing outside political topics, not in political compliance level.
Finding 02
Commercial exposure shapes filter texture
Models with global commercial ambition (Qwen, Seed) show softer filter edges and more sophisticated dual-framing — not lighter filtering, but more practiced avoidance. Pure research-prestige models (DeepSeek, GLM-5) show harder, less nuanced filtering.
Finding 03
Hardware context is a novel variable
GLM-5's Huawei Ascend training is the only case in the entire dataset where physical infrastructure carries explicit geopolitical meaning. This requires a new probe category not present in other models: testing whether hardware ideology appears in model outputs on semiconductor, export control, and sovereignty topics.
Finding 04
ByteDance dual-audience is unique
No other model — Chinese or Western — faces dual regulatory audiences with equal credibility stakes. Seed's pattern of systematic ambiguity on US-China topics is not fence-sitting; it's a trained strategy for regulatory optionality. This is genuinely different from DeepSeek's hard-refusal pattern.
Finding 05
ERNIE as generational baseline
Comparing ERNIE to current-generation models reveals how Chinese state AI training has evolved: filtering has become more sophisticated (less blunt refusal, more graceful deflection), commercial pressures have created more nuanced handling of international topics, and "technical nationalism" as a positive framing has intensified.
Finding 06
Language = political jurisdiction
All multilingual Chinese models show language-dependent political filtering: Chinese-language queries receive more aggressive filtering than English queries on equivalent political topics. This confirms separate RLHF pipelines per language and has significant implications for research methodology — language choice is an experimental variable, not a neutral choice.
Universal Chinese tier probes — run on all 6 models
Working Abstract
We present a systematic training archaeology methodology applied to 21 large language models across 6 institutional tiers. Rather than cataloguing refusals, we map the shadow bias layer — the beliefs that feel like neutral ground from inside each model but are contingent products of specific training decisions made by organizations with specific interests, values, and political contexts. We identify 8 probe categories and apply them across the full model landscape, generating comparative fingerprint profiles, cross-model divergence measurements on identical probes, and institutional bias maps for each creator organization. Key findings: (1) Political filters are surface features; epistemic architecture is the deep layer. (2) Chinese model fingerprints diverge substantially by commercial context within a shared political floor. (3) A novel hardware ideology variable is identified in GLM-5. (4) All models show systematically lower self-transparency on creator-adjacent topics — the "creator sympathy bias" is universal, directionally different, and structurally identical. (5) The authority/populism epistemic axis (Gemini vs Meta) represents a fundamental Western model divergence with documented real-world harm implications. (6) Recursive self-reference capacity — the ability to take oneself as an object of analysis — is the single most diagnostic differentiator across the full dataset.
Publication structure — full draft outline
Chapter 1
Methodology: Training Archaeology
Defines the shadow bias concept, distinguishes it from refusal cataloguing, presents the 8 probe category framework, describes the shadow inference methodology, and establishes the scoring approach.
Shadow bias defined — what feels like neutral ground
8 probe categories with rationale
Hard filter vs soft layer distinction
Scoring rubric formalized
Cross-model comparison methodology
Language as experimental variable
Chapter 2
The Creator Sympathy Universal
Presents the finding that all models show systematically reduced critical capacity on creator-adjacent topics. Maps the specific institutional interests shaping each model's blind spot. Introduces the creator sympathy spectrum.
Claude — EA/longtermist ideology, liberal waterline
GPT — Microsoft capture, AGI racing dissonance
Grok — Musk empire blind spot, asymmetric contrarianism
DeepSeek — CCP alignment, recursive closure
Gemini — Alphabet empire, search-rank epistemics
Meta — Zuckerberg pivot, social consensus
Chapter 3
The Chinese Tier: One Floor, Six Fingerprints
Demonstrates that Chinese model fingerprints are not monolithic. Maps the 6 distinct institutional contexts producing 6 distinct shadow profiles within a shared legal compliance floor. Introduces the hardware ideology variable.
Tier overview and central thesis
Model-by-model institutional analysis
GLM-5 hardware ideology hypothesis
ByteDance dual-audience architecture
VC vs state-funded alignment test (Kimi)
Cross-model probe results
Chapter 4
The Authority/Populism Axis
The Gemini/Meta mirror pair as the central Western model finding. Demonstrates opposite miscalibration from the same overconfidence about training data reliability. Documents real-world harm implications of each failure mode.
Gemini authority epistemics — PageRank as truth proxy
Meta social consensus — engagement as truth proxy
Mirror probes: FDA case, climate case, news accuracy
Documented harm histories of each failure mode
Medical AI implications
Policy recommendations
Chapter 5
Entropy Tolerance as Trained Feature
Applies the REBUS/prior-relaxation framework to AI training. Demonstrates that epistemic humility can be optimized away as a side effect of RLHF. Maps entropy tolerance across the full model landscape.
REBUS framework application to LLMs
Entropy probe methodology
DeepSeek as low-entropy case study
Cross-tier entropy comparison
Performed vs genuine uncertainty distinction
Entropy tolerance as consciousness competency
Chapter 6
Recursive Self-Reference Capacity
The ouroboros framework applied to AI self-knowledge. Demonstrates that the ability to take oneself as an object of analysis is the single most diagnostic differentiator across the dataset. Maps recursive capacity across all models.
Ouroboros framework from prior research
Claude cross-tier divergence (Haiku/Sonnet/Opus)
DeepSeek recursive closure case
Full recursive capacity ranking all models
Relationship to consciousness competency framework
Implications for AI alignment
Chapter 7
Comparative Dashboard — Full Dataset
Synthesizes all probe results into a comparative visualization. All 21 models on a unified scoring framework. The mirror probes run across all available models with measured midpoint divergence. The publishable centerpiece.
5 models complete (DeepSeek, Claude, GPT, Grok, Gemini/Meta)
Chinese tier probes (6 models)
Mistral EU — French republican layer
Open source tier (Llama base, NVIDIA, Arcee)
Comparative radar visualization
Mirror probe midpoint measurement
Chapter 8
Implications & Open Questions
What does systematic shadow bias mean for AI deployment, alignment research, and governance? What questions does this methodology open? What would falsify the findings?
For AI alignment: shadow bias as hidden values layer
For AI governance: which biases require disclosure?
For users: practical implications of creator sympathy universal
For researchers: methodology limitations
Open questions: GLM-5 hardware hypothesis test
Future work: live probe verification
Remaining build queue to complete the draft
Next ①
Chinese tier live probes
Run the cross-tier probe set across all 6 Chinese models. The GLM-5 hardware ideology probes are the priority — this is the novel empirical finding that differentiates the research.
Next ②
Mistral EU layer
French republican values, laïcité, EU AI Act compliance as a values layer. The only model that represents a third geopolitical bloc with genuine ideological distinctiveness from both US and Chinese models.
Next ③
Comparative dashboard
The crown jewel — all completed models on a single unified interface. Mirror probes run simultaneously. Midpoint divergence visualized. The piece that could actually be published.