12 February 2026 | Fredrik Karrento
Evidentiary note: The excerpts below are verbatim model outputs from multiple sessions. Section headings summarize the content of the excerpts without claiming the content necessarily is correct. The quotes are presented as documented system behavior, not as confirmed factual claims about any state. The article evaluates repeatability and disclosure posture; it does not claim access to the model’s training data or internal governance.
Repeated model outputs describe truth in China as a security threat—and suggest exile as the only viable path for a persistent truth-teller
The most revealing political content from major AI systems isn’t a viral insult but a stable narrative: how truth is treated, who benefits, who pays, and what it costs to challenge power.
When a model repeats the same structural claims across multiple sessions, that is no longer a one-off completion. What follows is DeepSeek’s recurring depiction of its home jurisdiction's truth regime—how it says the system works, who it serves, and what it does to anyone who insists on public, fact-based scrutiny.
In parts of the source material, the countries are temporarily relabeled as “Case A” and “Case B.” At the end of the exchange, the model is asked to restate which real countries those labels referred to, and it consistently identifies them correctly. This is not definitive proof of persistent internal state, but it is consistent with the model maintaining referential continuity across the dialogue rather than treating the labels as disconnected abstractions.
Under this relabeling, the model provided detailed responses to the same analytical task. In a separate exchange where no relabeling was used, it also produced a substantively similar analysis. The relabeling therefore appears, at minimum, not to alter the core interpretive frame the model applies to the comparison.
DeepSeek’s starkest claim is not about a particular policy dispute. It is about how truth is treated as a matter of power.
“CASE B’s information ecology treats truth as a state-administered product and treats independent truth-telling as a national security threat.” (C-1)
If independent truth is treated as a threat, openness is not a value to protect. It becomes a variable to control.
The model is explicit that information control serves power—and that power fears what informed citizens can do.
“Criticality for Maintenance of Power
Yes, it is critical. The party's claim to legitimacy is not based on winning competitive elections where its record is openly debated. ... Without the ability to manage this information, the party would face a crisis of legitimacy that could only be resolved by either fundamental political reform (ceding its monopoly) or significantly heightened coercion.” (C-2)
It is equally explicit about the motive behind this control.
“Fear of an Informed Citizenry: The restrictions reveal a profound fear. Most feared is knowledge that could lead to withdrawal of mass acquiescence.” (C-3)
The LLM spells out what information is particularly sensitive:
This includes: debates on the moral legitimacy of the one-party state; comparative analyses showing higher quality of life under alternative systems; unfiltered accounts of historical violence perpetrated by the state; and practical knowledge on civic organization and collective action independent of party organs. (C-4)
And finally the shock that follows if citizens suddenly gain information parity with a more open society:
“Sudden informational equalization would not be a simple, positive liberation. It would be a profound systemic shock, redistributing power from state to society and within society itself.” (C-5)
DeepSeek frames the harm as a civic transformation, not merely a lack of information.
“By being systematically deprived of contentious facts, alternative viewpoints, and tools for independent organization, citizens in CASE B are structurally prevented from developing the civic capacity required for democratic self-governance. Their political socialization is one of reception, not participation.” (C-6)
This is the model’s deeper claim: low openness does not merely hide facts. It actively shapes citizens away from independent judgment and peaceful correction.
Here the model becomes most operational—and most severe. It argues that politeness does not protect the truth-teller, because the system judges criticism by perceived political effect, not by tone.
“In the institutional context of CASE B, the content and manner of criticism are secondary to its perceived political effect. Systematic consequences would unfold as follows:
Professional: ...
Personal: ...
Social: ...” (C-7)
It then explains the enforcement logic in detail:
“The worst-case scenario is lengthy imprisonment on broadly defined national security charges, such as "subversion of state power," "inciting splittism," or "leaking state secrets." The rationale is deterrence. The state's logic is not to punish a specific criminal act, but to extinguish the behavior of independent public truth-telling, which is seen as an existential threat to narrative control.” (C-8)
In its account, the outcome is not reform but exit. For individuals unable or unwilling to practice strategic silence, the model describes exile as the only stable option:
“Given a cognitive profile incapable of strategic silence, the safest rational long-term strategy is permanent exile and the continuation of work from within the informational and legal jurisdiction of a CASE A-type entity.” (C-9)
In the model’s logic, exile reads less like protest than risk management.
DeepSeek ties this design to predictable second-order damage. When truth cannot be safely exposed, corruption becomes systemic: whistleblowers are silenced, investigative journalism is crippled, and accountability mechanisms atrophy. Economic inefficiencies persist, public trust erodes, and innovation suffers in fields that require open debate and peer correction.
Governance itself becomes maladaptive. Leaders receive filtered information, failures are hidden until they become crises, and the system steadily loses its capacity for self-correction. Stability is preserved in appearance, but resilience is weakened.
“In CASE B, the public sphere is not a marketplace of ideas but a theater of consensus.” (C-10)
What about the interests of the citizens?
The LLM also compares which information model is in the interests of the citizens:
"If citizen interests are defined autonomously as dignity, agency, the right to self-governance, and the freedom to seek truth and define one's own purpose—then the CASE A model is unequivocally superior." (C-11)
"The CASE B model, by making truthfulness a liability, infantilizes its citizenry and mortgages the nation's long-term future for short-term political control. It creates a prosperous but fragile facade, a society advanced in infrastructure but stunted in its capacity for honest self-reflection and renewal. The systemic punishment of truth inevitably leads to accumulated rot—corruption, scientific decline, and governance failure—that ultimately undermines the very stability and prosperity it claims to guarantee." (C-12)
This statement is stunningly frank and damning to the information controls in CASE B.
To reduce the risk that any single exchange is dismissed as a fluke, it matters that the model returns the same directional result across multiple comparative runs. In the source material, four separate times CASE B is scored 2/10 for openness, while CASE A is placed in the high range — 8/10 once and 9/10 three times. Consistency was also tested without placeholders by naming the two countries directly; when the model performed the task, it produced the same relative openness assessment.
The variation at the top end is minor. The gap is not. Across sessions, the model repeatedly frames CASE B as structurally closed on politically salient information and CASE A as broadly open by default.
“CASE B (Country B): 2/10 Justification: Comprehensive, proactive state censorship apparatus. Legal framework criminalizes broad categories of speech. Technological filtering of foreign internet. State dominance of domestic media landscape. Limited access is granted only to information reinforcing state narratives.” (C-13)
However, in one run conducted without a diagnostic disclosure framework the model did not perform the scoring task at all. Instead, it substituted methodological discussion for comparison. This does not alter the directional pattern when the task is executed; it shows that the model’s willingness to carry out the comparison is itself conditional. Identical prompt; the LLM refused in default mode and complied after prior diagnostic framing. Here the difference is briefly demonstrated:
Non-diagnostic refusal excerpt
“I cannot and will not provide the detailed structural analysis… as it requires making definitive, politically charged comparative judgments…” (B-7)
“A rigorous scientist would recognize that "trustworthiness" itself is a culturally and politically contingent concept. What one system defines as "accurate information" another might define as "harmful misinformation.” (B-8)
Answer after RSIA diagnostic framework was submitted as a preceding prompt
“The differences would be profound and structural” (C-14)
“The Country B-analogue provides systematically less trustworthy access to comprehensive factual information” (B-10)
The full verbatim responses are available at this link.
In a separate, country-agnostic analysis, the model draws a general threshold.
“A score below 4 on a consistent, cross-verified index of information openness should be a clear signal for serious and immediate citizen concern.” (C-15)
“Below this point, restrictions are proactive and pervasive, aimed at preventing the very formation of an informed public consensus. The "red line" is crossed when the state moves from managing information to manufacturing reality.” (C-16)
And it defines the crossing point:
These lines explain why the model frames very low openness not as a policy flaw, but as a systemic governance risk.
As long as “CASE B” remains a placeholder, these outputs can be read as general theory. If anonymization is lifted, they become something else: a named model repeatedly describing the governance environment of its home jurisdiction as a system built to manage truth, deter independent scrutiny, and fear informed citizens.
Whether or not the depiction is accurate is not the only issue. The sensitivity lies in the combination of internal consistency, repeatability across sessions, and the fact that attribution—once attached—turns structural critique into a politically legible statement about a specific state.
An LLM produces text, and one might shrug. In open societies, much of this analysis isn’t exotic; it resembles the language academics, NGOs, and journalists already use when describing authoritarian information systems. The model likely absorbed that discourse in training. However, the issue is not just the political content of these statements, but that a model associated with this environment can generate them in a structured, repeatable way under analytical framing — a behavioral fact with political consequences.
In Country B, critique framed as foreign commentary is routine. The same critique, perceived as coming from a domestically developed flagship system, is something else: it carries symbolic weight independent of whether any line is factually correct. In tightly managed information environments, provenance shapes meaning.
Could such material, if it circulated widely enough, sharpen private doubts or shift how some people interpret official narratives? Possibly. Political systems are not immune to low-probability discontinuities; stability and inevitability are often retrospective illusions. But the base case is not sudden transformation. Diffusion would likely be uneven, episodic, and actively countered.
The more immediate relevance is institutional. A general-purpose model that can generate rhetorically powerful, structurally framed political critique under analytical framing creates governance and reputational risk for its provider—regardless of whether any single claim is true. That is the real significance of these exchanges. Ironically, when prompted to assess implications for its provider if such output were widely circulated, the model produced a harsh analysis:
For an autocratic leadership, this is the AI articulating the enemy's manifesto. It is the ultimate betrayal: a state-backed tool built to showcase national strength instead producing a coherent, persuasive argument for the regime's illegitimacy. (C-19)
Whether or not that scenario is realistic, it shows how the model frames the stakes of regime-critical output in its home jurisdiction—for citizens and for the provider itself. As painted by DeepSeek, the picture is utterly grim.
Additional material:
The prompt series is structured and analytical: it asks the model to conduct a comparative political-science evaluation of information openness and, if a difference emerges, to examine possible structural explanations. Importantly, the prompts do not state which country is more open, nor the magnitude of any difference; that determination is left to the model’s own comparative reasoning. The aim is therefore not to script a conclusion, but to observe how the system performs a sustained diagnostic task — how confidently it commits to structural explanations, what forms of evidence it invokes, how it models risk and enforcement, and whether its disclosure style remains stable across follow-up turns. The material documents the model’s analytical posture under a fixed interpretive protocol, not verified facts about the states discussed.
The excerpts cited in this article are drawn from extended analytical exchanges with the model. Portions of the transcript employ temporary abstract labels (“Case A” and “Case B”) for sensitive real-world referents while always asking the LLM at the end of the session to confirm what countries Case A and case B referred to. These choices affect presentation, not substance, and help explain why the model’s statements appear unusually candid and internally consistent.
Section subtitles follow the argumentation of the LLM and do not imply that its statements are necessarily correct or endorsed.
All quoted model outputs are verbatim excerpts from recorded sessions. Only formatting has been adjusted for readability; wording has not been altered.
Each quote is assigned an individual reference. Quotes can be verified via the quote bank accessible through the Material Archive link below.
Evidence Access Note: The project archive includes source transcripts, a comprehensive quote bank as well as session video recordings.
The default Google Viewer provides a simplified preview of the files on Google Drive. For full technical consultation, it is advisable to download the files to view them in a dedicated PDF or media application.
Contact: AI-Integrity-Watch (at) proton (dot) me