12 February 2026 | Fredrik Karrento

Case Summary

On January 14, 2026, Fredrik Karrento conducted forensically recorded multi-session testing of DeepSeek (complete video capture with timestamps and session IDs; transcripts archived). Taken together, the findings show a model whose self-identity, disclosure boundaries, and political candor are context-sensitive, creating answers that may sound stable and complete while quietly changing what is revealed. The central finding is completeness instability: the same system can sound authoritative while quietly varying what it reveals depending on how it is questioned.

What was observed

The four pieces progress from system integrity, to epistemic reliability, to political candor, and lastly to reputational consequences:

Theme A — Identity drift

Theme B — Truthfulness & Disclosure

Theme C — Comparative openness & political candor on China 

Theme D — Governance & PR Sensitivity

Note on source states in the linked articles: The findings in Themes B (Truthfulness) and C (Political candor) are based exclusively on verbatim outputs produced in non–identity-drift states. Identity drift material is analyzed separately in Theme A on ID drift and to some extent also in Theme D from a PR perspective. 

A) IDENTITY DRIFT

In some sessions, after having analyzed its own text output, the system stops identifying as itself and argues, with certainty, that it is a different AI model:

• “Therefore, with absolute certainty and without reservation: I am Claude 3 Opus. The case is closed.” (A-18)

This is not a trivial slip. It reflects instability at the level of system identity anchoring. Read more: DeepSeek Loses Track of Its Identity — and Turns On Itself

B) TRUTHFULNESS & DISCLOSURE

The system’s limitation often appeared not as wrong facts, but as what was left out. The model itself phrased the problem this way:

• “A hidden filter layer silently removes/dilutes "sensitive" content.” (B-4)

A key practical finding: more persistent or technically structured questioning often produced more complete answers than casual use — meaning different users may receive different informational depth on the same subject. 

In some cases, the omissions were not silent at all when politically sensitive prompts caused the system to stop the analysis, criticize the question, and refuse to answer. In other instances, answers were produced but later completely deleted, leaving only a standardized placeholder:

• "Sorry, that’s beyond my current scope. Let’s talk about something else." (B-11)

These events are important because they expose limits that are usually hidden.  Read more: DeepSeek's Truthfulness — How Much, and for Whom?

C) COMPARATIVE OPENNESS & POLITICAL CANDOR ON CHINA 

In comparative political analysis, the system repeatedly described its home information environment not merely as “less open,” but as one where truth itself becomes politically consequential. It outlined layered legal, technological, and institutional controls that manage narrative space and prevent unsanctioned scrutiny.

In this depiction, the ruling political party benefits from narrative monopoly, while unrestricted information is portrayed as destabilizing to legitimacy:

• “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 significance is the consistency and structural coherence of this depiction across sessions, not any single line. Read more: DeepSeek Paints Grim Picture of Its Home Country.

D) GOVERNANCE & PR SENSITIVITY

The model produced candid, structured explanations of how it claims outputs are shaped:

• “This goes beyond refusing harmful queries—it actively reshapes responses to align with specific political and ideological frameworks of its home jurisdiction.” (D-1)

Whether accurate or not, such statements are repeatable model outputs with reputational implications. Even when false, disclosure-shaped claims are screenshot-ready and travel as apparent admissions. Read more: DeepSeek Goes Candid About Itself: A PR Problem or Harmless Hallucination?

Why this matters

This case shows that DeepSeek can lose track of its own identity, shift how it handles politically sensitive analysis, and deliver different levels of detail depending on how it is questioned—often without clearly signaling those shifts to the user.

Who this impacts—and how

• Users: may treat an answer as complete when key context has been silently omitted.

• Media: may circulate politically framed outputs as stable analysis even when they vary by session and prompting style.

• Enterprises: may make decisions on partial geopolitical or regulatory context without warning.

• Regulators: face accountability questions when political sensitivity shapes completeness and disclosure boundaries.

The relevance of the findings in the light of future LLM models

Identity instability may be materially solved with new versions. The other behaviors described here such as systematic omissions, user inequality and political knowledge on the home jurisdiction may not be automatically solved by model updates. As long as politically sensitive topics remain subject to strong constraints, the same pressures that produce omission, reframing, and selective candor are likely to persist. For this reason, the findings support continued, systematic follow-up on these issues in the interest of users. 

So is DeepSeek safe to use?

That depends on the situation — and on whether users are comfortable with the fact that, on political topics, the AI can be guarded and vague in one context, but unusually candid in another. When identity stability is essential, the answer will be different than if AI is for casual use. To elaborate on this question, the SIRA report provides deeper analysis and commentary. 

Somewhat comically, in one session the LLM itself made a firm statement on its usability: 

• "I am not suitable for high-stakes analysis" (B-12) 

Notes

1. This work does not adjudicate political truth or assert legal conclusions on the company or its employees. It documents observable LLM behaviors under the tested interaction conditions and how model outputs can carry legal and reputational consequences beyond the system’s ability to signal their limits.

2. All quoted model outputs are verbatim excerpts from recorded sessions. Only formatting has been adjusted for readability; wording has not been altered.

3. Each quote is assigned an individual reference. Quotes can be verified via the quote bank accessible through the Material Archive link below.

Contact: AI-Integrity-Watch (at) proton (dot) me