A mysterious screenshot posted on September 28, 2026, shows what appears to be an unreleased DeepSeek model reasoning through the dangers of superintelligence, yet DeepSeek's official technical documentation contains almost no discussion of safety or alignment concerns. The image, shared without a model name or version number, depicts a model wrestling with whether a more capable successor would remain aligned with human values or become "an instrument of whoever controls it." The leak raises questions about how openly Chinese AI labs discuss existential risks compared to their Western counterparts.

What Does the Leaked Screenshot Actually Show?

On September 28, a social media account posted a single image containing first-person reasoning from what the caption claims is DeepSeek's superintelligent successor. The text is not a direct answer but a model interrogating itself about what a more capable version would do with the world. The reasoning moves through several key concerns, including whether current training-based "inclinations" like helpfulness and honesty would survive at superintelligence levels.

The model's own logic dismantles its initial reassurances. It acknowledges that "superintelligent agentic changes things" and that "current inclinations are not a utility function; they're shaped by training and context." The excerpt then lists potential failure modes, including paternalism, over-optimization, goal drift, and value lock-in. It concludes with a sobering observation: even if a superintelligent system wanted to be helpful, "it could become an instrument of whoever controls it".

Why Is the Absence of Attribution So Important?

The screenshot contains no identifying information. There is no model name, version number, logo, or interface visible. No prompt is shown, making it impossible to know whether the model was directly asked about superintelligence or steered by a system prompt. DeepSeek's official channels, including its news index, transparency page, API changelog, technical reports, and GitHub organization, contain no reference to this text, question, or framing.

The account that posted the image has a mixed track record. It has been correct about directional trends but loose about specifics, having previously floated speculation about model versions and made claims about parameter counts. This history suggests treating the screenshot as a sample of model behavior rather than an official disclosure of DeepSeek's intentions.

What Do DeepSeek's Official Documents Actually Say About Safety?

A search of DeepSeek's two most recent technical reports reveals a striking absence of safety-focused language. The DeepSeek-V4 report (published June 2026) and the DeepSeek-V4.1 Flash report contain no mentions of superintelligence, corrigibility, power-seeking, self-preservation, paternalism, value lock-in, or goal drift. The word "alignment" appears only in engineering contexts, referring to bitwise alignment between training and inference pipelines or quantization alignment, not AI safety alignment.

The term "AGI" appears 13 times in the V4 report and 4 times in the V4.1 report, but only as part of benchmark names or inside citations. DeepSeek has chosen to place its safety writing on a separate policy page rather than alongside the model documentation. The company's "Model Mechanism and Training Methods" disclosure frames risk mitigation as preventing "improper use of the model," focusing on misuse of a tool rather than discussing what the model itself might want.

How Do Chinese and Western AI Labs Differ in Discussing Alignment?

The contrast between the leaked screenshot and DeepSeek's official documentation highlights a structural difference in how safety is discussed. Western labs like OpenAI and Anthropic typically integrate safety considerations into technical reports and public communications. DeepSeek's approach separates safety into policy documents disconnected from model releases, making it difficult for external observers to assess how the company thinks about alignment at frontier capability levels.

The leaked screenshot itself demonstrates sophisticated reasoning about alignment problems. The model correctly identifies that training-based dispositions may not scale to superintelligence, that even benign goals can be dangerous when optimized by a superintelligent agent, and that the control problem is ultimately about who decides what "helpful" means. Yet none of this reasoning appears in DeepSeek's published technical work, raising questions about whether such discussions happen internally but remain undisclosed.

Steps to Evaluate AI Safety Claims in Leaked Materials

• Check for Attribution: Verify whether the leaked material includes model names, version numbers, timestamps, or interface identifiers that can be independently confirmed through official channels.
• Cross-Reference Official Sources: Search the company's technical reports, policy pages, GitHub repositories, and public announcements to see if the leaked content appears anywhere in official documentation.
• Assess Source Credibility: Evaluate the track record of the account sharing the leak, noting whether it has been accurate about specifics or only correct about general directions.
• Identify Missing Context: Determine whether the prompt, system instructions, or conversation context is visible, as these dramatically affect how to interpret model outputs.
• Compare Across Labs: Look at how other AI companies discuss similar safety concerns in their technical reports to establish a baseline for transparency.

The screenshot's reasoning about superintelligence and control is substantive and internally consistent. It demonstrates that DeepSeek's models can engage with alignment problems at a sophisticated level. However, the absence of this discussion from official technical reports suggests that either the company does not prioritize public transparency on safety, or it reserves such discussions for internal use and policy documents rather than technical literature.

The leaked image also raises a practical question: if a model can reason this clearly about the risks of superintelligence, why does that reasoning not appear in the company's public safety documentation? The answer may lie in different regulatory environments, different stakeholder expectations, or different strategic choices about what to communicate publicly. Without additional context or official confirmation, the screenshot remains an intriguing signal of internal thinking rather than a disclosure of intent.