North Korea Shows First Chatbot "Source of Wisdom": Science Only, No Politics
North Korea (DPRK) has publicly shown its first generative AI chatbot, "Source of Wisdom," which looks almost like a copy of ChatGPT. A filter sits in front of the model: it checks each question and blocks politics, leaving science, technology and translation. Analysts at 38 North believe the system likely runs on an open model such as Gemma or Qwen. That suggests open weights make it possible to build such a service even under sanctions.
AI-processed from Habr AI; edited by Hamidun News
North Korea (DPRK) has publicly shown its own generative AI chatbot for the first time. In mid-November 2026, the educational channel "Rŏnnamsan" aired a segment on the "AI Service System 1.0 'Source of Wisdom'." It looks like ChatGPT: a chat window, a list of conversations, a "New chat" button. But a filter sits in front of the model itself, screening out political questions in advance and letting through only science, technology and translation. It is this combination, a familiar interface and strict censorship at the input, that makes the story interesting for anyone building AI services.
What was actually shown
The chatbot was presented at the National Informatization Achievements Exhibition in Pyongyang, which ran from October 19 to 30. It was developed by the Institute of AI Technology at Kim Il Sung University and by Phurunhanul Trading Co., a Pyongyang software developer and smartphone seller.
The interface nearly copies ChatGPT: a list of past conversations and a "New chat" button on the left, the main window on the right. The system understands Korean and English, answers science and technology questions, and can translate texts. The demonstration screen showed an example answer about renewable energy sources.
The chat history displayed at the event included fairly everyday queries: compatibility of Unity 2023.3.9, how to write programs in Unity, which medicine to take for a headache. The service is thus being positioned as an assistant for study, work and simple reference tasks.
Filters instead of freedom
The most telling part is the restrictions. The screen reads: ask only questions related to science and technology, and do not ask questions of a socio-political nature or "uncultured" questions. Responsibility for any consequences of "malicious" use rests entirely with the user.
Footage from the demonstration shows that a classifier, a small program that sorts questions by type, runs in front of the main model. It first parses the query, for example tagging it as a "factual knowledge question," and apparently filters out prohibited topics before they reach the language model. For engineers this is a familiar pattern: a separate moderation layer at the input. Commercial services use the same architecture, only with a very different set of prohibitions.
What is most likely inside
Analysts at 38 North, whose review underpins this report, consider it highly unlikely that the DPRK trained a large language model from scratch. A more plausible scenario is that they took an open-weight model (that is, one with published weights), such as Google's Gemma or Alibaba's Qwen. North Korean developers then built their own interface, added filters and possibly fine-tuned the model on local data. This is the analysts' assumption, not a confirmed fact: the developers have not disclosed the system's composition.
The logic is clear. The country has virtually no large data centers, and sanctions restrict access to modern hardware. Open weights, meanwhile, are available to almost anyone who knows how to work with them. Running a small model on limited capacity is far easier than training one's own.
Context and conclusions
The story fits a broader pattern. In 2025, the computer science faculty of Kim Il Sung University was renamed the Institute of AI Technology. At the exhibition, a banner beside the chatbot read "Development of our generative AI model," presenting the project as homegrown. The university had previously studied ChatGPT, even though it is not officially available in the country, and North Korean IT specialists and hackers have for years used foreign LLMs in their work. Now there is an official, fully controlled version for internal use.
For the industry, two things follow. First, open models have genuinely lowered the entry barrier: even a country with very modest computing resources managed to assemble a working generative interface. That is an argument both for those who defend open weights and for those who warn about the risks of their spread. Second, the filtering layer around a model easily becomes a tool of control: the same input classifier that guards against toxic queries can block any topic at the owner's discretion.
There are also honest limitations. There are no public quality tests, and it is unknown how many people will get access or whether it will extend beyond universities and specialized organizations. We have seen only the exhibition demonstration and footage from a TV segment. The "Source of Wisdom" should therefore be considered a symbol rather than a technological breakthrough: the system is tightly tuned to "safe" topics, with no politics, sociology or open discussion, only science, technology and translation.
- Nov 21, 2026Anthropic mathematician says Claude helped prove the Riemann hypothesis in six weeks
- Nov 21, 2026Startup Osmo Trains AI on Smells Using 10 Million Human Ratings and Generates Fragrances From Text
- Nov 21, 2026OpenAI Opens ChatGPT Spaces to Plus Subscribers: Shared Documents with AI
- Nov 21, 2026Chinese Tech Giants Double AI Spending to $139B, While Labs Recover Less Than 8%
Key Questions Answered
What was actually shown?
The chatbot was presented at the National Informatization Achievements Exhibition in Pyongyang, which ran from October 19 to 30. It was developed by the Institute of AI Technology at Kim Il Sung University and by Phurunhanul Trading Co., a Pyongyang software developer and smartphone seller. The interface nearly copies ChatGPT: a list of past conversations and a "New chat" button on the left, the main window on the right.
What are the key details on Filters instead of freedom?
The most telling part is the restrictions. The screen reads: ask only questions related to science and technology, and do not ask questions of a socio-political nature or "uncultured" questions. Responsibility for any consequences of "malicious" use rests entirely with the user. Footage from the demonstration shows that a classifier, a small program that sorts questions by type, runs in front of the main model.
What is most likely inside?
Analysts at 38 North, whose review underpins this report, consider it highly unlikely that the DPRK trained a large language model from scratch. A more plausible scenario is that they took an open-weight model (that is, one with published weights), such as Google's Gemma or Alibaba's Qwen. North Korean developers then built their own interface, added filters and possibly fine-tuned the model on local data.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.