OpenAI and Anthropic Restrict AI Model Distillation

Chinese researchers are using knowledge distillation to train smaller domestic AI models with outputs from leading U.S. systems. Examples include Shanghai Jiao Tong University researchers turning about 5,000 responses from OpenAI’s o1 into 10,750 bilingual examples for Alibaba’s Qwen2.5-72B in November 2024; Alibaba researchers using GPT-4 and GPT-4o to score and filter data in 2025; and a February 2026 project in which researchers from Shanghai Jiao Tong and Meituan used GPT-4o to create coding plans. Distillation is used internationally and is not illegal, but OpenAI and Anthropic prohibit using their models to train competitors. U.S. officials have raised concerns about large-scale use, and some researchers warn that dependence on the method could constrain original breakthroughs.

A separate fictional exercise presented at the Munich Security Conference depicts a dissident group using AI tools to design proteins, predict virus interactions with human cells and assemble DNA fragments to create a more dangerous enterovirus D68. Its imagined outbreak reaches 13 countries and eventually causes hundreds of millions of infections and millions of deaths before a vaccine is available. Scientists involved say the scenario reflects a realistic threat from current and near-term capabilities. The exercise omits details about how the group would spread the virus to avoid enabling misuse.

AI’s security implications also surface in a report on crypto. A theft from Binance, described as the largest in crypto history, began after an engineer installed a tampered package while looking for an AI coding tool. Attackers later used an internal Binance AI tool to reach other engineers and the wallet-signing service; the FBI attributed the theft to North Korea. Investigators have not shown that AI planned or conducted the attack, though attackers are adopting AI. The article also argues that blockchain-based assets can provide scarcity and durable records in a world of generated digital content, while a test found four of six AI models lost money trading crypto perpetuals; the worst lost about 60% in a little over two weeks.

Governments and educators are building practical AI skills. Andhra Pradesh launched the Vidya Vahini learning portal on November 11, National Education Day, with two free courses—Introduction to Artificial Intelligence and Applications of Artificial Intelligence and Quantum Technology. Each carries four academic credits and includes 40 modules. The state aims to reach more than 11 lakh higher-education students; the portal was developed with Acharya Nagarjuna University and support from Rashtriya Uchchatar Shiksha Abhiyan. In Indonesia, the Ministry of Creative Economy held a Banyuwangi workshop on November 16, 2026, for 80 to 100 developers and other participants. The CHAT 2026 program covered AI-assisted app building, business planning, dashboards and security testing, with suggested projects including tourism at Ijen, local MSME marketing and health-service access.

In the United States, Wilbur Wright College students Dakota and Lisa presented a Titanic passenger-survival machine-learning project at the National Summit on Artificial Intelligence, where they were the only student presenters among more than 200 attendees. Through the college’s IT club, they learned data cleanup, Amazon Web Services, machine learning and AI. Their professor, Gustavo Alatta, also presented research on AI infrastructure and STEM education. Separately, the Center for Digital Government’s 2026 Digital Cities Survey highlighted Corona, California, which cut permit plan-review time by 97% using AI, and Normal, Illinois, which uses drones for police, emergency response and public works. Cybersecurity remained cities’ top IT priority for the 12th year; 88% place AI governance under CIO or enterprise IT leadership.

AI adoption is also changing operations and prompting questions about human judgment and infrastructure. Healthcare service SuperDial says its AI agents have completed more than 10 million insurer calls, but it will now measure resolved requests rather than call volume. It says successful calls using payer voice automation average about four minutes, versus at least 17 minutes to reach a human representative. Its longer-term plan for agents to exchange trusted information directly remains unfinished and would require payer confidence in outside agents handling coverage decisions. In journalism, a University of Maryland study found about 9% of major newspapers used AI to create news content as of October 2025. A writer argues that synthetic narration can flatten reporting and that AI can introduce errors, harm, bias and accountability problems.

Growing demand from AI data centers is complicating Illinois’ pollution targets. Wind and solar construction is not keeping pace, and a state report recommends keeping 11 oil- or gas-fired peaker plants operating until 2033 to help prevent blackouts, rather than closing them in 2030 or 2031. Advocates criticize the recommendation, while state officials say they are not advising new gas plants and point to nuclear power and plant upgrades as sources of lower-carbon electricity. Lawmakers may debate data centers and power demand next month.

Key Takeaways

  • Chinese researchers used outputs from OpenAI models, including o1, GPT-4 and GPT-4o, to help train or refine smaller systems such as Alibaba’s Qwen; OpenAI and Anthropic terms prohibit using their models to train competitors.
  • Researchers warn that reliance on distillation could limit original AI breakthroughs, while U.S. officials have raised concerns about large-scale use of American model outputs.
  • A fictional Munich Security Conference exercise portrays AI-assisted engineering of a more dangerous enterovirus D68; participating scientists describe the scenario as realistic, while withholding transmission details.
  • The FBI attributed the Binance theft—initiated after an engineer installed a tampered package sought as an AI coding tool—to North Korea. Investigators have not shown that AI planned or ran the attack.
  • Andhra Pradesh’s Vidya Vahini portal offers two free, four-credit AI and quantum-related courses, each with 40 modules, and aims to serve more than 11 lakh higher-education students.
  • Indonesia’s CHAT 2026 workshop in Banyuwangi trained 80 to 100 participants in AI-powered app development, including security testing and potential projects in tourism, local business and health access.
  • SuperDial reports more than 10 million insurer calls by its AI agents and says automated successful calls average about four minutes, compared with at least 17 minutes to reach a human representative.
  • Wright College students presented a Titanic survival-prediction machine-learning project at a national AI summit; they were the only student presenters among more than 200 attendees.
  • Illinois’ state report recommends keeping 11 oil or gas peaker plants open until 2033 amid rising data-center electricity demand; officials say they are not advising new gas plants.
  • Corona used AI to reduce permit plan-review time by 97%; the 2026 Digital Cities Survey also found cybersecurity remained the top city IT priority for the 12th consecutive year and 88% place AI governance under CIO or enterprise IT leadership.

Alibaba and Meituan Researchers Use AI Distillation to Train Smaller Models

Researchers from Alibaba and Meituan used outputs from leading U.S. AI models to train smaller domestic systems. Their method, called knowledge distillation, transfers information from a large model to a smaller one, which can make training faster and more efficient. A recent study found that the approach improved model performance and efficiency. Chinese developers are increasingly exploring distillation as they seek more efficient training and address concerns about data security.

Chinese AI Firms Use OpenAI Models to Train Smaller Systems

Chinese researchers and technology companies are using outputs from advanced models, including OpenAI’s, to train smaller domestic AI systems through a process called distillation. Examples include Shanghai Jiao Tong University researchers using about 5,000 responses from OpenAI’s o1 model to create 10,750 bilingual examples for Alibaba’s Qwen2.5-72B model in November 2024. In 2025, Alibaba researchers used GPT-4 and GPT-4o to score and filter data for smaller Qwen models, and in February 2026 researchers from Shanghai Jiao Tong University and Meituan used GPT-4o to produce coding plans for training a domestic model. The article says distillation is not illegal and is used worldwide, but OpenAI and Anthropic terms prohibit using their models to train competitors. U.S. officials have raised concerns about large-scale use of American AI outputs, while researchers have warned that relying too heavily on distillation could limit original breakthroughs.

Scientists Warn AI Could Help Create a Deadly Biological Weapon

A fictional scenario presented at the Munich Security Conference describes a dissident group using AI tools to engineer a more dangerous version of enterovirus D68. The tools in the scenario help design proteins, predict how a virus could interact with human cells, and assemble DNA fragments. The imagined outbreak spreads to 13 countries, with the scenario later describing hundreds of millions of infections and millions of deaths before a vaccine becomes available. The story is fiction, but scientists involved in the study called the threat realistic based on current and near-term AI capabilities. The exercise avoided explaining how the fictional group would spread the virus because that information could be misused.

Crypto’s Role in an AI World Goes Beyond Automated Trading

An October 29 theft from Binance, described as the largest in crypto history, began after an engineer installed a tampered package while searching for an AI coding tool. Attackers later used an internal Binance AI tool to reach other engineers and the wallet-signing service, and the FBI attributed the theft to North Korea. The article says investigators have not shown that AI planned or ran the attack, but notes that attackers are adopting AI. It argues that Bitcoin and other blockchain-based digital assets can offer scarcity and records that cannot be changed in the way generated digital content can. It also cites a test in which four of six AI models lost money while trading crypto perpetuals, with the worst losing about 60% in a little over two weeks.

Andhra Pradesh Opens Free AI and Quantum Courses Through Vidya Vahini

The Andhra Pradesh State Council of Higher Education launched the Vidya Vahini digital learning portal on November 11, National Education Day. Its first phase offers two free courses, Introduction to Artificial Intelligence and Applications of Artificial Intelligence and Quantum Technology, each worth four academic credits and made up of 40 modules. The state aims to extend AI literacy to more than 11 lakh higher education students, while faculty can take courses or help develop them. The portal was developed with Acharya Nagarjuna University, Guntur, with support from Rashtriya Uchchatar Shiksha Abhiyan, and offers learning materials, assessments, credit tracking and other course tools.

AI narration raises concerns about journalism’s human touch

The writer argues that generative AI should not take part in journalism, describing how an AI voice made a heartfelt story from The San Francisco Chronicle sound dull and monotonous. Everlit provides AI narration to newsrooms, and a University of Maryland study in October 2025 found that about 9% of major newspapers use AI to create news content. The writer says AI can make mistakes, cause harm, reflect bias, and make it harder for journalists to be accountable. As an editor at Middle Tennessee State University’s student newspaper Sidelines, the writer urges student journalists to take a firm stance against AI and says human reporting still matters.

Indonesia trains regional developers to build AI powered apps

Indonesia’s Ministry of Creative Economy held an AI app development workshop in Banyuwangi, East Java, on November 16, 2026, through its Creative Hub for App Talent (CHAT) 2026 program. The event brought together 80 to 100 app developers, IT community members, academics, and industry participants. Attendees practiced building website apps and learned about vibe coding, business planning, product catalogs, dashboards, and security testing. Ministry director Tri Wahyudi said AI can speed up coding, design, and testing, but developers must understand user needs and build safe, useful products. Suggested project areas included helping tourists visit Ijen, marketing local MSME products, and improving access to health services.

SuperDial shifts focus from insurer calls to resolved requests

SuperDial says its AI agents have completed more than 10 million calls to insurers, and the company will now measure success by resolved requests rather than call volume. The service handles revenue cycle tasks for providers, health systems, RCM companies, and payers, first trying electronic systems, portals, or fax documents before making a call. When a call is needed, agents use information already gathered and ask only for what is missing. SuperDial says successful call attempts with payer voice automation average about four minutes, compared with at least 17 minutes when reaching a human representative. Its longer term goal is an AI network where agents share trusted information directly, but that plan is unfinished and would require payers to trust outside agents with coverage decisions.

Wright College students present Titanic machine learning project

Wilbur Wright College students Dakota and Lisa presented a machine learning project at the National Summit on Artificial Intelligence alongside Computer Information Systems Assistant Professor Gustavo Alatta. More than 200 educators, industry leaders, and innovators attended the summit, and the students were the only students to present. Through Wright’s IT club, they learned data cleanup, Amazon Web Services, machine learning, and AI, then built a model using Titanic passenger information to predict survival. Lisa said the presentation helped her make professional connections and overcome a fear of presenting. Alatta also shared research on infrastructure for AI development and STEM education, and the students are planning their next project.

Illinois data center power demand could keep polluting plants open

Growing electricity demand from AI data centers is making it harder for Illinois to meet its goals for reducing pollution and fighting climate change. Wind and solar power are not being built fast enough to meet that demand, while the state still has seven coal plants in central and southern Illinois. A state report recommends keeping 11 oil or gas peaker plants running until 2033 to help prevent blackouts, rather than closing them in 2030 or 2031 as expected. Advocates criticized the recommendation, including Mayra Mendez of Clean Power Lake County, who said Waukegan’s peaker plants create strong smells and a burning feeling in the nose. State officials are not advising new natural gas plants and say Illinois will rely on nuclear power and plant upgrades for more low carbon energy; lawmakers may debate data centers and power demand next month.

Digital Cities Survey highlights practical technology gains in 2026

The Center for Digital Government named its 2026 Digital Cities Survey winners on November 17, marking the survey’s 25th anniversary. Corona, California, used AI to cut permit plan review time by 97%, while Normal, Illinois, uses drones to support police, emergency responders and public works teams. Louisville, Kentucky, is inviting innovators to use AI to solve government problems. The survey found cybersecurity remained the top IT priority for the 12th year, and 88% of surveyed cities placed AI governance under CIO or enterprise IT leadership. It also reported that more than 80% offer skills development and certification programs, and over 60% have implemented advanced data analytics and low-code or no-code tools.

Sources

NOTE:

This news brief was generated using AI technology (including, but not limited to, Google Gemini API, Llama, Grok, and Mistral) from aggregated news articles, with minimal to no human editing/review. It is provided for informational purposes only and may contain inaccuracies or biases. This is not financial, investment, or professional advice. If you have any questions or concerns, please verify all information with the linked original articles in the Sources section below.

AI knowledge distillation OpenAI Anthropic Alibaba Qwen AI model training AI biosecurity AI-enabled cybercrime Binance North Korea AI education AI workforce training Machine learning AI in government AI governance AI customer service AI in journalism AI infrastructure Data centers Energy demand Cybersecurity

Comments

Loading...