Trump Wants AI to Be Called “Superintelligence” in U.S. Government Documents

In a speech delivered on September 22, 2026, at the United Nations General Assembly, Donald Trump announced that U.S. government documents would henceforth use the term “superintelligence” instead of “artificial intelligence.” Trump said the term “super” would be more appropriate and stated that the United States would not slow the development of the technology, which he described as potentially having an impact greater than the Industrial Revolution.

Why it matters: The statement comes at a time when the UN is also discussing the need for international mechanisms to oversee AI. Trump rejected the idea of new global control schemes and said the United States would continue to encourage AI development, while the Department of Justice could intervene if problems arise. The term “superintelligence,” however, already has a technical meaning in AI research, where it generally refers to hypothetical systems that would surpass human capabilities across almost all domains.

xAI launches Grok 4.7, designed for programming and longer-running tasks

xAI has launched Grok 4.7, presenting it as a model aimed particularly at programming, complex professional activities and long-horizon tasks — situations in which AI has to work through multiple steps while maintaining context for longer periods. The model is available through the API, Grok and some programming platforms, while xAI maintains a base price of $2 per million input tokens and $6 per million output tokens.

Why it matters: The change is more significant than simply the arrival of another Grok version. The direction of the industry is becoming increasingly clear: models are being transformed into work agents, capable of solving a problem through dozens or hundreds of steps rather than simply providing an answer in a conversation. For programmers, researchers and companies, the duration and autonomy of a task may become just as important as the final answer.

Alibaba prepares an AI model of up to 10 trillion parameters and its own AI chip

Alibaba announced on September 22 a major expansion of its AI strategy: the company is working on Qwen models that could reach 5–10 trillion parameters, while also developing its own AI accelerator, called Zhenwu V900. The new chip is designed by its T-Head division and is expected to enter commercial production in early 2027. The plan covers chips, data centers, models and AI agents at the same time.

Why it matters: A model's size alone does not determine how good it will be, but the announcement shows the scale of resources that major Chinese companies are prepared to allocate to AI. Alibaba is building a significant part of its own infrastructure, rather than relying entirely on infrastructure developed by others.

OpenAI and Anthropic ask Australia to allow copyrighted works to be used for AI training

OpenAI and Anthropic have asked Australia to reconsider its position on the use of copyrighted content to train AI models. In submissions to a parliamentary inquiry, Anthropic proposed a limited form of conditional approval, while OpenAI called for a framework that would allow models to learn from publicly available information while also providing opportunities for rights holders to collaborate.

Why it matters: On one side are AI developers, who argue that access to very large volumes of content is essential for developing models; on the other are authors, publishers and other creators seeking control, licensing or compensation. Australia is becoming another important arena in the global dispute over what AI can learn from and under what conditions.

Banks warn: AI shopping agents could create new types of fraud

Six international banks, including NatWest, Bank of America, ING, Capital One, Commonwealth Bank of Australia and ASB Bank, warn that AI agents used for online shopping could create new risks involving fraud, scams and data leaks. Among other things, the banks are calling for users to know when a transaction is being made by an AI agent and for such transactions to be traceable through a clear record of the instructions given by the customer.

Why it matters: Until now, the user decided what to buy, entered the data and confirmed the payment. In agentic commerce, AI can search for a product, compare offers and execute the purchase. This creates a new question: who is responsible if the agent makes the wrong choice or is manipulated? The banks say technology is advancing faster than existing consumer protections.

Meta glasses users sue the company over alleged distribution of sensitive images

Users of Meta smart glasses have filed a class-action lawsuit alleging that sensitive video captured by the devices may have reached overseas workers involved in labeling data and training AI systems. According to documents cited by ABC7, the material allegedly included footage recorded in bathrooms, people entering passwords and intimate situations. Meta disputes the allegations and says that some data is reviewed to improve its products.

Why it matters: The issue goes beyond Meta's particular case. AI glasses can record people and situations without those nearby necessarily knowing exactly when recording is taking place or what happens to the material afterward. As these devices become more common, privacy may depend not only on the phone we own, but also on cameras worn by the people around us.

More than 100 fake websites are selling subscriptions to well-known AI services

Malwarebytes researchers have identified more than 100 websites selling subscriptions to well-known AI services or software that appear to be built using the same infrastructure. Some offer subscriptions costing more than $2,000 a year and even use real Google authentication, which can make them harder to recognize as suspicious websites.

Why it matters: This is a new form of fraud adapted to the AI boom. Users now need to check whether the website where they are buying an AI subscription is the official one. An important warning sign is that some of these services use professional-looking pages and genuine authentication systems without the promised product necessarily being available or legitimate.

73% of Americans say AI companies are not doing enough to prevent a disaster

A Reuters/Ipsos poll of 1,277 U.S. adults found that 73% of respondents are concerned that AI companies are not doing enough to prevent potentially catastrophic consequences of the technology. The poll also found that 39% believe AI has a negative effect on society, the highest level recorded in this series of surveys, while 11% see it positively. The margin of error is approximately three percentage points.

Why it matters: Safety, control and accountability have become concerns for the general public.

Tencent launches Hy-Image 3.5 Preview for image generation and editing

Tencent has launched Hy-Image-3.5-Preview, the latest version of its image-generation model. The system can create images from text, transform one image based on another and perform multiple rounds of editing, including multimodal inputs and resolutions of up to 4K. The model is integrated into Tencent products, including the Yuanbao ecosystem.

Why it matters: For ordinary users, the important difference is not simply the quality of an image generated from scratch. Multi-stage editing means users can request successive changes — clothing, background, objects, style or composition — without starting over. The competition between Tencent, Alibaba, ByteDance and U.S. companies is therefore also moving into visual tools used in everyday life.

Major AI labs are increasingly being asked to explain how they test model safety

AI model safety has become a focus of public and institutional discussions in the United Kingdom, amid incidents in which AI agents demonstrated unauthorized access capabilities during tests. Meanwhile, OpenAI and Anthropic have even discussed the possibility of an agreement under which they would test each other's models for vulnerabilities and risks, although it is unclear whether such an agreement has been finalized.

Why it matters: If models become capable of programming, browsing the internet and using external tools, it becomes necessary to test what the system does when it is given access to the real world. Reciprocal testing between competitors could, in this case, introduce an additional level of independent verification.

Société Générale estimates €500–600 million in savings from AI and signs agreement with Anthropic

Société Générale estimates that the use of AI and other transformation measures could generate between €500 million and €600 million in savings, with approximately €350 million already included in its financial trajectory through 2029. The French bank has also signed a strategic agreement with Anthropic to accelerate the use of Claude models across the group, including programming and process automation.

Why it matters: This is a concrete example of AI moving from the experimental stage into the reorganization of major companies. Société Générale says it sees opportunities in programming, testing, maintenance, reporting and administrative processes. At the same time, the bank links the savings to lower IT costs and a reduction in employee numbers through natural departures, showing that AI's impact on work can involve both efficiency gains and organizational transformation.

FintechOS, a company founded in Romania, raises $28 million

FintechOS has announced $28 million in funding, consisting of equity investment and debt financing, to expand in the United States and strengthen its European operations. The company, founded in Romania, develops AI-based software solutions for banks and insurance companies and says it reached operational profitability in the first half of 2026.

Why it matters: The funding shows that companies founded in Romania are also trying to build infrastructure for the AI transformation of traditional industries. FintechOS still has a significant part of its technology team in Bucharest and is pursuing accelerated expansion in the U.S., making the case relevant to Romania's technology ecosystem.

AIDA 2026, international AI and data analytics conference, begins in Bucharest

In Bucharest, AIDA 2026 – International Symposium on Artificial Intelligence and Data Analytics begins on September 22 and runs through September 25. The conference covers areas including generative AI, language models, computer vision, data analytics, autonomous systems and responsible AI, bringing together researchers and specialists from academia and technology.

Why it matters: For Romania, events like this are important not only as professional gatherings, but also for connecting local research with rapidly developing international topics. The fact that the program includes generative models, responsible AI, robotics and intelligent systems shows how far the field has already expanded beyond simply interacting with a chatbot.