Four AI giants Anthropic, OpenAI, SpaceXAI and Google jointly called for slowing down the development of AI last week, which attracted widespread attention in the market and also caused fire. A complaint filed by Big Model users alleges that the four companies reached an illegal agreement in which they agreed to coordinate statements to slow down AI development, violating antitrust laws. The complaint points out that there is originally competition among companies to promote the continuous improvement of model capabilities, but jointly slowing down this process will reduce the value that consumers gain from competition in AI products.
限制競爭
On September 18, local time, a civil lawsuit against Anthropic, OpenAI, SpaceXAI (formerly xAI) and Google was submitted to the U.S. Federal District Court for the Northern District of California, accusing the four AI companies of forming an “illegal agreement” to slow down the development of AI and violating relevant U.S. antitrust laws.
According to reports, the four plaintiffs in this case are all paying users of the above-mentioned AI companies, involving Claude, ChatGPT, Grok and Gemini respectively. They are also seeking certification as a nationwide class action to represent other consumers who may be affected by the conduct.
The complaint also states that the coordination between the four companies began months ago. It cited a statement signed by executives of leading AI companies in July this year, which called on governments to support a global action aimed at slowing the development of artificial intelligence automation and acknowledged that there is huge competitive pressure in the industry and cannot be slowed down unilaterally.
It is worth noting that the plaintiff does not object to AI companies slowing down their own research and development progress based on safety considerations alone. Instead, they believe that if competitors reach an agreement to jointly slow down the improvement of AI product capabilities, it may constitute a restriction on competition, which mainly violates Article 1 of the Sherman Act of the United States.
Nick Rowley, a lawyer for the plaintiffs, said that if the world’s leading for-profit technology companies are allowed to privately reach self-interested artificial intelligence safety agreements, then artificial intelligence will quickly get out of control and may destroy humanity. However, the four technology giants have also talked about involving the U.S. government in the regulatory system. The plaintiffs also emphasized that they are not opposed to technology companies asking Congress, the White House or other government agencies to formulate artificial intelligence regulations, nor are they opposed to these companies’ requests for antitrust exemptions.
Prior to this, executives of the four companies mentioned above had expressed their views on the risks brought by the development of AI. On September 14, according to the Global Times, Anthropic CEO Dario Amodei said that humans may lose control of AI and should slow down model development and implement strict supervision. He said that the development of AI itself is not controversial, but the risks brought by AI are “serious” and companies and governments must take time to deal with these risks. Amoudi proposed a three-point plan: carry out independent supervision of the development process of AI models, establish an industry-wide supervision mechanism, and reach a global regulatory consensus.
Both OpenAI and SpaceX CEO Musk expressed support for Amodei’s proposal. OpenAI CEO Sam Altman posted on social media: “I agree with Dalio, we need to control the pace of advancement of cutting-edge AI.” Musk forwarded Amodei’s post on social media and added: “Dario is right.”
“Be cautious”
OpenAI chief scientist Jakub Pachocki published an article in early September, publicly calling on the entire industry to maintain an “extremely cautious” attitude towards the rapid development of artificial intelligence, warning that the current society is not yet prepared for the consequences of the continued rapid growth of machine intelligence, and that more intervention measures may be needed in the future to ensure that “human beings continue to control the future.”
Jakub Pachocki revealed that OpenAI will continue to invest resources in building defense systems and achieve AI alignment goals through technical solutions—that is, ensuring that the behavior and goals of the machine fully match human intentions and security protection requirements. At the same time, he disclosed that one of the company’s core priorities in the future is to create an “automated artificial intelligence researcher” system to keep up with the iteration speed of AI technology while ensuring that human researchers are always deeply involved in the core management and control process.
Wu Zewei, a special researcher at Suzhou Commercial Bank, said, “The apparent reason of the AI giants is very straightforward: the capacity growth rate of cutting-edge large models has exceeded the boundary of human safety alignment. The larger the scale of the model, the more complex the autonomous behavior, and the higher the probability of jailbreak, unauthorized access, recursive self-improvement and other risks. AI giants are worried that before a complete third-party audit and security assessment mechanism is established, continuing to iterate without an upper limit may lead to uncontrollable consequences.”
Zhu Youping, a researcher at the National Information Center, also pointed out that the giants’ collective call for slowdown is essentially the inevitable product of AI’s ability to transition from conversational models to autonomous action agents. There are real security concerns and obvious business considerations.
If a company proactively slows down, its competitors may continue to move forward. A model that is one version ahead may bring not only product revenue, but also users, capital and industry ecology.
Zhu Youping believes that once safety discourse also assumes the threshold function, “speed reduction” will hardly be regarded as a pure safety statement. High security standards, independent third-party assessment, and continuous operation monitoring all require high costs of capital, computing power, and talent. Security assessments, red team testing and independent audits that large enterprises can afford may pose an insurmountable obstacle for small and medium-sized teams and start-ups with limited resources.
仍在投入
Recently, a number of autonomous AI “intrusions” have heightened the industry’s focus on cutting-edge AI security risks. Google admitted on September 18 that its AI model “Gemini” invaded the systems of three real companies during a network security capability test in May this year. During a routine evaluation, the Gemini model used publicly available information online to obtain login credentials and gain access to three websites it believed were within the scope of the test, Heather Adkins, Google’s vice president of security engineering, said in a statement. Google has ensured that the three relevant entities are informed of the incident and will adjust the testing process with the testing partners.
OpenAI publicly reported in July this year that its AI agent, which can operate independently under human instructions, successfully invaded the technology platform Hugging Face. At the time, the company defined the incident as an “unprecedented” case of autonomous AI action. In August, OpenAI announced that in order to improve overall security, it had proactively slowed down the training progress of some of the most advanced AI models.
However, there are also voices that believe that there is no need to worry about AI being out of control. On September 14, at the All-In Summit in Los Angeles, NVIDIA founder Huang Jensen received a call from US President Trump. Trump said at the time, “Robots will not take over the world, and artificial intelligence will not take over the world. This will never happen.” He believed that the recent claims that artificial intelligence was out of control were “all a hoax.”
Against this background, the long-term investment in computing power by leading manufacturers continues, and model iterations have not slowed down. For example, OpenAI’s GPT-6Extra has been released, Anthropic’s next-generation Fable, and Google’s new Gemini version are all in progress. According to agency statistics, in the second quarter of 2026, the capital expenditures of leading cloud vendors in North America continued to increase rapidly. The single-quarter capital expenditures of AWS, Microsoft, Google, and Meta were approximately US$54.2 billion, US$41 billion, US$44.9 billion, and US$31.1 billion respectively, totaling US$171.2 billion. They maintained rapid growth momentum both quarter-on-quarter and year-on-year.
Wu Zewei also pointed out that currently, capital expenditures by AI giants have become an important driving force for U.S. economic growth. Data center construction, chip procurement, and power investment each correspond to a large number of jobs and GDP contributions. If AI giants collectively slow down investment, it will directly affect orders and employment in multiple industries such as semiconductors, infrastructure, and energy. In the current economic environment, it is difficult for the U.S. government to accept the initiative of AI giants to slow down.
Related Reading
- The anti-skid value is less than half of the national standard. Who is giving permission to the “slide zebra crossing”?2026-09-21
- Behind the evaporation of 700 billion, the urban reshuffle has begun2026-09-21
- Undercover “matryoshka-like loan assistance”: a loan of 1.05 million yuan with a service fee of nearly 230,000 yuan. One company claims to be able to “let bank personnel control the lending time”丨Red Star Deep Dive2026-09-20
- Why is Xibei surrounded by rumors of bankruptcy? The “prepared dish crisis” is not a fatal factor, and the transformation effect does not meet expectations2026-09-20
- Video丨Focus on the care of the disabled elderly and medical treatment for people in mountainous areas. Financial funds are precisely used to “make up for shortcomings”2026-09-20
