A researcher who has been working on pre-training research in the field of artificial intelligence within OpenAI and Anthropic for the last three years, announced that he resigned from his position at Anthropic. In a long statement published on X, the researcher argued that both companies did not act responsibly regarding advanced artificial intelligence. According to him, the industry is in an ever-accelerating competition to reach superintelligent systems that can improve their own capabilities. The researcher suggested that this race should not be considered only as a commercial competition between companies and that the systems that may emerge could have serious consequences in terms of security. It should also be noted that the evaluations in the statement reflect the personal views of the researcher and are not definitive conclusions regarding the future of artificial intelligence.
The point that the researcher particularly emphasized was the possibility that future artificial intelligence systems could achieve much broader capabilities than the models used today. He suggested that these systems could reach the level of being able to carry out cyber attacks, make very rapid progress in different fields of science and technology, and access resources in the real world. He also cited the developments seen in recent years in areas such as coding, scientific problem solving, vehicle use and autonomous task execution as basis for this view. However, there is no consensus in the research world about when or under what conditions current models will actually achieve superhuman general abilities. Therefore, the timeline and risk assessments put forth by the researcher represent one of the more conservative scenarios in AI security circles.
OpenAI and Anthropic’s AI security approach criticized
One of the most striking parts of the series was the distinction the researcher made between OpenAI and Anthropic’s in-house approaches. The researcher claimed that many people working at OpenAI had not adequately internalized the civilization-scale implications of advanced AI. On the Anthropic side, he argued that the risks were better understood, but the company was under pressure to develop the technology first, thinking that other organizations would not act responsibly enough. These evaluations are based on the researcher’s own experience and are not a direct summary of the companies’ corporate positions. Still, the statement shows how evident the tension between AI security and rapid model development can become for researchers working in this field.
The researcher also criticized the companies’ acceptance of the race to reach superintelligence in its current form. He argued that such a decision had far-reaching societal consequences that could not be made solely through the internal processes of private companies. He is particularly skeptical of the assumption that security and alignment efforts can be kept pace with the rapid development of model capabilities. The “alignment” mentioned by the researcher, or artificial intelligence alignment studies in Turkish, includes research that aims to keep the behavior of advanced systems compatible with human purposes and security limits. However, both the magnitude of the risks and which technical security methods will be sufficient for much more powerful models in the future continue to be one of the controversial topics in artificial intelligence research.
The statement brought up not only criticism but also possible policy options. The researcher thinks it may be possible for leading AI laboratories operating in the United States to agree on common rules regarding the pace of development. Stating that this alone may not be enough to stop competition on a global scale, the researcher argues that even costly options such as a temporary ban on further increasing model capabilities should be discussed when necessary. The recommendations here reflect the researcher’s preferred policy approach rather than an existing regulatory decision. Restricting artificial intelligence development brings security gains as well as challenges such as economic competition, scientific research and monitoring of developers in different countries.
The researcher’s last call was directly addressed to his colleagues working on advanced models. He said that before starting a reinforcement learning study on a system that could approach the level of super intelligence, it should be questioned whether the behavior of the system is sufficiently understood. He asked researchers to demand changes in working conditions and safety standards rather than accepting as a foregone conclusion that the technology will already be developed by others. This resignation alone does not prove that OpenAI or Anthropic’s work is safe or unsafe; Despite this, the fact that a researcher who has directly conducted pre-training studies expresses his concerns so openly in public adds a remarkable example to the security discussions within the sector. As the capabilities of artificial intelligence models progress, the center of the debate is not only which system is stronger, but also under what security conditions this progress should be sustained.
Join Channel