AI on threshold of creating new models on its own:Scientists express concern that human control over machine decisions may eventually slip away

The question about AI until now was how quickly it can do human work. But now the concern goes beyond that. Experts are asking – if AI starts creating next generation models on its own, will humans be able to control it? Recently, AI company Anthropic had said that there should be an option to slow down or temporarily halt frontier AI development. According to the company, more than 80% of the code published in May 2026 was written by AI. Anthropic’s co-founder Jack Clark estimates that by 2028, a system is possible that can create its new version without human help. This is called recursive self-improvement. That is, one model of AI creates another, the second creates the third, and the third creates the fourth. Each model becomes more capable than the previous one. MIT physicist and AI safety expert Max Tegmark warns that if the pace of development is not monitored, AI will surpass humans in decision-making. In such a scenario, the human role in governments, companies, and important institutions could weaken. Until a few years ago, AI only followed instructions. Now it is finding solutions. Google DeepMind’s AI system Alpha Evolve has suggested ways to improve data center operations and accelerate AI training. This is why a section of scientists believes that humans will remain research directors rather than researchers. They will only provide direction. However, some experts believe that this change will not happen overnight. AI still needs data centers, expensive chips, electricity, and new training materials. Its pace will remain limited for now. Currently AI is not creating successors, but is accelerating training Currently, no AI model can fully create its successor on its own. However, large models can create smaller models themselves. Former OpenAI researcher Andrej Karpathy trained a chatbot. GPT-2 was created by OpenAI in 2019. It took 168 hours then. Karpathy achieved the same result in 3 hours with 8 GPUs in one computer. Later, this task was given to an AI agent. It reduced it by another 18%.