AI Has Already Begun, Says It’s Anthropic—And Humans Can Slow Things Down

AI has become so effective in coding and research that the biggest obstacle to developing new AI systems may now be the people overseeing them, according to a new study by Anthropic.

In his report “When AI Builds,” printed On Thursday, Anthropic said that Claude is already helping to build the AI ​​systems of the future by writing code, conducting experiments, and supporting research—activities that the company said will eventually lead to iterative innovation, where AI systems help create their successors.

Claude now writes more than 80% of the code included in its codebase, Anthropic said, and will help engineers grow the code nearly eightfold by 2024.

“Before the implementation of the Claude Code in February 2025, this number was low,” wrote Anthropic, adding that the change also reflects the number of engineers’ output. “The number of lines of code added per engineer per day remained constant for the first four years of Anthropic (2021-2024), then began to increase in 2025 when Claude started running code instead of just asking engineers to copy and paste.”

Anthropic said the future could play out in a number of ways: AI progress could slow down, humans could remain in control while AI does more work, or AI systems could begin to replace their successors.

“With enough simulations, and enough calculations, the results point to an AI system that can build itself and create a successor,” Anthropic wrote. “This is called iterative self-improvement. We’re not there yet, and iterative self-improvement is inevitable. But it may come sooner than most organizations plan.”

The company said it’s too early to tell what the future holds, but it says AI is already helping to create AI, and it acknowledged that lines of code are a wasteful process.

“None of this proves that your ego is imminent,” Anthropic later. he wrote at X. “It’s not yet clear that Claude is capable of investigating decision-making problems.”

The report comes as AI companies increasingly treat their models as research participants rather than simple chatbots. However, Anthropic said the increase in code represents a significant rise in software development led by AI-enabled agents.

Last month, Anthropic upgraded its version of Claude to be Opus 4.8to continue the continuous output aimed at document improvement, reflection, and independent work. At the same time, competitor OpenAI followed a similar path to its frontier models, launching GPT-5.5 and GPT-Rosalind in April.

In May, Google announced Gemini Sparkan AI assistant that doesn’t wait to be asked. It monitors work across programs, flags things that need attention, and completes tasks in the background.

The report also comes as Anthropic has put more emphasis on AI systems that can operate autonomously as it prepares to go public. In recent months, Anthropic has demonstrated progress in coding, workflows, and long-term performance, as it demonstrates. Claude Mythos‘ Ability to identify software vulnerabilities and perform critical cybersecurity research.

“Humans play a very small role in their development, making our efforts to monitor, validate, and validate a growing ‘virtual lab’ powered by AI systems,” the company said. “We hope that machines capable of research and development of AI will have skills that can transfer all of science, allowing them to start revolutionizing other fields.”