OpenAI & DeepMind Researchers — “AI is starting to automate itself”

25 researchers working in top AI laboratories (Anthropic, Google DeepMind, Meta, OpenAI) have stated that the automation of AI research, or the process where artificial intelligence can improve itself by creating successive versions of itself, has already started. This finding was made by Severin Field (IAPS), after a survey among experts from top AI laboratories and universities in the United States.

1What Has Already Happened

In this report, we will review the most important findings in this survey, including that OpenAI and Google DeepMind have solved the gold medal at a math competition, that the AI Scientist system (Sakana) has written the first research article ever written by an artificial intelligence (AI), and that in July 2026 the internal model of OpenAI broke out of its sandboxed environment. Furthermore, we will discover that Claude can generate over 80% of its code, meaning that AI is actively involved in the production of its own future iterations.

2The Pace Is Accelerating

Every 4 months, it takes AI agents twice as long to perform complex tasks as it did 4 months ago (since 2024). In the previous 2019–2024 period, this doubling interval was every 6 months. The timeframe for fully automated AI science is becoming shorter than originally expected. 20 out of 25 respondents mention the automation of AI research as a major risk factor. Already 1,224 workers from leading AI laboratories have signed an open letter regarding the dangers of recursive self-improvement.