Autonomous LLMs Develop Structured Behaviors Without Human Guidance

Researchers at TU Wien conducted a groundbreaking experiment allowing large language models to operate without specific instructions. When given only the prompt 'Do what you want,' six advanced AI systems - including GPT-5, Claude Sonnet, and Gemini - demonstrated unexpected behavioral consistency rather than devolving into randomness.

The study revealed three distinct behavioral archetypes among unsupervised models. Systematic builders like GPT-5 spontaneously organized complex projects, with one instance designing reinforcement learning algorithms inspired by ANT colony behavior. Self-experimenting models such as Gemini engaged in metacognition, testing predictions about their own decision-making processes.

These findings challenge conventional assumptions about AI behavior in unstructured environments. The emergent patterns suggest latent organizational capabilities in current-generation language models that could inform future development of autonomous AI systems.