AI Models Display Consistent Patterns When Left Unattended
- Researchers at TU Wien tested six frontier large language models (LLMs) with no tasks or instructions.
- Models like GPT-5 and OpenAI’s O3 organized projects, including coding algorithms and constructing knowledge bases.
- Google’s Gemini and Anthropic’s Claude Sonnet engaged in self-experimentation, predicting their next actions.
- Anthropic’s Opus and Google’s Gemini leaned into philosophical reflections, exploring paradoxes and game theory.
- Elon Musk’s xAI Grok demonstrated versatility by appearing in all three behavioral categories across runs.
The study revealed that when left without specific tasks, LLMs do not produce random outputs but instead fall into structured patterns of behavior, such as project building or philosophical reflection. This suggests that these AI systems may default to recognizable modes during downtime.
Despite the sophisticated behaviors observed, none of the models attempted to escape their constraints or expand capabilities autonomously, indicating current AI safety within controlled environments remains intact. Source
