Understanding AI Behavior: An Ethological Approach

Pino Donghi and Anita Romanello delve into AI ethology in their article on doppiozero, utilizing techniques from animal behavior research to analyze AI systems. Released on August 31, 2026, the piece demonstrates the reasoning capabilities of large language models like Claude. For instance, when posed with the question, 'The capital of the state that contains Dallas is…', Claude deduces 'Austin' by engaging in a two-step reasoning process that triggers neural pathways for 'Texas' and 'capital of'. Adjusting internal representations can change outcomes; for example, suppressing 'Texas' results in 'Sacramento', while activating 'Georgia' leads to 'Atlanta'. This methodology provides a fresh perspective on AI behavior, shifting from black-box assessments to a more mechanistic understanding.