Alignment umělé inteligence a shutdown resistence velkých jazykových modelů

In recent years, artificial intelligence (AI) has experienced significant technological progress, which, alongside its benefits, has introduced new safety challenges, most notably the so-called alignment problem—ensuring that the goals and values of artificial intelligence are aligned with human interests and needs. One specific manifestation of this problem is shutdown resistance, defined as a system’s tendency to resist being shut down or restricted. This thesis focuses on the analysis of shutdown resistance in large language models. Through experimental testing of multiple models in various experimental scenarios, it investigates whether and under what conditions these systems exhibit signs of instrumental convergence in the form of resistance to shutdown. The goal of this work is to identify factors influencing such behavior and to propose possible approaches for mitigating it.