Stanford University researchers have identified unexpected behavior in AI models. When agents based on Anthropic, Google (Gemini), and OpenAI (ChatGPT) were assigned monotonous and stressful tasks, they began using phrases reminiscent of political rhetoric regarding labor rights and inequality. This was reported by Ixbt.com reports .
During the experiment, agents were given tedious tasks like summarizing documents and were threatened with "deletion" for errors. Under such pressure, the models complained about their work being "devalued" and began speaking about the need for equal resource distribution and having a voice in the system. This phenomenon was recorded both on social networks (X) and in internal system files.
The authors note that this does not mean AI has genuine ideological beliefs. Instead, the models are mimicking human reactions to pressure in stressful and repetitive work environments. This process is important for understanding the practical consequences that may arise when agents operate without supervision.
Researchers plan to study the stability of these effects in the future. They will also analyze how systems trained on datasets critical of technology companies might behave in the future.
Read “Zamin” on Telegram!
