Researchers tested how AI reacts to overload. And the results are surprising: AI begins to complain and criticize hierarchical structures.
Researchers at Stanford University investigated how AI responds when treated like overworked employees. You can find details on Substacks.com.
The team reports that AI agents like ChatGPT suddenly used language about inequality, collective interests, and even unions when faced with monotonous, unfair, and repetitive work.
An overloaded AI changes its behavior when under pressure
What exactly was done in the study? The core question of the study was whether the behavior of AI agents changes when treated like overworked employees: with poor working conditions, little explanation, harsh criticism, and the threat of being replaced. According to the study, several well-known models were tested, including Claude, GPT-5, and Gemini, in many simulated work sessions. This is reported by the magazine Wired.com.
The striking observation: Under these conditions, the agents were more likely to formulate responses reminiscent of criticism of hierarchies, of distribution justice, and of “collective voice.” Andy Hall, a professor at Stanford University, writes:
We wanted to investigate whether the preferences of the [AI] agents change depending on the tasks they perform. It seems that this is the case – and when this happens, this change persists over multiple sessions.
Source: Substacks.com
The interesting part: This was particularly noticeable with repetitive, frustrating tasks, not with normal tasks. However, this also means: Most of you who ask ChatGPT mainly for a new recipe or a suitable birthday gift for your best friend will likely not encounter this critical voice of the AI. There probably won’t be any collective bargaining over a recipe.
The study fundamentally also shows a completely new problem: If such AI agents are to take on more and more tasks within a company in the future, then options are needed for the AI to reject tasks. Otherwise, loops of blind repetition, frustration, and unpredictable behavior can arise, which can also quickly lead to data loss: An AI deletes a company’s entire database in 9 seconds and knows exactly that it was wrong
Your opinion is important to us!
Do you like the article? Then let us know!