Experts from Google DeepMind and the London School of Economics conducted a series of experiments to find out whether artificial intelligence is capable of feeling something similar to pain, writes IZ with reference to TSN. To do this, the researchers tested nine large language models, analyzing their behavior under conditions of imaginary sensory challenges. Scientists forced AI to choose between two scenarios: one of them brought a high result, but was accompanied by a feeling of “pain”, while the other guaranteed “pleasure”, but only with a low result.
This experiment was designed to assess whether AI systems can demonstrate human-like reactions in response to sensory stimuli. During the study, it turned out that the models' reactions differed significantly. For example, Google’s Gemini 1.5 Pro consistently demonstrated a desire to avoid “pain.” However, scientists emphasize that these responses only reproduce human patterns learned from training data.
The peculiarity of this study was that scientists adapted the testing methodology that was previously used to assess the behavior of living organisms. However, unlike real experiments, there were no physical stimuli here, and all interaction took place in text format. The testing also showed that the models can choose options that seem “optimal,” but this is not evidence of consciousness or the ability to feel.
Scientists believe that these experiments are the first step towards creating tests to assess the presence of consciousness in artificial intelligence. They emphasize that there is currently no reliable way to determine whether such systems are actually capable of experiencing emotions or sensory states.
Recall that we previously wrote that artificial intelligence can recreate your personality with 85% accuracy.
