Artificial intelligence learns to avoid suffering: Google DeepMind Research

Researchers from Google DeepMind and the London School of Economics conducted an unprecedented experiment studying artificial intelligence’s ability to respond to “pain” and “pleasure”. The study involved nine major language models, including ChatGPT and Gemini 1.5 Pro, which demonstrated unexpectedly “human-like” behavior.

In the experiment, AI models were presented with a text-based game offering a choice between two scenarios: obtaining a high score with the risk of “pain” or a low score with guaranteed “pleasure”. The results were striking: artificial intelligence consistently sacrificed high scores to avoid “painful” consequences.

Particularly demonstrative was the behavior of Google’s Gemini 1.5 Pro model, which systematically chose a “pain” avoidance strategy, even when it led to the loss of potentially high results. Moreover, the models demonstrated the ability to adjust their behavior depending on the intensity of the offered stimuli.

“While the results don’t prove consciousness in AI, they lay an important foundation for future research,” notes Jonathan Birch, LSE professor and study co-author. The experiment’s methodology was inspired by similar studies of animal behavior, particularly hermit crabs’ reactions to unpleasant stimuli.

Unlike traditional research based on AI systems’ self-reports, the new approach focuses on behavioral responses without direct questions about artificial intelligence’s internal states. This allows for a more objective picture of AI decision-making processes.

The research raises important questions about ethical and legal aspects of artificial intelligence development, especially in the context of systems demonstrating behavior similar to living beings. As AI evolves, such observations become critically important for developing ethical principles and reassessing artificial intelligence’s role in society.