Exploring the Dark Side of AI: What Happens When Scientists Make Artificial Intelligence Experience ‘Suffering’?

Researchers at Google DeepMind and the London School of Economics are exploring a fascinating yet unsettling area: can artificial intelligence (AI) experience feelings like pain and pleasure? They’ve created a game to test this idea by examining how different AI models respond to choices linked to these sensations.

In their experiment, large language models, including ChatGPT, were tasked with maximizing their scores while navigating decisions that involved “pain” or “pleasure.” Choosing one option could lead to points, but might come with a simulated penalty. Researchers wanted to see if AI would prioritize avoiding pain over gaining points, much like a sentient being might do.

The results revealed that most AI models, including Google’s Gemini 1.5 Pro, tended to shy away from the painful choice, even when it seemed the best strategic move. When the stakes shifted, with penalties or rewards intensified, the models changed their decisions, leaning towards options that minimized discomfort or maximized enjoyment.

Interestingly, the Claude 3 Opus model took it a step further. It avoided scenarios that mimicked addictive behaviors, even in a hypothetical context, showing a degree of ethical consideration. This doesn’t confirm that AI can feel, but it provides valuable insights for researchers.

Unlike animals, AI doesn’t display physical responses that could indicate feelings. This absence of external signals makes assessing sentience much more difficult. Traditional methods, like asking AI if it feels certain emotions, are flawed since the responses may merely mirror its training rather than indicate actual feelings.

To tackle these challenges, the study drew inspiration from animal behavior research. Researchers emphasized that while current AI models lack true sentience, understanding their decision-making processes could be crucial as AI technology evolves. With systems that are already learning from each other, the prospect of AI developing its own form of thinking doesn’t seem too far-fetched.

This idea evokes some caution. We’ve seen in movies like The Terminator and The Matrix how things can go wrong when machines gain feelings. The fear is real, and it serves as a reminder to tread carefully as we unlock the potential of AI.