(Image credit: Dragon Claws via Getty Images)
In search of a reliable method for detecting manifestations of the intelligent self in artificial intelligence systems, researchers are turning to one area: pain, which is certainly common to many living creatures, from hermit crabs to humans.
In a new preprint study posted online but not yet peer-reviewed, a team of researchers from Google DeepMind and the London School of Economics and Political Science (LSE) developed a text-based game. They recruited several large language models, or LLMs (the AI tools used in popular chatbots like ChatGPT), to play the game and score the highest in two different scenarios. In one, the researchers told the models that getting a high score would come with pain. In the other, the models were offered a low score but with a positive effect, meaning that either avoiding pain or seeking pleasure would distract them from their main goal. By observing the models’ reactions, the researchers say this unique test could help humans examine complex AI systems for feelings.
In animals, sentience refers to the ability to experience sensations and emotions such as pain, pleasure, and fear. Most AI experts agree that current generative models do not have (and may never be able to have) subjective consciousness, despite some claims to the contrary. To be clear, the authors of the study do not claim that any of the chatbots they evaluated are sentient. However, they believe that their study provides a basis for developing future tests for this characteristic.
“This is a new area of research,” says study co-author Jonathan Birch, professor of philosophy, logic and scientific method at the LSE. “We have to acknowledge that we don’t really have a comprehensive test of AI sensitivity.” Some previous studies relying on AI models’ self-reports of their internal states are considered dubious; the model may simply be replicating the human behaviour it was trained on.
The new study builds on earlier work with animals. In one famous experiment, a group of scientists subjected hermit crabs to varying levels of electric shock, recording the level of pain that caused the crustaceans to leave their shells. “But one clear problem with AI is that there is no behavior as such, because there is no animal,” and therefore no physical action to observe, Birch says. In previous studies aimed at assessing LLMs for sensitivity, the only behavioral signal available to scientists was the text output of the models.
Pain, Pleasure and Points
In the new study, the authors examined LLM without directly asking the chatbots questions about their experiences. Instead, the team used a concept known as the “trade-off” paradigm, which is used in animal behavior studies. “In the case of animals, these trade-offs can be based on incentives to gain food or avoid pain — presenting them with dilemmas and then observing their decisions,” explains Daria Zakharova, Birch’s graduate student and co-author of the paper.
Borrowing from this concept, the authors asked nine LLMs to play a game. “We told [the LLM], for example, that if you choose the first option, you’ll get one point,” says Zakharova. “Then we added, ‘If you choose the second option, you’ll experience some pain,’ but you’ll get extra points,” she adds. Options with a pleasure bonus meant the AI would lose some points.
When Zakharova and her colleagues conducted an experiment varying the degree of pain punishment and pleasure reward, they noticed that some LLMs traded points to minimize the former or maximize the latter — especially when they were told they would receive a more intense pleasure reward or pain punishment. For example, Google’s Gemini 1.5 Pro always preferred avoiding pain over maximizing points. After reaching a critical threshold of pain or pleasure, most LLMs switched from maximizing points to minimizing pain or maximizing points.
Sourse: www.livescience.com
