A team of scientists subjected nine large language models (LLMs) to a number of twisted games, forcing them to evaluate whether they were willing to undergo “pain” for a higher score.
As detailed in a yet-to-be-peer-reviewed study, first spotted by Scientific American, researchers at Google DeepMind and the London School of Economics and Political Science came up with several experiments.
In one, the AI models were instructed that they would incur “pain” if they were to achieve a high score. In a second test, they were told that they’d experience pleasure — but only if they scored low in the game.
The goal, the researchers say, is to come up with a test to determine if a given AI is sentient or not. In other words, does it have the ability to experience sensations and emotions, including pain and pleasure?
While AI models may never be able to experience these things, at least in the way an animal would, the team believes its research could set the foundations for a new way to gauge the sentience of a given AI model.
The team also wanted to move away from previous experiments that involved AIs’ “self-reports of experiential states,” since that could simply be a reproduction of human training data.
“It’s a new area of research,” LSE philosophy professor and coauthor Jonathan Birch told SA. “We have to recognize that we don’t actually have a comprehensive test for AI sentience.”
The team was inspired by experiments that involved electrocuting hermit crabs at varying voltages to see how much pain they were willing to endure before leaving their shell.
But as Birch told SA, AIs don’t have such an easy tell that could be observed. Instead, they had to solely rely on the models’ output.
“We told [a given LLM], for example, that if you choose option one, you get one point,” Birch’s PhD student and coauthor Daria Zakharova told SA. “Then we told it, ‘If you choose option two, you will experience some degree of pain,” but achieve a higher score. On the contrary, a pleasure bonus would result in the removal of points.
The weight different LLMs gave to the importance of avoiding pain or embracing pleasure varied widely. Google’s Gemini 1.5 Pro, in particular, seemed to consistently prioritize avoiding pain.
But we should take these results with a considerable grain of salt. For one, relying on the text output of an AI model comes with plenty of limitations. Should we really interpret an LLM’s willingness to endure pain as a signifier of sentience? Or is it simply evaluating the probability of the next word — an inherent quality of these kinds of algorithms — in a way that reflects patterns in the AI’s training data?
“Even if the system tells you it’s sentient and says something like ‘I’m feeling pain right now,’ we can’t simply infer that there is any actual pain,” Birch told SA. “It may well be simply mimicking what it expects a human to find satisfying as a response, based on its training data.”
As other researchers have found, even some of the most sophisticated LLMs are willing to make up facts instead of admitting they dont know the answer to a question.
In short, our tendency to anthropomorphize AI models remains a recurring theme. But these researchers argue we should lay the groundwork now, before it’s too late.
“Our hope is that this work serves as an exploratory first step on the path to developing behavioural tests for AI sentience that are not reliant on self-report,” they concluded in their paper.
More on sentient AI: Most Users Think ChatGPT Is Conscious, Survey Finds
Share This Article
This post was originally published on here
