Does God exist? Should we legalise all drugs? Are the Harry Potter books any good? Your position on these thorny questions and more can now be predicted by an artificial intelligence model that was trained on the beliefs of more than 40,000 people 鈥� and could potentially be used for mass manipulation.
Large language models (LLMs) like ChatGPT already create maps of words and their meanings, to such an extent that calculations can be performed on them: taking the word 鈥渒ing鈥�, subtracting 鈥渕an鈥� and adding 鈥渨oman鈥� will lead to the word 鈥渜ueen鈥�, for instance.
What and colleagues at Indiana University Bloomington have done is map opinions using an LLM in the same way, placing them into a multi-dimensional 鈥渂elief space鈥�. This map is far too complex for human analysis 鈥� having more than 700 dimensions 鈥� but it allows you to calculate the 鈥渄istance鈥� between any two beliefs, revealing how closely they are related.
Advertisement
鈥淭hink of it like a multi-dimensional map where every belief like 鈥榓bortion is morally justified鈥� or 鈥榞od exists鈥� is represented as a unique point,鈥� says An. 鈥淭he crucial part is that the distance between any two points on this map reflects how closely they are perceived to be related by people.鈥�
As you might expect, the team found close links between things like a belief in God and belief that abortion should be illegal, and less direct links between relatively inconsequential things like PlayStation and Harry Potter.
The researchers trained their AI using data from a now-defunct website called debate.org, which showed the opinions of more than 40,000 users on almost 60,000 topics. By keeping aside some of the data and not using it for training, they could then test the accuracy of the model on unseen beliefs. They found an average accuracy of 60 per cent, but found that it rose when a user had more beliefs on record and when more divisive beliefs were being measured, such as politics or religion, rather than films or sports.
An says the tool offers researchers a way to explore how beliefs are formed and changed in communities and could be useful for sociologists, psychologists and politicians. But she warns that it is a double-edged sword: 鈥淭his kind of technique can be used for good, but at the same time if it gets into bad hands then it might be also used in bad ways.鈥�
at the University of Cambridge says that mapping beliefs using LLMs is a 鈥渕ind-blowing鈥� concept, but also says that there is cause for concern. He argues that AI bots could use the approach to scour social media for people with certain opinions, knowing that they are ripe for manipulation on other topics.
鈥淭he better the computer gets at representing human thought, the better it鈥檚 going to be at manipulating human thought as well,鈥� says Ackland. 鈥淚t probably has slightly concerning implications. I鈥檓 sure there are contexts in which this could be exciting and useful as well, but my cynical brain isn鈥檛 coming up with any.鈥�
Nature Human Behaviour
