New study says AI poses no existential threat to humanity

As LLMs have grown exponentially more powerful in recent years, the possibility of them acquiring ‘emergent’ capabilities such as reasoning or planning has been an ongoing concern. Indeed, the largely ‘black box’ way in which LLMs operate has fed into this idea, with even those working on AI often unable to explain some of the capabilities that LLMs such as ChatGPT have been able to display. 

However, the new study claims to have put these fears to ease, demonstrating that there is no reasoning taking place within LLMs. According to the research, what sometimes appears to us as reasoning is in fact simply in-context learning (ICL), where LLMs can ‘learn’ new information based on a set of examples provided to them. Over the course of thousands of experiments, the research team showed that the ability to follow instructions (ICL) - combined with memory and linguistic proficiency - accounts for both the capabilities and limitations exhibited by today’s LLMs.

“The fear has been that as models get bigger and bigger, they will be able to solve new problems that we cannot currently predict, which poses the threat that these larger models might acquire hazardous abilities including reasoning and planning,” said study co-author Dr Tayyar Madabushi, a computer scientist at Bath University.

MORE ON AI

“This has triggered a lot of discussion – for instance, at the AI Safety Summit last year at Bletchley Park, for which we were asked for comment – but our study shows that the fear that a model will go away and do something completely unexpected, innovative and potentially dangerous is not valid.”

Led by Professor Iryna Gurevych at the Technical University of Darmstadt in Germany, the study conducted experiments to test the ability of LLMs to complete tasks that models have never encountered before. While there was no evidence for these ‘emergent’ abilities, the researchers were keen to point out that today’s AI still comes with significant risks, just not the existential threat we’re familiar with from science fiction and which many believe is looming today.

“Our results do not mean that AI is not a threat at all,” said Professor Gurevych.  

“Rather, we show that the purported emergence of complex thinking skills associated with specific threats is not supported by evidence and that we can control the learning process of LLMs very well after all. Future research should therefore focus on other risks posed by the models, such as their potential to be used to generate fake news."

The study was published as part of the proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) – a leading international conference in natural language processing.