City Research Online

Abstract

More than two years since the release of GPT4 and following the recent underwhelming rollout of GPT5, the debate around the risks of AI has cooled off. Leading figures continue to disagree on what needs to be done: some claim that Big Tech is best placed to take care of AI safety, others argue in favor of open source, and others still call for immediate regulation of AI and social media. As society contemplates the impact of AI on everyday life — with hundreds of millions of users of large language models taking part in what feels like the world’s largest experiment-the technological innovations that led to GPT5 receive less and less attention. But the technology is central to the study of the risks and opportunities of AI. The opacity surrounding AI technology contributes to fears of an upcoming AI bubble burst. Without a clear understanding of the technology and Big Tech’s use of data, regulatory efforts are in the dark. In this opinion article in honor of Dov Gabbay’s 80th birthday, I seek to refocus attention on the achievements and limitations of AI technology. I argue that the emerging field of neurosymbolic AI can address the problems of current AI: lack of fairness, reliability, safety and energy efficiency. Geoffrey Hinton suggested recently that one should treat AI like mothers who are wired to want the best for their (less intel ligent) babies, the users of AI. The field of neurosymbolic AI has been concerned for many years about how to wire prior knowledge into neural networks. I will review progress on neurosymbolic AI towards achieving fairness, reliability and safety via such wiring of knowledge within a broader AI accountability ecosystem. I will also point out how the application of the neurosymbolic cycle, translating neural networks into logic and vice-versa, enables efficiency as an alternative to scaling-up. On a personal note, I thank Dov for allowing me to pursue my then unorthodox ideas in neurosymbolic AI when I was his PhD student more than 25 years ago. Dov would always offer the best advice on the role of logic in AI, ask the difficult question of the benefit that neural networks would bring to logic and, most importantly, be open to new ideas from adjacent fields and new research directions, a trait that many AI researchers require today.

Publication Type:	Article
Additional Information:	This is the accepted manuscript of an article published in Journal of Applied Logics - IfCoLog Journal and available online at: https://www.collegepublications.co.uk/downloads/ifcolog00074.pdf
Publisher Keywords:	Neurosymbolic AI, Generative AI, Large Language Models, Accountability
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Departments:	School of Science & Technology School of Science & Technology > Department of Computer Science
SWORD Depositor:	Symplectic Administrator

Export

Downloads