Knowledge Evolution and Scientific Breakthroughs triggered by AI Hallucinations - A Paradigm Shift?

The interdisciplinary impact of artificial intelligence (AI) in science has been especially empha- sized by the fact that both, the Nobel Prize in Physics and in Chemistry in 2024 have been awarded for pioneering research with results, decisively based on artificial neural networks. The core of the excelling achievement in chemistry is described as: capturing of the full compu- tational understanding of living matter at atomic level (Abriata, 2024). An interesting detail behind this highly acclaimed success, is that one of the laureates had praised AI hallucinations to be the designers of de novo proteins (Anishchenko, 2021). AI hallucinations are defined as incorrect or misleading results generated by AI models. They are particularly associated with large language models, chat bots and computer vision tools, implementing generative AI. Gen- erative AI models are known to occasionally create nonsensical or altogether inaccurate outputs which originate in data interpretation, that deviates from what is expected or intended. They are usually welcome in domains such as imaginary and visionary art but they can have signif- icant negative consequences in practical applications. Hallucinating AI systems can fabricate nonexistent imaging results to justify a medical diagnosis, can misinterpret patient informa- tion, delivering plausible but incorrect conclusions about health status, news bots can spread falsehoods undermining mitigation efforts in public emergency situations, to name but a few of many known cases. However, AI hallucinations prove now to have a positive impact on biochemistry research. A conceptual model, widely used within information science and knowl- edge management, is the so called DIKW pyramid, describing the continuum of relationship between data, information, knowledge, and wisdom, putting wisdom at the highest level in the hierarchy (Cotton, 2023). AI systems lack human wisdom. They do not solve problems via understanding context or using ideas of their own. They work with predefined inputs and in the case of generative AI, they generate new patterns some of which may deviate from the knowledge implemented in the algorithm or even defy the wisdom of the algorithm designer. Still, they can prove to be compatible with reality as is the case with de novo proteins. AI hal- lucinations could then be viewed as glimpses into a future, one yet to be created, for instance when introducing man made proteins and organisms into the existing biosphere. Epistemolog- ical questions arising from the perspective that creative mistakes of AI can promote science more effective than human ideas will be discussed. Possible risks in connection to a rapid application of in silico results in structural biology, created mostly with machine learning, will also be considered.