Scientifc AI as Value Proposition - Some Inescapable Methodological Requirements

Scientifc AI as Value Proposition - Some Inescapable Methodological Requirements

The revival of AI began when Generative AI got introduced a couple of years ago. Large Language Models for text generation and foundational models for image, audio, and video generation stimulated the use of AI and set free a wave of gigantic financial investment announcements, in particular in the US. While AI infrastructure thus is boosting, critical questions about the potential return on these investments come up. One route to value creating AI is Scientific AI as a novel way of knowledge generation. Talking about Artificial Intelligence though calls for some clarification with regard to the concept of ‚intelligence‘ first. 

Basically, the term intelligence is used for the capabilty of getting to know something. It is primarily understood as a human capability, but turns up in the concept of Artificial Intelligence as well. In philosophy of science we can find lots of reflections on the process of getting to know something, or knowledge generation. Fundamental is Kant’s work on the epistemological principles that govern the possibility of gaining knowledge (Erkenntnis) – reason (Vernunft) respectively mind (Verstand) and sensual perception or experience (Erfahrung), and his distinction of thing-in-itself (Ding an sich) or being (Wesen) and their appearence or phenomena (Erscheinung) as principle limitations of knowledge or understanding (Wissen). Observing these principles is the very first requirement Scientific AI should comply with. 

A second set of requirements addresses scientific methodology. Well established in natural sciences, but also desirable in less rigorous scientific context are methodological approaches as there are securing the epistemic coherence in the formulation of a theory or model, proving the logical validity of inferences or implications, and verifying the significance of statistical findings gained from the analysis of empirical data. 

For the sake of trustworthiness and reliability of AI-based decision systems or assistants, finally, a third requirement shows up, that of conceptual clarity with regard to the terms data, information, and knowledge. Because better foundation of an AI-supported decision in this sense needs better understanding of the underlying rationale or process of reasoning, and better algorithmic realization in the AI system. The reasoning and final decision process rests upon data (parametrized features of an investigated thing or fact), their evaluation and contextualization into an information base, and then the conceptualization into decisive knowledge.

Aiming to keep up with others in the chase for value-creating AI, and to unleash the – scientific and economic – potential of AI, the next step to take is to go beyond GenAI and head toward SciAI. There is much to do in addition to strong infrastructure investment. 

Norbert Jastroch, https://orcid.org/0000-0002-4046-450X