Taming AI and becoming responsible for what we have tamed

?

"To me, you are still nothing more than a little boy like a hundred thousand other little boys. But if you tame me, then we shall need each other. To you, I shall be unique in all the world. To me, you will be unique in all the world."

Antoine de Saint-Exupéry, The Little Prince

?

The adoption of AI in healthcare is a complex and multifaceted issue. While there are numerous stories of remarkable successes in terms of medical breakthroughs, technological advancements, and improvements in patient well-being, there is an equal number of cautionary tales. Perhaps as a society, we need to develop the necessary skills and expertise to navigate this intricate landscape effectively.

A process that holds promises

However, even social learning itself is multidimensional. S?rensen [1], while exploring this concept's application in technology development and use, noticed that it holds three promises.

Firstly, social learning theory equips us with tools to analyze how cultures adopt and are shaped by technologies. It highlights that significant changes often occur during the process of using a technology, and rushing this process might not be beneficial. Finally, the theory offers valuable insights for developing more effective and comprehensive strategies to regulate technology in society.

Today we’d like to invite you to take a look at a very interesting publication. Robin Williams et al. used S?rensen's framework to analyze the social learning involved in applying diagnostic AI and described it in a paper titled “Domesticating AI in medical diagnosis” [2].

AI in medicine – a profound connection?

Researchers explore the social learning?framework through four case studies:

• Automatic Recognition of Brain Lesions
• Digitising ECGs and AI-enabled Atrial Fibrillation (AF) Prediction
• Rolling out Computed Tomography scanning tool for Lung Nodules
• Hospital AI procurement strategies: From AI apps to platform

The publication meticulously outlines the conditions and challenges that AI model developers faced in each case study – or rather: journey. However, we would like to focus on the points where these narratives intersect with social learning theory.

The authors distinguished three key points derived from S?rensen’s framework:

• Learning by doing. The early stages of integrating AI into clinical settings involve a "learning by doing" approach. This includes initial experiments to develop effective AI models and showcase their potential for improving healthcare practices. While many promising experimental collaborations have emerged, only a small fraction translate into successful and long-lasting innovations. Why? One of the answers might be related to investment interest. It currently prioritizes diagnostic tools over areas like health service organization, even though the latter might offer greater health benefits overall and might be a better first step towards AI domestication.
• Learning by interacting. The key to successful AI applications lies in their real-world implementation and ongoing development. This approach involves integrating these tools into everyday clinical workflows to prove their clinical effectiveness and commercial viability. An example of AI-powered lung cancer detection using CT scans was used to illustrate the concept. Healthcare professionals rigorously evaluated the tool's performance, applying the same strict standards they use for their own work. This process revealed the tool's strengths and weaknesses. For instance, it might be very reliable in identifying clear cases but less effective in handling ambiguous ones. In this way, healthcare professionals established a way to responsibly use the AI tool based on its real-world performance, even if the inner workings of the algorithm remain unclear. (We focused on an explainable AI – iXAI – in medical imaging in one of our previous articles.)
• Learning by regulating. The initial attempts to integrate specific AI tools in healthcare have paved the way for a larger phase of social learning and experimentation. This phase focuses on finding the most effective institutional arrangements to safely, effectively, and affordably implement AI in clinical settings. The research highlights the role of informal regulation and "domestication" by healthcare providers and clinical professionals. The authors also noted that the social learning priorities have shifted. Initially, the focus was on proving the value of individual AI tools. Now, the emphasis is on building the infrastructure and systems needed to effectively manage, implement, and maintain a growing number of AI applications on a large scale.

?

As we've seen, implementing AI in healthcare is an ongoing journey of social learning. We've moved from initial tool development to a crucial stage of building the infrastructure – but also building trust. Taming, in this sense, involves patience, vulnerability, and a willingness to open oneself up to new possibilities.

?

References:

[1] S?rensen, Knut H.: Learning Technology, Constructing Culture: Socio-technical Change as Social Learning, STS Working Paper No. 18/96.

[2] Williams R. et al.: Domesticating AI in medical diagnosis, Technology in Society, Volume 76, 2024, 102469, ISSN 0160-791X, https://doi.org/10.1016/j.techsoc.2024.102469.