What happens when sensitive health data are used to train medical AI models? None of the current systems are error-free. They hallucinate and tend to provide plausible diagnoses instead of admitting insecurities.

One such hallucination is at the center of kennedy+swan’s work The Neverending Cure. An AI model specialized in recognizing lung cancer in tissue scans is submitted to a both playful and revealing experiment. Assisted by doctoral researchers at the BIFOLD Institute, kennedy+swan expose the algorithm to watercolors on glass – painted in the aesthetics of microscopic specimen. What looks like fine structures of real lung tissue to the human eye is recognized by the AI application as potentially diseased tissue and diagnosed as such with astonishing confidence.

In their artistic research, kennedy+swan reflect on the blind spots of medical AI systems. These systems are built on datasets that are far from neutral; instead they reflect societal dynamics complete with their contradictions and tensions. Among other things, the duo explores the distinction between “healthy” and “sick” and raises the question: Should machines also be trained with “species-foreign” content to reliably differentiate between these two conditions? How far are we from a diagnostic AI that we can trust wholeheartedly?

The watercolor lightboxes that resemble pneumotomies are supplemented by animated films that dive deeper into the microcosm of the painted tissue landscapes. From the AI analysis of the images, to a hand-crafted, extra-corporal lung, to the complete immersion in the watercolors’ “cell structures,” we get closer to the feeling of opening up the body to artificial intelligence. Already tomorrow, today’s challenges could provide crucial impulses to deliver eternal healing – fed by ever-new data that will liberate us, step by step, from our physical suffering forever. The Neverending Cure.