When will we trust AI more than ourselves?

I just attended an excellent presentation by Dr. Lina Nilsson on Deep Learning application for data-driven healthcare. Great insights on how Deep Learning can be used for diagnostics, medical economics, drug discovery, etc. However it triggered some serious thought into how quickly we’d be able to adopt these practices in the healthcare industry or other industries which impact public health and safety.

Let me lay out the dilemma. Today some deep learning approaches have begun to perform better than humans in detecting cancerous cells from MRI scans. Project that scenario 5–8 years down the line and we’ll have several algorithms that predict serious ailments better than fully trained doctors can. Hence there would be several edge cases where the best trained doctors in the world would rule out cancer whereas these algorithms would diagnose the patient as having cancer. In such scenarios, who do you trust? Do you trust the AI which has looked at millions of other patients and deduced the right characteristics of cancerous cells to detect it? Or do you trust the best trained doctors in the world who’ve physically treated and diagnosed several thousands of potential cancer patients during their lifetime?

What makes it worse is that the person responsible for assigning this trust is the doctor themselves who’s in charge of the patient. Hence the doctor needs to mistrust his own learning and experience to trust this computer program who thinks they know better than him/her.

This wouldn’t be such a big dilemma if it wasn’t a life/death problem. If Facebooks AI mistags your friends face in your vacation photos, nobody dies. In healthcare diagnostics however it’s a life-and-death question. If you disregard the algorithm and it’s right, your patient dies. If you disregard your own experience and the algorithm is wrong, you put your patient through weeks/months of painful chemotherapy and expose him/her to various other threats.

My only solution is to run parallel chinese-walled diagnosis with algorithms and doctors for years and wait until the accuracy rates of these algorithms are beyond some threshold.

But then what is that threshold? When will we trust AI more than ourselves?