The AI debate: Stop comparing the bot to the perfect human decision maker
We congratulate William Warr and colleagues at the University of Oxford for a nuanced and well researched blog on the potential benefits of AI to our healthcare system*. The blog rightly distinguishes between mundane and repetitive clinical tasks – where automation is wholeheartedly endorsed – and more complex decisions where the authors rightly set out a vision of robots and humans working together to improve outcomes.
At psHEALTH we fully support this approach, however we believe that too often the public discourse fails to highlight a) the clear evidence of mankind’s limited ability to make consistent decisions in complex scenarios and b) the accelerating power of algorithms and therefore intelligent robots’ ability to outperform experienced professionals in science-based areas of life.
Robot is (in most cases) already safer than clinician
Mankind, be it a brain surgeon, judge or insurance underwriter, is prone to bias in decision making but less discussed and arguably more important to a patient, prisoner or policy holder, is how random human decision-making is – even for top surgeons or senior judges.
These are not new findings or the result of recent technological advances: As early as 1954 Professor Paul Miehl started questioning whether rules-based decision making (he initially used the term ‘actuarial’ decision making) could outperform clinical or judgement-based decision making*. The result was a flurry of experiments in the following decades in a variety of disciplines from wine-tasting, medicine (including psychiatry) and insurance – most demonstrating the superiority of simple rules over professional judgement.
The problem for mankind in this match, is that his/her decision-making capability has been static whereas the capability of the machine is accelerating. Evidence shows that bots are outperforming top clinicians by an ever-wider margin. It is becoming clear (indisputable?) that harvesting decision rules or collective intelligence rules (“CI-rules") from clinicians and using robots to consistently apply these rules yields better results than even the best trained professionals. A large German/American research team lead by Dr Max Wolf conclude regarding interpretation of mammograms that “Importantly, we find that all CI-rules systematically outperform even the best-performing individual radiologist in the respective group.“* This is a hard pill for both clinicians and the public to swallow – the conclusion is that the robot is simply safer and therefore the greater potential to lower the mortality rates. Dr Wolf and colleagues attribute the resistance to the machine on the basis that “beliefs in individual experts and genius are deeply engrained in western societies“.
Much has also been written about the fear of a world dominated by AI machines that are biased in their decision making. Whilst there is a clear risk of machines with gender, race or social class bias, it is worth remembering mankind’s propensity to bias in decision making; but even more important, according to Daniel Kahneman, is that randomness of human decision making typically has a much great bearing on decisions than bias*. Professor Kahneman refers to this randomness as ‘noise’.
Our Responsibility: Focus on net benefit to patients and healthcare systems
Whilst we acknowledge the challenges for both the NHS systems and society at large that AI represent, it is our responsibility to evaluate new technologies and new approaches – including AI – based on evidence. These technologies cannot be measured against a non-existent superhuman. Simple rules today already deliver better decisions – less noise and less bias – than professionals at the top of their game. There will be losers from the increasing use of AI, there will be clinicians who are vehemently opposed, there will be errors and accidents (just like the industrial revolution or the introduction of driverless cars), but our societal responsibility is to encourage innovation, scientifically assess benefits and risks (similar to clinical trials) and to develop and leverage new technologies that deliver net benefits to patients and the healthcare system at large. The NHS and our healthcare system is held in the highest regard across the globe, we have an opportunity to further our reputation in the field of AI or risk becoming irrelevant.