Following a lecture at the Oxford Martin School; 13th Feb 2014. Task specific machines have been with us for a long time. Sometimes they go wrong. Who to blame? Well it is worth pointing out that slaves and employees have been around for a long time, so liability rules have become quite predictable. A machine is just the agent of the master/owner/user no matter how “intelligent”. What’s different about AI? In AI the machine develops and refines its own optimum manner of performing a function. Of course it does this within the limits of its methods of action, the feedback it gets from taking those actions and the efficiency of the optimisation software. Provided the system is well designed, then the scope of potential harms can be factored out. The designer is to blame. If the system is misused, the user is to blame. But Utility maximisation is not always the social optimum. A self-driving car might find the optimum route but it is a route which would be ethically unac

Thoughts inspired by speakers at the Amlin conference “Systemic Risk of Modelling” 11th Feb Definition of systemic risk was far from decided. The general idea No single definition of systemic risk will cover all situations. Usually what people mean is that there is a mismatch between actual and imagined risk and that the effects of realising this mismatch are contagious. The effect is then realised at a system level. In banking, several mechanisms have been proposed. But in essence they all have the same origin – assets don’t have the value they were supposed to have. Banking has the innate feature that asset values are leveraged and so the effect of mispricing is automatically amplified; an example of positive feedback, especially if further leveraging is used to cover the discrepancy. Errors are sometimes corrected by selling into a falling market, another cause of positive feedback. So, what we are looking for in insurance is: a mechanism of mispricing, where multiple in

Evidence from: http://www.bermudareinsurancemagazine.com/news/praedicat-will-mine-for-footprints-of-casualty-risks “Every mass litigation that we investigated, was preceded by peer-reviewed science by some number of years. We realised that a hypothesis that some product, chemical, substance or exposure can result in bodily injury first emerges years before general acceptance within the scientific community, before public awareness and before the first litigation.” Praedicat aims to identify these trends and pinpoint liability exposures years before they hit re/insurers. The identification engine will employ a text-based data mining system that aggregates the results of peer-reviewed scientific literature. The system examines new hypotheses regarding potential casualty exposures and applies analytics and scoring around these results to determine how the science will evolve over time. Aim: “develop a comprehensive and scalable solution for identifying emerging risks, supported by footpri