Posted: 29 May 2019 05 min. read

Avoiding the Singularity

Greater intelligence of computers

We stand now at a crossroads and have the choice between ceding discussions to AI or building a human-centred society. Algorithms are becoming ubiquitous, and we are, in our trusting way, handing over responsibility for important decisions to algorithms.

This may be an informed choice based on the belief that human error is the cause of so many problems and that automating processes can only improve outcomes. That is indeed the case for many repetitive, simple procedures. Computers are fundamentally rational, but it is not possible to reduce the world to a rational place. It takes out the beauty of the world. No cost/benefit analysis will lead Hillary up Everest or Joyce to write Finnegans Wake.

What we need to work out is a way of optimising the relationship between HI and AI, Human Intelligence and Artificial Intelligence. The fact that AI can beat any human at chess and GO is seen as a sign of greater intelligence. But games are essentially finite and it is the infinitude of the real world that means all computable processes will come up short. Board games are a lot simpler than real life. There is something, intuition or instinct, in being human which extends beyond rational analysis. HI and AI are complementary; to paraphrase Garry Kasparov, he accepts he can be beaten by a computer, but he and a computer together will be stronger.

AI, when done right, can address human bias. For example, many studies in the US indicate that résumés with names indicating non-white heritage attract fewer call-backs than résumés with white-sounding names. But algorithms can also produce biased outcomes and if they have an aura of infallibility, our natural scepticism, which may allow us to adjust for bias, may not come into play.

Algorithm bias can arise from learning data: a hiring algorithm trained on past successful recruits will result in selecting more of the same. If we do an image search of the internet for “CEO” we will mostly be presented with pictures of men; the data will lead an algorithm to predict that the future will look like the past. As humans, we can conjure visions of the future that differ from the past and strive to achieve Utopian visions; that’s a human trait I want to keep, I don’t want to hand over decision making to computers. Computers can be programmed to develop alternative futures, but there can be no insight or understanding of the human condition applied by machines. So AI, if approached as a purely technical discipline, can merely amplify human culpability.

What we need are processes that ensure algorithms meet certain standards. There has been a lot of discussion on what standards algorithms should meet, and there is developing regulation.  Unacceptable discrimination must be avoided, and transparency of algorithmic processes is desirable.  Among other things Europe’s General Data Protection Regulation (GDPR) provides a “right to explanation” of decisions made by automated systems. Other regulators around the world are following suit. 

There are precedents which can inform our approach to the challenges of unacceptable discrimination and lack of transparency in algorithmic processes.  Our approach is to use actuarial frameworks to design, implement and monitor algorithms so as to enable human ownership of algorithm processes and ensure algorithm outcomes are consistent with corporate and social objectives.

Overviewing algorithmic processes requires a range of technical and governance skills.  The actuarial profession is trained to take the reasonable person’s perspective to balance conflicting interests, and have the technical aptitude, credentials and standards of practice required to take ownership of the appropriateness of algorithmic outcomes.  Our process unlocks the black box and uses established procedures from data science to specify and explain how algorithms arrive at particular outcomes.

Meet our author

Rick Shaw

Rick Shaw

Partner, Consulting

Rick is a partner of Consulting and part of the Actuaries practice. He has extensive overseas and Australian experience, and is recognised internationally for his work on capital modelling, regulatory systems and pricing and valuation. Rick’s primary focus is developing management information systems and integrating capital models into companies’ decision making. He has also advised regulators on actuarial valuation standards and capital model approval.