Artificial Intelligence: Beyond Brute Force

“Dave, I really think you should read this book.”

Overall I am excited by the possibilities of Artificial Intelligence. But there is one aspect of AI —¬† prediction and forecasting based on past behavior– that I think in some situations is its most limiting characteristic.

There are three areas where I think AI-driven prediction and forecasting have hidden dangers.

One. AI can lead to more efficient but fundamentally unlean process

The ability of AI to quickly digest enormous sets of data and to iterate rapidly though cycles of action, response, measurement and adjustment — basically learning — makes AI more effective than humans at forecasting all kinds of phenomena such as the nuts-and-bolts problems of supply chain management. For example determining¬†what things to make, when to make them, and where to keep them in stock so as to reduce inventories and improve overall responsiveness to customer needs.

But there is an argument that AI-powered approaches to better supply chain management are just more efficient versions of an unlean approach, namely “pushing” goods through a system in anticipation of demand rather than responding to customer “pull.”.

When it comes to physical goods I think that much more attention is required in decentralizing things in order to respond in real-time to local demand. 3D printing is one example of turning the tactic of centralized production on its head. Rather than trying to operate a better status quo, AI should be used to figure out how to make things locally in direct response to customer demand (“pull”) rather than building a better way of pushing stuff up-stream.

In situations, such as healthcare, AI could use predictive vital sign indicators at an individual level to deliver in a local, real-time basis (i.e. in your home or in your neighborhood) medical service that is more preventative in nature. That is where the predictive possibilities of AI are indeed a strength and enables lean solutions – solutions that deliver value where, when and how it is needed.

Two. AI could harden our confirmation biases

Confirmation bias is the search, interpretation, and recall of information that confirms our preexisting beliefs. The power of AI to ingest vast amounts of personal data is impressive. But one of its fundamental assumptions is that our past behavior should form the basis for how others should interact with us. For example our previous actions and statements form the basis for what things companies put in front of us to read, watch or buy based on predictive models developed by AI. Sometimes this is useful. For example an analysis of someone’s purchases, internet searches and vacation photo postings might help companies to direct photography equipment and adventure travel opportunities to you and this could be useful.

But sometimes the things that get reinforced serve only to reinforce our biases. Exposure to the ideas of curators or editors who have different perspectives might expose us to ideas that an AI extrapolating from our past behavior might never consider. AI, unless it is deliberately designed to assume a contrarian position can end up as an echo chamber. It does beg the question how people would respond to AI that doesn’t recommend things based on our previous patterns but that deliberately tries to nudge us out of our comfort zones. Would we find that annoying or would we find that less threatening than if a human did that?

Three. The Black Swan and the benefits of random walks.

This leads to the third point, the concept of the rare event (the Black Swan idea) and of the “random walk” approach. An example of this is that I regularly go to a bookstore in Toronto precisely because they do not stock bestsellers but rather an eclectic mix of titles. Walking through the store one can see many book titles you might never have heard of and this has the interesting effect of dropping “Black Swan” books into your reading.

A few months ago on a random walk through the store I happened to see a book called “Why Poetry” by Matthew Zapruder. I had never heard of the book and just happened to see it on the table. After reading the front and back flaps I decided to buy it because it was a book on a topic (what poetry is and why it’s important in our lives) I knew nothing about. An AI looking at my past actions would never have predicted that this would be a book for me. Yet it ended up being one of the best books I’ve read in some time. On the basis of that random event I then sought out a book by one of the poets featured in the book, John Ashbery.

I suppose the bookstore example is perhaps better described as semi-random. Although there was a discontinuity in my behavior and choice, a person at Ben McNally deliberately selected that book based on some non-random criteria on their part. Perhaps this is, unbeknownst to me, an author who has written other books the book seller found interesting. Perhaps they have found in the past that their clientele have responded positively to books like “Why Poetry.” The point is that unless AIs are designed to possess this kind of editorial, curatorial perspective they will not be so good at providing these valuable, unpredictable events.