I grew up in quasi-socialist ’80s India and we were taught that you don’t upgrade things easily and you ALWAYS unwrap presents really carefully so that you can re-use the wrapping paper. An admirable principle for shoes, less so for knowledge. Over the years, I’ve ignored a pretty big upgrade in my work: understanding the role that AI & machine learning have to play in behavioral science.
Recently, I was introduced to Causal AI which was eye-opening because what I thought drove behavior was not reflected in the data – causing (no pun intended) me to think harder. This study, by Surgo Ventures, for example, tackles a distressingly common problem – in Uttar Pradesh in India (a state with a population the size of Brazil), 20% of mothers don’t deliver their babies in hospitals leading to birth complications. You may think, as I did, that the problem is simple: hospitals are too far away. Behavioral science tells us: make it easy, thus, reducing distance is important. Ergo, build more hospitals. Well, it turns out that distance to the hospital played no role where a mother delivered her baby. Rather it is the absence of a delivery plan. This kind of insight can help decision makers optimize resources: invest in expanding transport rather than build hospitals. And it can outline a clear role for behavioral science: design a good implementation plan, to help expectant mothers and families stick to their delivery intentions.
Causal AI is only one example of the intersection between Artificial Intelligence (AI), machine learning (ML) and behavioral science. Both AI/ML and behavioral science can optimize human judgment & decision making: behavioral science by understanding the why and AI/ML by using patterns of past behavior.
Quick aside: in this article, I’m going to use the terms AI/ML as one umbrella term though I understand them to be distinct. This USAID report was most helpful: AI is the use of computers for automated decision making that normally requires human intelligence. Machine learning (ML) is a subset of AI that uses algorithms to give computers the ability to learn without being explicitly programmed.
I see two directions in which AI/ML can upgrade behavioral science: optimize the “where to compete’ and “how to compete”. (There is also how our knowledge of human biases can help reduce algorithmic biases which I’ll discuss in a later post).
First, AI/ML can help us figure out “where to compete”. Before even designing interventions, which populations and areas should we best focus our limited resources understanding and designing for? AI/ML can use large datasets to identify hotspots and vulnerable areas so as to make it clear where behavioral interventions can have the greatest impact. In this case study, IDinsight used machine learning to help Educate Girls to predict where girls were likely to be out-of-school without resorting to comprehensive but very expensive surveys. The team can then focus their expertise on what needs to be done to solve this problem in those focus areas.
Second, AI/ML can help behavioral scientists figure out better “how to compete”. They can do this through personalized nudging where ML helps to understand some patterns that are person specific and others that are situation specific and provides the appropriate nudges for this. Personalized nudging can be around (a) personalizing the content – different messages for example, for different segments or (b) personalizing the delivery channel – SMS for one, phone for another, depending on people’s preferences.
Of course in all this there are very real ethical and practical challenges as the USAID report I referenced earlier discusses in-depth: data availability, sustainability of models, privacy, regulatory and system challenges. Given these challenges and the high start-up costs, I am inclined to think this suggestion sensible: (read broadly but) invest in a few areas and use cases to really get a handle on where AI/ML can be used for good.
And now, to go and find that wrapping paper I so carefully stored away…