Can causal AI improve the decisions we make?

Artificial intelligence has gained ground in our society and has become a topic of conversation in Danish homes. It is often referred to under the abbreviation AI.

The technology can be used in many contexts. Yet, around the dinner table, people often joke about the strange answers artificial intelligence sometimes gives. The reason is simple: AI works much like a calculator. It looks for patterns in data, but it does not understand the context in which the data exists.

A new research project will try to provide greater insight into how companies can get the most out of the type of artificial intelligence that rests on causality and is abbreviated as cAI.

»The primary goal is to come up with guidelines that can help companies set up their causal thinking, their causal AI pipeline. This means giving companies advice on how they can best bring in their human experts,« says Paul Hünermund, Associate Professor at the Department of Strategy and Innovation at Copenhagen Business School (CBS), who will lead the research project.

Causal AI depends on human expertise. It is the experts who provide the knowledge about causes and relationships between data, which is the foundation that allows artificial intelligence to deliver reliable and trustworthy results.

The causal artificial intelligence is being tested in supermarkets

Paul Hünermund, who has received a Sapere Aude grant from Independent Research Fund Denmark, collaborates with the Dutch grocery group Ahold Delhaize during the research.

Ahold Delhaize's most well-known brand is the Albert Heijn store chain. Here, they want to reduce the amount of food waste in particular, but also waste, such as packaging.

The collaboration provides an opportunity to test the efficacy of causal artificial intelligence in an environment where it is actually applied.

»With causal AI, we bring in the human factor as part of a dynamic process, where the AI system may make a recommendation, and then the human has the final decision and authority. So, unlike classical AI, you can't train causal AI algorithms in a fully data-driven way. You have to bring in some kind of knowledge that comes from the outside,« explains Paul Hünermund.

The need for expert knowledge raises new questions. Because what does human expert knowledge actually consist of? Who will deliver it?

For example, if you want to reduce the waste of bread in a supermarket, you can choose to change the layout of the store. Many stores also use dynamic pricing, which means that the price of bread drops as the clock approaches closing time. So, different types of expertise can come into play, from the store manager to the logistics expert.

»There is no good research out there and no good guidelines for how to do this in practice. Many companies that we talk to are interested because they want to obtain exactly this kind of causal knowledge. But they are kind of stuck in how to bring people in and how to structure these human-in-the-loop processes. That's what we are trying to explore,« says Paul Hünermund, who will be hiring a PhD student and a research assistant to help him during the three-year project.

The assessment of cAI rests on experiments

The research will be divided into three parts, and the first part is about setting up different teams of experts. Together, they should be able to see both the big picture, such as a logistics expert, and the details of the individual store, which a store manager will be able to do.

Once the teams of experts are in place, the second part of the project will focus on how good the knowledge of the different teams of experts actually is.

»Here, our goal is to evaluate whether expert teams' performance can be improved, for example by comparing mixed expert teams with teams of employees from the same department. The most rigorous way to test such hypotheses is through randomized controlled trials,« Paul Hünermund elaborates.

The third and final part of the project examines the actual value of causal AI — specifically, whether a causal approach leads to better decision-making. Ultimately, the goal is to determine whether it can deliver on its promise of reducing food waste and improving business operations.

During the project, the project group will collaborate with Maastricht University in the Netherlands and, in an advisory role, the Causal Artificial Intelligence Lab at Columbia University in New York, USA.

Resources are often wasted due to a lack of knowledge

Paul Hünermund emphasizes that there are two prominent reasons why causality and causal AI are important, also at the societal level.

One is pragmatic. We want to understand what works. Many resources are often wasted by not knowing exactly what kind of policies or managerial initiatives have the highest impact,« says Paul Hünermund and concludes:

»The second reason is more philosophical. We want to understand why things work the way they do. For example, why the number of shark attacks increases when ice cream sales increase. So, it's more of a scientific approach to forming theories. But they all have to do with causality. We really need to understand cause and effect. That's also how we humans think.«