Talking about measuring well-being, what are some concrete examples of how new methods can be used to investigate these areas?
In my lab we develop proxies for measuring well-being in a variety of areas, from the relational to the environmental. Specifically, one approach we have found very promising is to measure diversity in complex social systems, particularly on socioeconomic diversity. One of the best proxies identified is the mobility of people: not so much the amount of movement, but the ways and patterns in which people move. For example, territories with a high presence of “returners” (people who move in a repetitive and routine way often following the home-to-work route) tend to be economically more disadvantaged, while territories with a high proportion of “explorers” (people who move in a varied and dynamic way without necessarily repeating patterns) are generally wealthier. This index of well-being can be applied at different levels: municipal, provincial and regional. In order to study these fields, we rely on data from many sources; for example, the widespread use of mobile telephony has allowed us to obtain extremely extensive data while respecting the privacy of participants. In one of our publications in Nature, we studied the prevalence of routinely mobile versus overall mobility of people, expecting a continuum but finding instead an extreme bias. There is an almost anthropological tendency to move toward one of these archetypes (returner or explorer), and during different stages of one’s life it is possible to move from one to the other, but it is very rare to find people who are somewhere in between.
There has been a lot of talk about artificial intelligence in recent years, what do you think are the main problems with these technologies and possible strategies to solve them?
We are currently developing the idea of “social artificial intelligence”, that is, we are trying to develop new mechanisms to measure the collective effects of AI and potentially design solutions. Essentially, AI-based tools have been developed from a selfish perspective, to improve the performance and output of the individual often at the expense of the collective. This can have significant socio-economic and environmental impacts. As a concrete example, working on a traffic digital twin in Rome, we simulated the impact of heavy use of navigation apps. If an increasing percentage of drivers followed the suggestions of a representative mix of navigators currently on the market, emissions would increase rather than decrease because traffic would be concentrated on a smaller number of roads. Looking at a map, it shows that the road arcs traveled are concentrated into almost separate areas. This trend is potentially worrisome because, in addition to an increase in emissions (which in our simulations potentially go up to double) we also see that emissions would concentrate in specific areas creating inequality problems. These problems related to a reduction in choices related to AI suggestions are also found in other fields, for example, suggestion algorithms in e-commerce increase the diversity of purchases of the individual, but reduce the overall diversity of the market, also creating a concentration effect. Popular products are often amplified while niche products are further penalized. One possible solution could be the inclusion of ocular randomizers in recommendation systems. Just the introduction of a reasonable level of randomness alone makes the effects described above disappear or even push them in the opposite direction.
However, you talked about developing tools and strategies to harness AI in an “altruistic,” community-focused way.
If some degree of coordination were introduced and AI were used to promote cooperation, benefits could be achieved that are as yet completely unexplored. Going back to the traffic example, if algorithms tried to promote more collectively effective solutions, emissions could be drastically reduced according to our simulations. I believe that acting on this area could lead to major reductions in information pollution, which has a significant impact on the polarization and radicalization present in our society. Creating tools that transparently promote nudging practices, influencing individual choices to try to promote diversification and collective well-being, could be important to better embrace the complexity of social systems. People naturally have a tendency to conform as much as possible; excessive conformity means that we do not adequately take advantage of all the opportunities that a complex system offers. We are unable to see how making a nonconformist choice can actually sometimes benefit us greatly, and even in areas such as remanufacturing, trying to diversify processes and raw materials could lead to better efficiency and collective benefits. Of course, there is no need to try to maximize the random factor in every case, but to better understand the trade-offs between diversification and standardization.
How can the techniques discussed, and the study of complex systems in general, be used to overcome some of the biggest obstacles that the energy transition will cause, namely the loss of flexibility in response to peak moments due to the inherent characteristics of renewables versus fossil fuels?
I believe it is important to use both machine learning and more generally the study of complex systems to ensure ideal conditions in the transition from the current hierarchical model to the distributed model that would be observed in a renewable-dominated power grid. A well-designed distributed grid at the topology level even when deficient at the level of some nodes, which turn out to be less resilient, could be more stable overall. In fact, the centralized model on which we now rely can come under strain in the event that a sequence of events damages the connections or operation of power plants that are at the top of the hierarchy, as can be seen from some of the blackouts that have occurred in recent years in the East Coast of the United States. In a network with widespread structure, single catastrophic events would no longer be a risk. In this area, AI could be used to build models that better predict demand structure and better estimates of the productivity of individual or diffuse small-scale plants. There are some current studies thinking about using each resource creatively, for example should we be able to further improve efficiency and promote the spread of many electric cars, some groups are hypothesizing the potential of using cars as a tool to move energy around the grid.
How can data mining and network science contribute to the management of complex systems to identify areas that could have the greatest impact on climate?
The use of data mining and network science can help us identify critical points, and this can further lead us to better design networks, improving overall efficiency. This approach uses complex multi-level networks, where we study complex systems by creating networks that can deal with either social, economic, or demographic aspects. This type of studies helps us identify what some potential crisis points might be and react to extraordinary situations. These kinds of approaches will be indispensable for reimagining topics such as energy transportation and the mobility of people in an optimal way, and they will depend especially on the increasing availability of data.
What are the main challenges related to bias in artificial intelligence models?
Historically in network science and data science studies, great attention has always been paid to the underlying biases of the models constructed. This has changed recently with the proliferation of AI models that are effectively black boxes. In our lab, we are trying to superimpose a level of “explainability” on deep learning and large language models (such as ChatGPT) and are trying to identify and make clear what biases these tools contain, in order to protect especially the most at-risk categories. At the same time, we are trying to develop new models that are more transparent from the beginning. We are not the only ones and it is possible to find a number of best practices already; after all, the absence of transparency in the most popular business models is not a necessary feature, although it is a pervasive one. Another major problem is that the popularization of these tools has been such that several companies have been pushed to spread models even when they were still immature, with consequences that are difficult to understand at the moment.
There also needs to be a broader reflection on this issue, as ill-considered use of tools such as large language models by people who are not aware about their limitations can cause great problems. For example, going back to the discussion of how AI can lead to a loss of diversity, one must keep in mind that most of these tools function by following a monolithic worldview, functioning mostly in English and promoting extreme homogenization, leading almost to a form of information colonialism.
Photo credits: NASA