Coordinating human and agent behaviour in collective risk scenarios

Resumo

Collective hunting, antibiotic abuse, vaccination hesitancy, climate action or even coordinating the population in order to comply to the covid19 regulations are some of the several examples of collective endeavours that entail a collective risk. These complex scenarios correspond to social dilemmas that are subject to many uncertainties, ranging from the actual risk of the catastrophe occurring, to the amount of collective effort required to avoid it or even the time available to take the required measures. Additionally, such dilemmas are no longer constituted solely of human-human interactions, but also, increasingly often, of human-agent interactions through socio-technical systems already deployed at different levels of society. This adds a new layer of complexity which is still mostly unexplored. In this thesis, we perform and analyse a series of behavioural economic experiments in a collective risk dilemma (CRD), which abstract the previously listed scenarios wherein participants are confronted with a choice: contributing sufficiently over several rounds to achieve a collective target that avoids the catastrophe, or defect and assume that others will make sufficient contributions, and thus, aiming to maximise one’s personal gains. We use this framework to gain new insights on how uncertainties affect the outcome of collective risk scenarios and the behaviour of its participants, as well as how these results are affected by the presence of (artificial) autonomous agents. To this end, we perform a series of behavioural economic experiments divided in two parts: i) first, we study the effect of uncertainty about the time available to avoid the catastrophe (timing uncertainty), ii) then, we investigate how the presence of autonomous agents affect the choices of participants in the CRD and the outcome of the dilemma. The first set of experiments is motivated by previous experimental results which explore the effect of uncertainties about the impact (or risk) of the disaster (impact uncertainty), as well as the required contributions to avoid it (threshold uncertainty). These studies indicate that people only tend to contribute to avoid the disaster once they perceive the risk to be high, yet even in this case, cooperation collapses once the placement of the collective target is uncertain. Nevertheless, there was still no experimental results about the effect of timing uncertainty. We hypothesise that, as happens for threshold uncertainty, timing uncertainty will produce significative behavioural changes in the CRD. The second study is motivated by the increasing importance of autonomous agents in our society. These agents are already present in many call centres (e.g., chat-bots), in social networks (e.g., Twitter bots) and autonomous vehicles. In our experimental study, we have designed three experiments in which, respectively, participants have to delegate, customise or interact in a hybrid human-agent group. Previous experimental results indicate that human participants tend to program agents to which they will delegate their actions with pro-social behaviours. Yet, these results seem to depend on the context of the strategic interaction. We hypothesise that delegation to autonomous agents will also increase group success in the CRD.