Awad2018MoralMachine-A

#multipleEntries

There are multiple notes about this text. See also: Awad2018MoralMachine-A

Edmond Awad, Sohan Dsouza, Richard Kim et al. "The Moral Machine experiment."

Bibliographic info

Awad, E., Dsouza, S., Kim, R. et al. The Moral Machine experiment. Nature 563, 59–64 (2018). https://doi-org.proxy.library.uu.nl/10.1038/s41586-018-0637-6

Commentary

This research by Awad et al. created the biggest survey ever done on global moral preferences for decisions made by ESDiT-shared - OLD 2021-11-03/esditHuman/self-driving cars in case of life-threatening dilemmas. Thereby, this research facilitated a way to quantify societal expectations about the ethical principles that should guide machine behavior on a scale never seen before. The set-up of this survey also carries some weaknesses with it, for instance, the self-selection bias and the simplified version of factors are weaknesses that are discussed in the sections below. Another weakness that is worth mentioning is the design of the interface which is set up in a playful way, thereby risking gamification: people might fill out this survey as a game, thereby not taking the survey as seriously anymore, which might influence their answers. However, it is understandable that the survey is set up this way, as this interface probably did result in the immense popularity of the survey, thereby contributing to the huge amount of answers that are collected.

Excerpts & Key Quotes

Moral Machine self-selection bias

"..we designed the Moral Machine, a multilingua online 'serious game' for collecting large-scale data on how citizens would want autonomous vehicles to solve moral dilemmas in the context of unavoidable accidents."

Comment:

One problem with the way this survey is set up is that it involves a self-selection bias: individuals choose themselves to go to the moral machine page and 'play' the game. This might imply that the moral decisions are based on a sample of people that might be the most interested in self-driving cars, thereby biased towards a particular world view on these machine dilemmas. It might have been better to set up the survey in a different way, for example through probability sampling, thereby reducing the risk for a self-selection bias. On the other hand, this non-probability way of sampling did result into a huge amount of responses, namely 39.61 million decisions from 233 countries, thereby also reducing the chance of the self-selection bias.

Cultural differences

Comment:

One interesting finding from this study that will have a huge impact on the development of self-driving cars is the fact that preferences are not global but differ per area in the world. As the quote states, Asian countries do not have such a strong preference to spare younger characters over older as other parts of the world have. How would we implement moral preferences in autonomous cars then? We might need to design machine software containing ethics based on the country where the car drives so that the car is most in line with the ethics from that particular country. However, this might be hard to implement in practice as cars drive all around the world and cross borders all the time.

"In the main interface of the Moral Machine, users are shown unavoidable accident scenarios with two possible outcomes, depending on whether the autonomous vehicle sweres or stays on course ... focuses on nine factors: sparing humans (versus pets), staying on course (versus swerving), sparing passengers (versus pedestrians), sparing more lives (versus fewer lives), sparing men (versus women), sparing the young (versus the elderly), sparing pedestrians who cross legally (versus jaywalking), sparing the fit (versus the less fit), and sparing those with higher social status (versus lower social status)."

Comment:

This passage shows that the Moral Machine experiment has a quite simplified interface compared to reality: participants could only choose between two scenarios, whilst, in reality, there will often be more than two possibilities. Furthermore, the moral decisions only contained nine factors. The factors are presented as opposites, however, in real life, the difference in these factors will often be not so big as presented in the survey. For example, fit vs non fit will not always be visible. In addition, gender is now constrained to two options, man or woman, however, as genderqueer becomes more common in certain parts of the world and thereby on the street view, the survey might need to update their presentation of gender. In general, it might have been good to extend these nine factors and to extend the inputs within these factors. However, as it is also important to get a high response rate, it is also good to simplify the street view to only a few factors as we need to take the length of the survey into consideration.