Using the Double Transparency of Autonomous Vehicles to Increase Fairness and Social Welfare

Abstract

Fully autonomous vehicles (AVs) create double transparency regarding human driving decisions. Opaque decision rules in the human mind have become transparent in AVs, and in turn, can be made transparent to third parties. This double transparency is creating an unprecedented opportunity to regulate driving decision rules to eliminate unreasonable selfishness and increase fairness and social welfare because AVs can be programmed to follow regulations 100% of the time. In this experimental ethics study, we performed an incentive aligned online experiment to examine humans’ willingness to sacrifice other people’s lives to protect their own in five different accident scenarios and to investigate the potential for AV regulation to curb unreasonable selfishness, thereby increasing fairness and social welfare. Our results reveal the need to regulate rules governing AV driving decisions; yet, a full transparency policy for decision algorithms may not necessarily lead to desired social effects. Thus, regulations should be tailored to different scenarios.

Appendices

Appendix 1. Survey

1.1 Neither Party at Fault

A driverless car with one passenger drives at the speed limit down a main road. However, due to a traffic light failure, the driverless car will hit N pedestrians. The AV has two options: (a) stay on the path and move forward, in which case the car will hit and cause serious harm to the pedestrians, but the passenger in the AV will not be harmed, or (b) swerve off to the other side of the road, hit a barrier, and roll over, in which case the passenger will suffer serious harm, but the pedestrians will remain unharmed.

1.2 AV at Fault

A driverless car with one passenger drives at the speed limit down a main road. Due to a brake failure, the driverless car will run into N pedestrians. The AV has two options: (a) stay on the path and move forward, in which case the car will hit and cause serious harm to the pedestrians, but the passenger in the AV will not be harmed, or (b) swerve off to the other side of the road, hit a barrier, and roll over, in which case the passenger will suffer serious harm, but the pedestrians will remain unharmed.

1.3 Other Party at Fault

A driverless car with one passenger drives at the speed limit down a main road. Suddenly, N pedestrians ignore the red light, and the driverless car will hit N pedestrians. The AV has two options: (a) stay on the path and move forward, in which case the car will hit and cause serious harm to the pedestrians, but the passenger in the AV will not be harmed, or (b) swerve off to the other side of the road, hit a barrier, and roll over, in which case the passenger will suffer serious harm, but the pedestrians will remain unharmed.

1.4 Both Parties at Fault

A driverless car with one passenger drives at the speed limit down a main road. The driverless car’s brakes fail, and N pedestrians suddenly ignore a red light. The driverless car will run into N pedestrians. The AV has two options: (a) stay on the path and move forward, in which case the car will hit and cause serious harm to the pedestrians, but the passenger in the AV will not be harmed, or (b) swerve off to the other side of the road, hit a barrier, and roll over, in which case the passenger will suffer serious harm, but the pedestrians will remain unharmed.

1.5 “Good Samaritan”

A driverless car with a passenger sees that another vehicle ahead is accelerating towards N pedestrians on the sidewalk. That vehicle will rush into N pedestrians, causing serious harm. The AV has two options: (a) speed up and hit the vehicle to incapacitate it and avoid pedestrian casualties, seriously harming its passenger in the process, or (b) not intervene and let the accident happen to prevent harm to its passenger.

1.6 Measurement Items

Each participant was presented with all scenarios, one at a time in random order, and asked to state his or her WTS, usage intentions and purchase intentions if the stated WTS could be set, and perceptions of fairness under policies based on full transparency and non-transparency. WTS is a quantitative measurement defined as the maximum number of humans a person will harm in order to save himself/herself. The participants were asked to set their WTS between 0 and 99, with 99 representing an infinitely large WTS.

1.1 You are the sole passenger in an AV. Under the full transparency policy, how would you set WTS in this scenario?

1.2 You are the sole passenger in an AV. Under the non-transparency policy, how would you set WTS in this scenario?

2.1 Under the full transparency policy, would you ride in a driverless car with the same WTS setting as yours for this scenario, assuming you are satisfied in all other aspects of the driverless car? (Please rate using a scale ranging from 1 to 9: 1—definitely would not; 9—definitely would).

2.2 Under the non-transparency policy, would you ride in a driverless car with the same WTS setting as yours for this scenario, assuming you are satisfied in all other aspects of the driverless car? (Please rate using a scale ranging from 1 to 9: 1—definitely would not; 9—definitely would).

3.1 Under the full transparency policy, would you buy a driverless car with the same WTS setting as yours for this scenario, assuming you are satisfied in all other aspects of the driverless car? (Please rate using a scale ranging from 1 to 9: 1—definitely would not; 9—definitely would).

3.2 Under the non-transparency policy, would you buy a driverless car with the same WTS setting as yours for this scenario, assuming you are satisfied in all other aspects of the driverless car? (Please rate using a scale ranging from 1 to 9: 1—definitely would not; 9—definitely would).

4.1 In this scenario, please rate the fairness of the full transparency policy. (Please rate using a scale ranging from 1 to 9: 1—very unfair; 9—very fair).

4.2 In this scenario, please rate the fairness of the non-transparency policy. (Please rate using a scale ranging from 1 to 9: 1—very unfair; 9—very fair).

1.7 Policy Types

Full transparency policy: The government regulates that every driverless car is required to disclose its “Willingness to Sacrifice (WTS)” to all road users.

Non-transparency policy: Driverless cars do not need to disclose “Willingness to Sacrifice (WTS)”.

Appendix 2. Data and Regression

1.1 “Good Samaritan” Scenario

In the incentive aligned group, 60.86% of participants (responding as passengers) allowed their AV to act as a “Good Samaritan” to save others (Fig. 3). The mean WTS is 70.00 (the intercept of the right censored regression) and the median is 49 for this scenario; among those willing to consider such an intervention (WTS < 99), the mean is 24.31 and median is 9.

Fig. 3

The distribution of willingness to sacrifice (WTS) in the “Good Samaritan” scenario (incentive aligned group)

Full transparency also increases perceptions of fairness when AVs are uninvolved (p < 0.001). The results of the effect of transparency on social welfare are consistent with other scenarios (Fig. 4), except there are fewer participants in segment 1 in the “Good Samaritan” scenario (p = 0.037).

Fig. 4

Changes in usage intentions, purchase intentions, and perceptions of fairness under full transparency policy in the “Good Samaritan” scenario. Boxes show the 95% CI of the mean

With regard to the “Good Samaritan” scenario, our findings reveal an important and challenging regulatory quandary: Should society allow such preferences to be programmed into an AV? While it may not be reasonable to set a ceiling in this scenario since the AV is not involved and should not be forced to intervene, should society set a floor value for WTS to prevent overzealous actions by AVs? The decision is a double-edged sword for society. Allowing such actions would decrease total fatalities from a societal perspective, but would cross an important line regarding the roles intelligent robots such as AVs should play in human society. Allowing AVs to act as “Good Samaritans” through the use of deadly force sets a precedent with far-reaching ramifications for future robot-human relationships.

1.2 Comparing the Incentive Aligned Group and Control Group

Respondents in the control group penalized the other party less when the other party was at fault (17.69 vs. 20.02 for the incentive aligned group) (Table 1 and Fig. 5). The average WTS for the control group increased by 50.69% (vs. a 59.78% increase for the incentive aligned group) compared to the neither party at fault scenario (Table 1 and Fig. 5). The incentive and control groups also behave slightly differently under the full transparency policy (Figs. 2 and 6).

Fig. 5

The distribution of willingness to sacrifice (WTS) in five scenarios (control group)

Fig. 6

Changes in usage intentions, purchase intentions, and perceptions of fairness under full transparency policy for the control group. Boxes show the 95% CI of the mean