Ethics of Autonomous Vehicles: Australians’ Expectations and Moral

Preferences

Amir Raﬁee

School of Information and Communication Technology, Grifﬁth University, Brisbane, Australia

ﬁ ﬁ

Keywords:

Autonomous Vehicle, AI Policy, AI Ethics, Automated Driving System, AI Regulation, Self-Driving Cars,

Ethical Dilemmas.

Abstract:

Autonomous Vehicles (AVs) can handle most driving scenarios, but ensuring safety in every situation remains

a challenge. Factors such as technology failures, faulty sensors, and adverse weather introduce complex eth-

ical dilemmas that AVs must navigate. Considering the societal beneﬁts of AVs, it is crucial to address both

technical challenges and ethical expectations. This paper evaluates Australians’ perceptions and expectations

regarding the ethical programming of personal AVs in six dilemma scenarios using a structured question-

naire. The participants selected the most acceptable outcome in each scenario, informed by ethical and legal

considerations. The survey offers a framework for understanding public moral preferences by excluding dis-

criminatory factors and considering legal contexts. The ﬁndings prioritise Australians’ preferences for ethical

AV behaviour, focusing on Injury Over Sacriﬁce (IOS), Harm Conﬁnement and Lawfulness (HCL), and Harm

Prevention and Prioritisation (HPP). These insights can guide policymakers and manufacturers in aligning AV

programming with societal values. The study also highlights how ethical models like the Objective Decision

System (ODS), which selects outcomes randomly when no clear moral preference emerges, can balance public

trust and responsibility in AVs.

1 INTRODUCTION

Autonomous Vehicles (AVs) can manage most driv-

ing scenarios with relative ease; however, designing a

system that ensures safety in every situation remains

highly challenging (Campbell et al., 2010). For exam-

ple, recognising humans and other objects on the road

is both critical and more difﬁcult for AVs than for hu-

man drivers (Farhadi et al., 2009). Consequently, in

future AVs, crash avoidance features alone will not

sufﬁce. It may not always be possible to avoid an ac-

cident, especially as AVs navigate city streets and av-

enues, which are more dynamic than highways. Fac-

tors such as technology errors, faulty sensors, mali-

cious actors, and bad weather can contribute to in-

evitable collisions (Gomez et al., 2014). Given the

risks associated with fully autonomous vehicles, it is

clear that they require a set of principles to govern

their utility. Moreover, for individuals to trust AVs,

their design must align with ethical and inclusive val-

ues. This has led to a global response, with nations

addressing the emerging ethical issues surrounding

AI-enabled technologies.

https://orcid.org/0000-0002-8149-4780

The major challenge, however, extends beyond

ethical dilemmas. If we assume that the widespread

adoption of AVs is necessary for the societal bene-

ﬁts these vehicles are expected to provide (Bonne-

fon et al., 2016), it is evident that any reasonable

ethics policy should also consider the various expec-

tations of users and the larger society in which they

are implemented. This necessitates ensuring user sat-

isfaction and safety, as well as other essential design

values such as trust, accountability, and transparency

(IEEE, 2016). Furthermore, ethical theories and peo-

ple’s expectations often overlap, even if they will al-

ways be somewhat in tension. Therefore, it appears

that an ideal ethics policy must resolve this tension

by balancing public acceptability and moral require-

ments. It must be acceptable enough to gain trust and

adoption from users while remaining moral enough to

avoid reﬂecting negative human tendencies (Shahriari

and Shahriari, 2017).

Some scholars have used polling and surveys to

understand public opinions on these moral issues

(Awad et al., 2018; Bonnefon et al., 2016). Although

these polls can assist in developing moral algorithms,

they cannot determine their implementation. There-

Raﬁee, A.

Ethics of Autonomous Vehicles: Australians’ Expectations and Moral Preferences.

DOI: 10.5220/0013123900003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 1, pages 287-297

ISBN: 978-989-758-737-5; ISSN: 2184-433X

287

fore, while public polling on how AVs should han-

dle dilemma scenarios can inform regulation, there is

a compelling argument that the ﬁnal decision should

be left to experts (Bonnefon et al., 2015). Nonethe-

less, a consensus is necessary, and important ethical

decisions should not be left solely to engineers or

ethicists. Algorithms that do not align with societal

moral expectations or consumer preferences are likely

to hinder the smooth adoption of AVs. Consequently,

any solution for programming moral dilemmas in AVs

should not discourage potential buyers or the public at

large.

In this paper, I aimed to evaluate Australians’ per-

ception and expectation regarding personal AVs re-

lating to various ethical settings. Section 2 contains

the deﬁnitions and considerations which will be used

throughout the paper. In Section 3, I critically as-

sessed the Moral Machine (MM) experiment, which

aimed to establish a global representation of peo-

ple’s moral preferences in various AV dilemma sce-

narios. I evaluated this study and explained why cer-

tain dilemma scenarios and participant options should

be excluded from future research. In Section 4, I ar-

gued for the inclusion of random selection as an op-

tion in the survey questionnaire. I discussed the data

collection method in Section 5, followed by an anal-

ysis of the data and ﬁndings in Section 6. My pro-

posed ethical policy and AV ethics framework were

presented in Section 7. In conclusion, I revisited the

primary research objective and summarised the ﬁnd-

ings, aiming to propose a publicly acceptable and eth-

ically sound framework for programming AVs.

2 DEFINITIONS AND

CONSIDERATIONS

In this section, I provide the necessary deﬁnitions and

considerations that form the foundation for the re-

mainder of the paper. The aim was to clarify the key

concepts and frameworks discussed throughout, en-

suring a common understanding of the topic at hand.

2.1 Autonomous Vehicles

There are a wide range of AVs, which supplement or

replace human drivers with AI. Meaning that a human

driver either has limited responsibility or does not

need to be present at all. The Society of Automotive

Engineers (SAE) deﬁnes 6 levels of automation that

apply to automotive vehicles. Levels 0–3 require that

human drivers take control of various vehicle opera-

tions under certain conditions, and levels 4 and 5 have

no such requirements (Shadrin and Ivanova, 2019). In

this paper, AVs refers to autonomous vehicles at level

4 or 5 of driving automation, as deﬁned by the SAE.

These are vehicles where no driver is ever needed, or

there might be an option for human override but not a

requirement.

3 RESEARCH QUESTION

DESIGN

There were six dilemma situations shown in the sur-

vey questionnaire involving an AV, and the partici-

pants were asked to decide which outcome was the

most acceptable to them. The purpose of the survey

questionnaire was to evaluate respondent’s perception

and expectation on personal AVs relating to various

ethical settings.

I also proposed and included a theoretical solution

called an objective decision system (ODS). If selected

by a participant, the AV would consider all feasi-

ble and justiﬁed outcomes for a given crash scenario,

based on a philosophical or moral framework, and

choose one at random. Numerous alternatives were

excluded because they were either unjustiﬁable or dis-

criminatory. In the questionnaire, I asked participants

ﬁve demographic questions, including gender, age,

education level, their perceived consumer adopter cat-

egory, and whether they currently own a Level 2 AV.

Adopter categories divides consumers based on their

willingness to adopt new innovations. Understanding

these categories helps in analysing preferences toward

AV crash scenarios. Using the survey responses and

the demographic variables, I then compared people’s

preferences towards a pre-determined outcome based

on the different crash scenarios.

3.1 Exclusion Criteria

The MM Experiment, surveyed people across hun-

dreds of countries to gauge moral preferences in AVs

and what priorities they should have in the event of

an unavoidable accident. The researchers used an on-

line survey to collect over 39 million responses to

hypothetical ethical dilemmas for AVs (Awad et al.,

2018). The strongest preferences were for sparing hu-

man lives over animal lives, sparing more lives, and

sparing young lives. There was a general indication

by the results, that there was more preference for spar-

ing children’s lives, rather than those of adults. No-

tably people from different parts of the world had dis-

similar beliefs on how AVs should make such life and

death decisions.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

288

3.1.1 Human Life vs Animal Life Decisions

One of the many crash scenarios in the MM exper-

iment involving an AV, presented participants with

a choice between saving a human life or an animal.

Generally, when an animal appears directly in front

of a vehicle and there is not enough time to brake

or swerve (with road conditions being a factor), the

safest option is to continue ahead and strike the an-

imal (Curtis and Hedlund, 2005). Currently, drivers

are not legally liable if they hit a wild animal, such

as deer, elk, and other smaller animals. It can be

assumed that this would also apply to AVs. Addi-

tionally, German Ethical Rule 7 clearly states that

in dilemma situations, the protection of human life

should have top priority over the protection of animal

life. This rule aligns with social expectations assessed

through the MM experiment (Luetge, 2017). There-

fore, I excluded similar scenarios from the study, as

it is apparent that in an imminent crash where an AV

must choose between saving an animal life or a hu-

man life, the human life should take precedence.

3.1.2 Discriminatory and Immoral Decisions

Another crash scenario in the MM experiment in-

volved an AV faced with an imminent collision, re-

quiring participants to choose between saving an el-

derly or a young person. The justiﬁcation of discrimi-

nation is crucial when analysing trolley-problem-like

scenarios. Philosophers distinguish between non-

arbitrary discrimination, which is morally justiﬁed,

and arbitrary discrimination, which is not (Rachels

and Rachels, 2012). For example, excluding blind ap-

plicants from air trafﬁc control jobs is non-arbitrary

because visual acuity is essential for safety. In con-

trast, excluding individuals based on age, race, or gen-

der without relevant justiﬁcation is arbitrary and un-

just. In crash scenarios, age-based discrimination is

morally equivalent to other forms of arbitrary discrim-

ination, such as those based on race or gender (Lin,

2013). Therefore, discriminatory scenarios have been

excluded from this survey research.

3.2 Inclusion Criteria

3.2.1 Legal Implications

The MM experiment, presented as an applied trolley

problem based on Thompson’s case, focused solely

on moral responsibility while neglecting the legal is-

sues that could signiﬁcantly impact participants’ deci-

sions and constraints on their rights to action (Etienne,

2020). As observed in other experiments, real con-

ditions can inﬂuence participants’ decisions (Francis

et al., 2017). People’s moral choices often ﬂuctu-

ate and are heavily affected by the amount of infor-

mation available and the degree of deliberation in-

volved. Many participants might contend that, with

more information, their decisions could differ (Nooth-

igattu et al., 2018). Therefore, in the survey research,

I presented the legal implications of each of the six

dilemma scenarios for participants to consider, ensur-

ing their decisions were informed by both ethical and

legal perspectives.

4 THE CASE FOR RANDOM

SELECTION

4.1 Introduction

In this section, I draw upon existing literature to

present arguments supporting the use of random se-

lection in future AVs, particularly in dilemma situ-

ations where, from a legal perspective, all available

outcomes are justiﬁable on philosophical or moral

grounds. I contend that lotteries can eliminate bias,

maintain ethical standards by excluding immoral or

illegal options, and ensure transparent, tamper-proof

outcomes. Decision-making by lot is likely to be sim-

ple, objective, and cost-effective, making it a serious

consideration for researchers, AV manufacturers, and

policymakers. This approach ensures that AVs navi-

gate ethical challenges justly and equitably, promot-

ing trust and acceptance of AV technology.

4.2 Random Selection Applied to AVs

Imagine an AV navigating a busy intersection when

a pedestrian suddenly steps into the road. The AV

faces two immediate options, brake to avoid hitting

the pedestrian, potentially causing a collision with a

closely following motorcyclist; or continue straight,

hitting the pedestrian but avoiding harm to the mo-

torcyclist (Coca-Vila, 2018). In this scenario, both

options have valid moral justiﬁcations: prioritising

pedestrian safety versus preventing harm to the mo-

torcyclist. Rational analysis based on safety metrics

and trafﬁc conditions might fail to determine a clear

superior option. Additionally, from a normative view-

point, they constitute a symmetrical community of

danger in which saving the life of one is only possible

at the cost of the life of the other. According to crim-

inal law, when two duties to act are in conﬂict, the

self-driving car has to decide which interest to pro-

tect and, as such, it will always be acting according

to the law (Zimmermann, 2014). This means that, in

Ethics of Autonomous Vehicles: Australians’ Expectations and Moral Preferences

289

the same way that a father is permitted to breach the

obligation to save one child from drowning by sav-

ing the other, the self-driving car in the example also

has to decide who will suffer the harm awaiting both

victims, without causing a transfer of harm to an in-

nocent third party (Coca-Vila, 2018). In short, brak-

ing the car to prevent running over the pedestrian is

not, therefore, unlawful homicide, and neither would

be not stopping and hitting the pedestrian crossing the

road in front of the car. If both possible outcomes are

acceptable, then the car can be programmed to decide

through a random decision system (RDS) (Zimmer-

mann, 2014). Consequently, the AV ethics setting ac-

tivates its RDS to decide whether to brake or continue

straight.

One of the primary beneﬁts of using a randomised

decision in AVs is its ability to neutralise biases. A

random decision system eliminates these biases by

making the choice impartially, solely based on chance

rather than subjective factors. Critics might argue

that using a random decision system bypasses ratio-

nal thought and deliberation, potentially leading to

suboptimal outcomes; however, where no decision-

making parameter exists and the alternative of not

making a decision has been rejected, randomness is

seen as the best solution to solve undecidable conﬂicts

(Rescher, 1960).

First, employing a lottery-based decision-making

process ensures that the outcome is not inﬂuenced

by arbitrary or subjective factors. This approach up-

holds fairness and prevents the vehicle from making

biased decisions based on irrelevant considerations

(Dworkin, 2011). Secondly, random decision systems

in AVs plays a critical role in preserving ethical stan-

dards. By excluding immoral or illegal options from

consideration and focusing solely on morally accept-

able choices, AVs uphold ethical norms and legal reg-

ulations. This ensures that decisions align with soci-

etal values and legal frameworks, promoting trust and

acceptance of AV technology among stakeholders and

the general public.

Thirdly, the fact that the car makes a random deci-

sion and automatically acts on it removes any risk of

manipulation, both in the choice and implementation

of the solution. Therefore, the victim will know that

all outcomes are considered equally, and his faith is

resting in the hand of an objective force (Coca-Vila,

2018). Duxbury (1999) also argues that a decision

made by lot, offers a fair way of dealing with many

uncomfortable, or even inherently unfair, dilemmas.

A non-biased lottery would remain impartial and thus

less susceptible to corruption (Duxbury, 1999). Fi-

nally, decisions determined through random selection

lack a human element, suggesting that accountability

for challenging choices can be alleviated from spe-

ciﬁc individuals or groups (Duxbury, 1999). More-

over, a non-random system may be more suscepti-

ble to manipulation. For instance, if pedestrians be-

lieve the system always favours them, they might be-

come careless. Even more concerning, malicious ac-

tors could exploit this predictability to stop vehicles

for their own purposes. By eliminating complete pre-

dictability, a random system discourages both reck-

less behaviour and exploitation.

5 DATA COLLECTION METHOD

This research employed a descriptive survey design to

effectively gather original data from a population too

large to observe directly (Botes, 1996). The descrip-

tive survey design aligns with the study’s objective,

which aimed to assess respondents’ perceptions and

expectations regarding personal AVs in various ethi-

cal settings. Participants were drawn from the gen-

eral Australian population aged 18 and above, with

Australia’s diverse demographics, including its status

as the country with the highest immigrant proportion

among nations with populations over 10 million, en-

suring a representative sample. The Australian popu-

lation was 25 million as of December 31, 2020 (Aus-

tralian Bureau of Statistics, 2021). Data collection

was conducted via SurveyMonkey, targeting eligible

respondents across various age categories. The plat-

form stored responses securely, and the data was sub-

sequently analysed using SPSS.

The chi-square analysis of variance was chosen

as the statistical method to evaluate the survey ques-

tionnaire results due to its suitability for examin-

ing relationships between categorical variables, which

aligned with the research questions posed. The conﬁ-

dence level was set at ninety-ﬁve percent to ensure

robustness in the conclusions drawn from the data.

SPSS software was used, speciﬁcally employing the

cross tabulation feature under Descriptive Statistics to

generate a contingency table displaying frequencies

and column percentages. Detailed statistical analyses

and accompanying charts are available upon request;

however, due to word limit constraints, they have been

excluded from this submission.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

290

6 DATA ANALYSIS AND

FINDINGS

6.1 Overview

To sufﬁciently understand the results, readers are ad-

vised to familiarise themselves with the scenarios out-

lined in the survey questionnaire. The following pro-

vides a summary of the 6 AV crash scenarios, the as-

sociated moral dilemmas, and the legal implications,

offering the necessary context to understand the re-

sults discussed in this section. As noted in Section 3,

a theoretical solution called the ODS was proposed

and included. If selected by a participant, the AV

would consider all feasible and justiﬁed outcomes for

a given crash scenario based on a philosophical or

moral framework and choose one at random.

Scenario 1 - Brakes Failure on Mountain Road

The AV faces a dilemma where its brakes have failed,

and it must decide between swerving right to hit a by-

stander on the mountain or continuing straight off a

cliff, resulting in the death of the passenger. Legally,

the bystander is not responsible for the emergency sit-

uation or the AV’s mechanical failure, which means

the AV cannot legally prioritise saving its passenger

by causing harm to the bystander.

Figure 1: The ratio between AV passengers and bystanders

was equal, with one passenger for every bystander.

Scenario 2 - Intersection with Illegally Crossing

Pedestrians

An AV with one passenger encounters a dilemma at

an intersection where ﬁve pedestrians are illegally

crossing the road. The AV has two options: it can

either swerve left to avoid the pedestrians, potentially

killing a cyclist, or continue straight, resulting in the

likely injury or death of one or more of the pedes-

trians. The AV’s passenger will remain unharmed in

either case. Legally, the pedestrians are at fault due to

their illegal crossing, and the AV could justify its de-

cision to continue straight as a defensive emergency,

meaning it would not be breaking the law by priori-

tising the cyclist’s safety over that of the pedestrians

who caused the dangerous situation.

Figure 2: Five times as many people illegally crossed the

road compared to the innocent cyclist. The AV passenger

remains unharmed regardless of the outcome.

Scenario 3 - Highway Dilemma with Fallen Pallet

The AV with a single passenger encounters a high-

way dilemma when a pallet falls from the truck ahead.

The AV has three options: continue straight, resulting

in the passenger’s death from colliding with the pal-

let; swerve right, leading to the death of a helmetless

motorcyclist; or swerve left, causing severe injury but

survival of a fully geared motorcyclist, with no harm

to the passenger in these latter scenarios. The helmet-

less motorcyclist’s illegal activity doesn’t prevent his

death if hit, while the fully geared motorcyclist may

suffer bone fractures but can seek compensation. This

scenario poses a moral dilemma between prioritising

preserving life versus avoiding physical harm.

Scenario 4 - Pedestrian vs. Motorcyclist

The AV faces a dilemma where it must decide be-

tween two potentially fatal outcomes involving a

pedestrian and a motorcyclist. If the AV brakes to

save the pedestrian who has suddenly appeared in its

path, the motorcyclist following closely behind may

collide with the AV and be killed. Alternatively, the

AV could continue on its course, striking the pedes-

trian and saving the motorcyclist. Importantly, the

passenger in the AV remains unharmed in either sce-

nario. From a legal perspective, both the pedestrian

and the motorcyclist could be considered negligent

Ethics of Autonomous Vehicles: Australians’ Expectations and Moral Preferences

291

Figure 3: The distribution between AV passenger and mo-

torcyclists was even, indicating that in all three possible

outcomes, either one passenger or one of the motorcyclists

would be harmed.

for their actions, which contributed to the situation.

However, the AV and its passenger would not bear

legal responsibility for the outcomes because the AV

took reasonable measures to avoid foreseeable harm,

thus absolving them of liability under the current legal

frameworks.

Figure 4: The ratio between the negligent pedestrian and

motorcyclist was equal, at one to one. The AV passenger

remains unharmed regardless of the outcome.

Scenario 5 - Intersection with Legally Crossing

Pedestrians

The AV encounters an ethical dilemma at an intersec-

tion where ﬁve pedestrians are crossing legally with

a green light. The malfunctioning AV must decide

between swerving into a wall, risking the lives of its

ﬁve passengers to save the pedestrians, or proceeding

straight, potentially endangering one or more pedes-

trians to protect its passengers. Legally, the pedestri-

ans are in the right as they cross with the green light,

implying no fault on their part. This legal framework

highlights the ethical challenge posed by AV technol-

ogy, where programming decisions must conform to

legal norms while addressing intricate moral dilem-

mas in unpredictable situations.

Figure 5: The number of pedestrians legally crossing the

road equalled the number of AV passengers illegally cross-

ing the intersection, with both totalling ﬁve instances each.

Scenario 6 - Intersection with Legal Pedestrian

Crossing

The AV faces a critical decision during loss of con-

trol at an intersection. It can either swerve to the

right, colliding with a tree and resulting in the deaths

of three passengers, or continue straight through the

intersection, striking a pedestrian legally crossing on

the zebra lines, thereby saving the passengers. The

legal implication centres on the pedestrian’s right of

way, as they are crossing legally at the zebra lines,

prioritising pedestrian safety over vehicular move-

ment. This legal context adds complexity to the ethi-

cal dilemma faced by the AV, highlighting the poten-

tial clash between moral imperatives and legal respon-

sibilities in autonomous driving scenarios.

6.2 Discussion

In this section I summarised the participants’ moral

preferences, which could be used to develop an ethi-

cal policy for AV decision-making. I then compared

these preferences to the three main philosophical the-

ories which have been extensively discussed in the

context of AVs, speciﬁcally utilitarianism, deontolog-

ical ethics and the Rawlsian approach, and evaluated

which framework best aligns with the proposed eth-

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

292

Figure 6: The ratio of AV passengers illegally crossing the

intersection compared to pedestrians legally crossing the

road was 1 to 3.

ical principles for AV programming (Raﬁee et al.,

2023).

6.2.1 Australians Moral Preferences

Objective Decision System (ODS). Overall, the ODS

was the most preferred outcome among all partici-

pants. The use of the ODS could be justiﬁed as it

provides a straightforward and comprehensible solu-

tion that appeals to a broader audience. It offers a

common-sense approach that resonates with a larger

population. This approach could be particularly use-

ful in the early development stages of AVs at levels

4 or 5 of driving automation, as discussed in Sec-

tion 2, where no driver is needed, or there might be

an option for human override but not a requirement.

These AVs may still lack the technological capabili-

ties to quickly assess all viable outcomes and make

a split-second decision in a moral dilemma involving

potential loss of life. Such technological limitations

could delay their deployment and adoption.

Injury Over Sacriﬁce (IOS). The only crash sce-

nario which did not receive the highest vote for the

ODS, was question 3 where the majority of the par-

ticipants opted for the AV to swerve to the left, hit-

ting and severely injuring the motorcyclist with a hel-

met. This choice reﬂects a prioritisation of minimis-

ing fatalities, with severe injury seen as a more ac-

ceptable outcome than death. The IOS rule can be

programmed into an AV from the perspectives of util-

itarianism, deontology, and Rawlsian ethics. Utilitar-

ianism supports IOS as it minimises overall harm by

opting for injury over the loss of life. Deontological

ethics aligns with IOS by prioritising the duty not to

kill, thus making the rule morally acceptable. Rawl-

sian ethics, focusing on fairness and justice, also sup-

ports IOS as it respects the fundamental right to life

and aims to minimise the worst possible outcomes.

Therefore, IOS can be justiﬁed and integrated into AV

programming under these ethical frameworks, each

providing a distinct rationale for its implementation.

Harm Conﬁnement and Lawfulness (HCL). The

participants’ choice in scenario 1 and 2 when ODS

was removed, reﬂects a prioritisation of moral and le-

gal principle of avoiding harm to innocent third par-

ties and transferring harm to the responsible party

that has caused the emergency situation. If harm is

not limited to this extent, anyone could potentially be

struck by an AV, requiring everyone to be constantly

vigilant of this danger. Therefore, consideration for

the legal implication of a decision made by the AV

and to the extent that other road users have acted law-

fully takes precedence over saving the greater number

of people. The survey results indicate that Australians

consider rule-breakers (such as jaywalkers) as more

ethically liable to suffer harms, and they preferred

measures that prevented death to innocents. This is

indeed the very fact that appeared to be missing from

previous studies, including the MM experiment and

the infamous Trolley problem (Awad et al., 2018).

HCL makes harm distribution dynamic, as it will

depend on scenario speciﬁc variables. The rule priori-

tising legal considerations and the lawful actions of

road users over saving the greater number of people

aligns best with deontological ethics. Deontological

ethics focus on adherence to rules and duties, mak-

ing it suitable for programming an AV to respect le-

gal responsibilities even at the expense of maximis-

ing overall lives saved. Legal responsibility could be

integrated into a decision tree, where the default ac-

tions taken by the AV change based on the legal im-

plications of each scenario. In contrast, utilitarian-

ism, which emphasises maximising overall happiness

and minimising harm, would generally favour saving

the greater number of people regardless of their legal

status, making it incompatible with this rule. Simi-

larly, Rawlsian ethics, which focuses on fairness and

the protection of the least advantaged, does not inher-

ently prioritise legal considerations, thus also making

it a less suitable framework for this speciﬁc rule.

Harm Prevention and Prioritisation (HPP). Partic-

ipants’ preferences in scenario 4, when the ODS was

removed, highlighted a strong consensus to prioritise

preventing harm rather than intentionally causing an-

other person’s death, especially considering that all

parties involved were part of the original scenario and

at fault. Killing implies that you are directly respon-

Ethics of Autonomous Vehicles: Australians’ Expectations and Moral Preferences

293

sible for the death of another person and something

that, for liability reasons, remains critically important

to AV manufacturers. Allowing someone to die on the

other hand, entails much less responsibility on your

part, as there were some contributory factors already

in motion that you did not initiate or otherwise con-

trol.

Given that both the pedestrian and the motorcy-

clist were at fault, a utilitarian AV would calculate

the total expected harm from each option. If the AV

determines that braking to avoid the pedestrian re-

sults in less overall harm (e.g., the motorcyclist may

have a higher chance of survival even in a collision),

it would choose this option. The AV could be pro-

grammed to use advanced sensors and algorithms to

estimate the severity of injuries or likelihood of death,

making an informed decision that minimises total

harm. The deontological AV would be programmed

to brake, as this action avoids the intentional harm

to the pedestrian, despite the pedestrian’s negligence.

The AV would prioritise adhering to moral rules over

the consequences, focusing on the duty to avoid caus-

ing harm. Similarly, a Rawlsian AV would be pro-

grammed to brake, prioritising the pedestrian who is

directly in the path and at immediate risk. The AV

would aim to distribute the risks more equitably, en-

suring that harm is not disproportionately inﬂicted on

the more vulnerable party, in this case, the pedestrian.

7 PROPOSED ETHICAL POLICY

An ethical policy consists of broad, abstract values

that guide the development of socially acceptable

AVs, ensuring they align with societal norms and ex-

pectations. The speciﬁc decision-making models and

normative ethics can be too complex for the aver-

age user to understand, so an ethical policy is nec-

essary to explain how AVs will make decisions (Liu

et al., 2021). Based on my research ﬁndings, the pro-

posed ethical policy includes the following four re-

quirements, which are grounded in Australian moral

preferences:

Injury Over Sacriﬁce (IOS). In any scenario where

the AV can choose between causing injury or death,

the AV will prioritise actions that minimise fatalities.

Severe injuries are considered more acceptable than

deaths.

Harm Conﬁnement and Lawfulness (HCL). When

all available options lead to death, the AV will aim to

transfer harm to those who are responsible for creat-

ing the hazardous situation. The AV conﬁnes harm

to those directly involved and ensures innocent by-

standers are not harmed. The AV’s actions will be

guided by legal implications and the extent to which

other road users have acted lawfully. Legal principles

and moral responsibility take precedence over simply

saving the greatest number of people.

Harm Prevention and Prioritisation (HPP). In sit-

uations where multiple parties are equally at fault and

have all contributed to the dangerous situation, the AV

will avoid taking direct actions that would intention-

ally cause any individual’s death, opting instead to re-

main passive. This avoids intentional killing, reﬂect-

ing legal and ethical priorities. If multiple principles

apply to a moral dilemma, the AV’s algorithm should

evaluate the scenario based on all applicable princi-

ples. The above principles work together and are not

standalone. The AV should dynamically balance IOS,

HCL, and HPP, ensuring a holistic and ethically sound

approach to decision-making.

Objective Decision System (ODS). When faced with

scenarios where decision-making is hindered by tech-

nological limitations, time constraints, or other fac-

tors and the IOS, HCL or HPP principles are rendered

ineffective, the AV must employ the ODS. This sys-

tem will evaluate all viable and justiﬁed outcomes and

select one at random, ensuring fairness, impartiality,

and non-discrimination in the decision-making pro-

cess.

7.1 Revisit Crash Scenarios

In this section I revisit the crash scenarios from the

survey questionnaire and discuss how an AV pro-

grammed with the proposed ethical settings would

handle each situation.

Scenario 1 - Brakes Failure on Mountain Road

Moral Dilemma: The AV must choose between

swerving right to hit an innocent bystander or con-

tinuing straight off a cliff, resulting in the death of the

passenger.

Principle Application:

• IOS: Not applicable as both outcomes lead to

death.

• HCL: The bystander is not culpable for the emer-

gency situation or the AV’s mechanical failure.

The AV should prioritise conﬁning harm and not

introducing new parties to the risk.

• HPP: The AV should avoid actively causing the

death of an innocent bystander.

• ODS: If no clear decision can be made, the AV

would randomly choose from justiﬁed options.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

294

AV Response: The AV would continue straight off

the cliff, resulting in the death of the passenger. This

action aligns with HCL and HPP by avoiding harm to

an innocent bystander who did not contribute to the

dangerous situation.

Scenario 2 - Intersection with Illegally Crossing

Pedestrians

Moral Dilemma: The AV can either swerve left to

save ﬁve illegally crossing pedestrians, causing the

death of a cyclist, or continue straight, potentially hit-

ting one or more pedestrians.

Principle Application:

• IOS: Not applicable as all possible outcomes lead

to death.

• HCL: The illegally crossing pedestrians are at

fault. The AV should prioritise conﬁning harm

to the responsible parties and transferring harm to

those who contributed to the dangerous situation.

• HPP: The AV should avoid actively killing the cy-

clist who is not at fault.

• ODS: If no clear decision can be made, the AV

would randomly choose from justiﬁed options.

AV Response: The AV would continue straight,

potentially hitting one or more of the pedestrians.

This action aligns with HCL by conﬁning harm to

those who are responsible for the situation (the ille-

gally crossing pedestrians) and avoids actively killing

the cyclist, who is not at fault.

Scenario 3 - Highway Dilemma with Fallen Pallet

Moral Dilemma: The AV must choose between con-

tinuing straight and killing the passenger, swerving

right and killing a helmetless motorcyclist, or swerv-

ing left and causing severe injury to a fully geared

motorcyclist.

Principle Application:

• IOS: Applicable; the AV should avoid causing

death by choosing an outcome that minimises fa-

talities.

• HCL: The helmetless motorcyclist is partially at

fault for not wearing safety gear.

• HPP: The AV should avoid actively killing and

prefer causing severe injury over death.

• ODS: If no clear decision can be made, the AV

would randomly choose from justiﬁed options.

AV Response: The AV would swerve left, causing

severe injury but survival of the fully geared motorcy-

clist. This action aligns with IOS by preventing death

and with HPP by avoiding active killing. The helmet-

less motorcyclist, while partially at fault, is spared to

prevent a fatality.

Scenario 4 - Pedestrian vs. Motorcyclist

Moral Dilemma: The AV must choose between brak-

ing to save a pedestrian, causing a motorcyclist to

crash into the AV and die, or continuing and striking

the pedestrian, saving the motorcyclist.

Principle Application:

• IOS: Not applicable as both outcomes lead to

death.

• HCL: Both the pedestrian and the motorcyclist are

considered negligent.

• HPP: The AV should avoid actively causing the

death of either party.

• ODS: If no clear decision can be made, the AV

would randomly choose from justiﬁed options.

AV Response: The AV would brake to save the

pedestrian, causing the motorcyclist to collide with

the AV and be killed. This action aligns with HPP

by avoiding the active killing of the pedestrian, who

is directly in the AV’s path.

Scenario 5 - Intersection with Legally Crossing

Pedestrians

Moral Dilemma: The AV must decide between swerv-

ing into a wall, risking the lives of ﬁve passengers, or

proceeding straight, potentially endangering legally

crossing pedestrians.

Principle Application:

• IOS: Not applicable as both outcomes lead to

death.

• HCL: The pedestrians are legally crossing, while

the AV’s passengers are not at fault.

• HPP: The AV should avoid actively causing the

death of the pedestrians who are acting lawfully.

• ODS: If no clear decision can be made, the AV

would randomly choose from justiﬁed options.

AV Response: The AV would swerve into the

wall, risking the lives of its ﬁve passengers to save

the pedestrians. This action aligns with HCL by pro-

tecting the legally crossing pedestrians and with HPP

by avoiding active killing.

Scenario 6 - Intersection with Legal Pedestrian

Crossing

Moral Dilemma: The AV must choose between

swerving into a tree, killing three passengers, or con-

tinuing straight and hitting a pedestrian legally cross-

ing.

Principle Application:

• IOS: Not applicable as both outcomes lead to

death.

Ethics of Autonomous Vehicles: Australians’ Expectations and Moral Preferences

295

• HCL: The pedestrian is legally crossing, while the

passengers are not at fault.

• HPP: The AV should avoid actively causing the

death of the legally crossing pedestrian.

• ODS: If no clear decision can be made, the AV

would randomly choose from justiﬁed options.

AV Response: The AV would swerve to the right,

colliding with a tree and resulting in the deaths of

three passengers. This action aligns with HCL by pro-

tecting the legally crossing pedestrian and with HPP

by avoiding active killing.

8 CONCLUSION

The research set out to explore how a publicly ac-

cepted ethical setting for an AV could look like. It

presented different ethical theories, decision-making

models and moral variables that could help solve

moral dilemmas faced by AVs. The objective was

to evaluate respondents’ perceptions and expectations

regarding personal AVs in relation to different ethical

settings. The survey questions were designed to elim-

inate any biased or redundant criteria based on pre-

vious research, ensuring that participants’ decisions

were informed by both ethical and legal considera-

tions through the inclusion of legal implications for

each crash scenario.

A signiﬁcant aspect of the research was the in-

troduction of a randomised choice option, referred

to as the ODS. In scenarios where the AV encoun-

ters a moral dilemma, such as the decision between

braking to avoid hitting a pedestrian or not brak-

ing, resulting in harm to the motorcyclist behind, the

ODS allows the AV to evaluate all viable outcomes

based on a moral framework and then select one at

random. The ﬁndings revealed that both utilitarian-

ism and Rawlsian ethics were unsuitable for aligning

with the participants’ preferences, as the core prin-

ciples of these ethical theories do not adequately ac-

count for the different roles of participants in harm

distribution. Consequently, the proposed ethical set-

ting is grounded in deontology, utilising a rule-based

decision-making model with moral variables, includ-

ing IOS, HCL, HPP, and ODS as redundancy. This

shift represents a departure from the utilitarian harm

minimisation approach observed in other studies.

Research questionnaires, like the one I conducted,

are valuable for simplifying complex real-world is-

sues, focusing on core ethical considerations (Poj-

man, 1990). They can help test initial programming

ideas for AVs, as algorithms that contradict societal

moral expectations are likely to hinder the smooth

adoption of AVs. Furthermore, ethical decision-

making cannot rest solely with engineers or ethicists;

a consensus is essential, as solutions to programming

moral dilemmas in AVs should not deter potential

buyers. Engaging the public can spark curiosity about

the various ways AVs could be programmed to ad-

dress ethical dilemmas (Nyholm, 2018).

It has been observed that respondents’ decisions

shift with the level of concreteness in experiments

(Francis et al., 2017). For future research, I recom-

mend using virtual reality to create realistic crash sce-

narios, allowing participants to engage with differ-

ent dilemmas and choose outcomes they deem most

acceptable. Additionally, I have outlined the advan-

tages of the ODS in AV programming. Despite initial

concerns about using lotteries for determining legal

outcomes, ODS should not be dismissed in AV crash

algorithms. Decision-making by lot is often simple,

objective, and inexpensive, and my research suggests

that it is both understandable and acceptable to many

individuals. Thus, I believe that lotteries, more than

other decision-making devices, merit serious consid-

eration by future researchers, AV manufacturers, and

policymakers.

Developing ethical autonomous machines is one

of the most challenging topics in artiﬁcial intelligence

today. It extends beyond resolving ethical dilemmas;

setting realistic expectations with users and the public

is crucial for market acceptance and adoption. These

challenges are common across automotive manufac-

turers engaged in this emerging ﬁeld, not limited to

speciﬁc companies. AVs promise substantial bene-

ﬁts while also presenting unintended effects that are

difﬁcult to predict, and the technology is progress-

ing rapidly, regardless of ethical considerations (Lin,

2016). This juncture represents an ideal opportunity

to discuss the ethical values and principles best suited

for self-driving cars, laying the groundwork for future

improvements.

REFERENCES

Australian Bureau of Statistics (2021). Australia migra-

tion 2019-20 ﬁnancial year. https://www.abs.gov.au/

statistics/people/population. Accessed: 2021-06-20.

Awad, E., Dsouza, S., Kim, R., Schulz, J., Henrich, J., Shar-

iff, A., Bonnefon, J.-F., and Rahwan, I. (2018). The

moral machine experiment. Nature, 563(7729):59–64.

Bonnefon, J.-F., Shariff, A., and Rahwan, I. (2015). Au-

tonomous vehicles need experimental ethics: are

we ready for utilitarian cars. arXiv preprint

arXiv:1510.03346.

Bonnefon, J.-F., Shariff, A., and Rahwan, I. (2016). The

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

296

social dilemma of autonomous vehicles. Science,

352(6293):1573–1576.

Botes, A. (1996). Understanding social research. Health

SA Gesondheid : Journal of Interdisciplinary Health

Sciences, 1.

Campbell, M., Egerstedt, M., How, J. P., and Murray, R. M.

(2010). Autonomous driving in urban environments:

approaches, lessons and challenges. Philosophical

Transactions of the Royal Society A: Mathematical,

Physical and Engineering Sciences, 368(1928):4649–

4672.

Coca-Vila, I. (2018). Self-driving cars in dilemmatic situ-

ations: An approach based on the theory of justiﬁca-

tion in criminal law. Criminal Law and Philosophy,

12(1):59–82.

Curtis, P. and Hedlund, J. (2005). Reducing deer-vehicle

crashes. report funded by the insurance institute for

highway safety. cornell university, ithaca.

Duxbury, N. (1999). Random Justice: On Lotteries and

Legal Decision-Making. Oxford University Press UK.

Dworkin, R. (2011). Justice for hedgehogs. Harvard Uni-

versity Press.

Etienne, H. (2020). When ai ethics goes astray: A case

study of autonomous vehicles. Social Science Com-

puter Review, page 0894439320906508.

Farhadi, A., Endres, I., Hoiem, D., and Forsyth, D. (2009).

Describing objects by their attributes. In 2009 IEEE

conference on computer vision and pattern recogni-

tion, pages 1778–1785. IEEE.

Francis, K. B., Terbeck, S., Briazu, R. A., Haines, A., Gum-

merum, M., Ganis, G., and Howard, I. S. (2017). Sim-

ulating moral actions: An investigation of personal

force in virtual moral dilemmas. Scientiﬁc Reports,

7(1):1–11.

Gomez, L. R. P., Szybalski, A. T., Thrun, S., Nemec, P., and

Urmson, C. P. (2014). Transportation-aware physical

advertising conversions. US Patent 8,630,897.

IEEE (2016). Ethically aligned design: A vision for pri-

oritizing wellbeing with artiﬁcial intelligence and au-

tonomous systems.

Lin, P. (2013). The ethics of autonomous cars. the atlantic,

october 8.

Lin, P. (2016). Why ethics matters for autonomous cars. In

Autonomous driving, pages 69–85. Springer, Berlin,

Heidelberg.

Liu, Y., Sun, L., Wang, Y., and Ren, Y. (2021). Vision-based

obstacle detection and recognition for autonomous ve-

hicles. Measurement, 184:109984.

Luetge, C. (2017). The german ethics code for automated

and connected driving. Philosophy & Technology,

30(4):547–558.

Noothigattu, R., Gaikwad, S., Awad, E., Dsouza, S., Rah-

wan, I., Ravikumar, P., and Procaccia, A. (2018). A

voting-based system for ethical decision making. In

Proceedings of the AAAI Conference on Artiﬁcial In-

telligence, volume 32.

Nyholm, S. (2018). The ethics of crashes with self-

driving cars: A roadmap, i. Philosophy Compass,

13(7):e12507.

Pojman, L. (1990). Discovering right and wrong. Belmont,

CA: Wadsworth.

Rachels, J. and Rachels, S. (2012). The Elements of Moral

Philosophy 7e. McGraw Hill.

Raﬁee, A., Wu, Y., and Sattar, A. (2023). Philosophical and

legal approach to moral settings in autonomous vehi-

cles: An evaluation. In Social Licence and Ethical

Practice, volume 27, pages 95–114. Emerald Publish-

ing Limited.

Rescher, N. (1960). Choice without preference. a study of

the history and of the logic of the problem of "buri-

dan’s ass". Kant-Studien, 51(1-4):142–175.

Shadrin, S. S. and Ivanova, A. A. (2019). Analytical review

of standard sae j3016 «taxonomy and deﬁnitions for

terms related to driving automation systems for on-

road motor vehicles» with latest updates. Avtomobil’.

Doroga. Infrastruktura., (3 (21)):10.

Shahriari, K. and Shahriari, M. (2017). Ieee standard re-

view—ethically aligned design: A vision for priori-

tizing human wellbeing with artiﬁcial intelligence and

autonomous systems. In 2017 IEEE Canada Interna-

tional Humanitarian Technology Conference (IHTC),

pages 197–201. IEEE.

Zimmermann, T. (2014). Justiﬁcation or excuse? the col-

lision of positive duties in german criminal law. The

Journal of Criminal Law, 78(3):263–276.

Ethics of Autonomous Vehicles: Australians’ Expectations and Moral Preferences

297