The OYI Review · One Young India Press

The Ethical Framework for Autonomous Driving

By Bhavya Sharma, Jayshree Periwal High School

Published 2025 · Reviewed and updated 2026 by One Young India Review

Abstract

The rapid development of autonomous vehicle (AV) technology introduces ethical questions that go well beyond conventional automotive safety. As artificial intelligence (AI) systems take on greater responsibility for life-or-death decisions on public roads, society must confront fundamental questions about moral programming, liability, and the translation of human ethical principles into algorithmic logic. This analysis synthesises current research, legal frameworks, and technological developments to examine the ethical implications of AI in autonomous vehicles and to set out a practical direction for building morally defensible autonomous transport. The central argument is that the well-known trolley problem, while useful for provoking public debate, is the wrong frame for real engineering decisions. Vehicles do not select between predetermined victims. They make sequential decisions under uncertainty. A stronger framework therefore rests on dynamic risk minimisation, a legally grounded duty of care to all road users, transparency, and honesty about the technical limits of AI. Recent deployment data, updated regulation, and a growing body of empirical ethics research all point in the same direction: ethically sound autonomous driving is technically feasible, and it is a precondition for public trust rather than an optional extra.

Introduction to Autonomous Vehicles

The arrival of autonomous vehicles represents a genuine shift in transport technology, changing the relationship between human judgement and machine decision-making in safety-critical moments. In a conventional car, a human driver makes split-second choices from instinct and moral intuition. An autonomous vehicle must instead be programmed in advance with explicit rules that govern its behaviour in complex traffic. This move from human to algorithmic moral agency is the source of most of the difficulty, because it forces designers to state, in code, principles that human drivers apply without ever articulating them.

To describe this progression, the Society of Automotive Engineers (SAE) defines a six-level classification, from Level 0 (no automation) to Level 5 (full automation). Each level marks an increase in machine autonomy and a corresponding shift in where ethical responsibility sits. Most systems sold to consumers still operate at Level 2, where the human must supervise continuously. That picture is now changing at the top of the scale. In 2023 Mercedes-Benz became the first manufacturer to obtain certification for a Level 3 system, DRIVE PILOT, for use on public roads in the United States, first in Nevada and then in California, where the vehicle may take over the driving task at limited speed on qualifying motorways while the driver attends to other things. At the same time, driverless Level 4 services are carrying paying passengers. This mix of Level 2, Level 3, and Level 4 on the same roads is exactly what makes liability attribution and ethical programming so demanding, because human and machine responsibility now coexist within a single traffic system.

The ethical consequences of AVs reach beyond individual moral judgements into questions of social justice, resource allocation, and technological governance. Research consistently finds that public acceptance depends heavily on whether the ethical programming is seen as fair and transparent. The MIT Moral Machine experiment, which gathered close to 40 million moral decisions in ten languages from participants across 233 countries and territories, sharply challenged the idea that a single set of universal moral rules could be written into every car. Its results revealed systematic cross-cultural differences in ethical preference, which complicates any search for a one-size-fits-all solution.

Theoretical Foundations of AI Ethics in Autonomous Systems

The Limits of the Trolley Problem

For years, debate about AV ethics has been dominated by the trolley problem, a philosophical thought experiment that forces a binary choice: harm one person, or allow harm to reach many. A growing body of work argues that this framing is fundamentally unsuited to real autonomous driving. The trolley problem assumes a level of certainty and predictive power that no current or foreseeable AV possesses.

Writing for the Brookings Institution, Heather M. Roff argues that autonomous vehicles work through sequential decision-making under uncertainty rather than one-shot choices between known outcomes. Such vehicles reason with models like Partially Observable Markov Decision Processes (POMDPs), in which the system has incomplete knowledge and acts on probability distributions, not explicit moral calculations over fully specified victims. As Roff puts it, an AV is "working in an environment where the system has incomplete knowledge and uncertainty, and is working from probabilities". This mismatch between the assumptions of the thought experiment and the operating reality of the machine is the core reason the trolley problem misleads more than it clarifies. Ethical frameworks for AVs must therefore be grounded in what the technology can actually do.

Dynamic Risk Assessment and Ethical Decision-Making

Contemporary research treats dynamic risk assessment as the cornerstone of ethical AV behaviour. Rather than programming a vehicle to choose in advance between different groups of people, emerging frameworks focus on minimising overall risk while respecting legal and ethical constraints. This reflects the fact that an AV must decide continuously, on incomplete information, so its ethics must function under uncertainty rather than assume it away.

Empirical work supports this shift. In a representative study published in Scientific Reports in 2024, Sebastian Krügel and Matthias Uhl asked participants how an AV should distribute risk in ordinary traffic, using interactive graphical scenarios rather than stark life-or-death dilemmas. Preferences deviated significantly from simple collision avoidance, and participants were often willing to accept risk themselves for the benefit of other road users. The authors argue that the ethical question is not confined to rare emergencies but is present in the permanent redistribution of risk among road users during everyday driving. That finding reframes the whole problem: ethics is embedded in normal driving behaviour, not reserved for a hypothetical crisis.

The Stanford team led by Chris Gerdes offers a complementary framework that grounds AV ethics in existing traffic law and the social contract between road users. On this view, an AV should uphold a duty of care to everyone on the road and should do everything within the physical limits of the vehicle to avoid a collision. In Gerdes's formulation, "the only basis for an AV to break the law should be that it is necessary to avoid a collision", and even then the vehicle may not act in ways that violate the duty of care owed to others. When a conflict is unavoidable, the system should address the party who created it, meaning the party who first breached their own duty of care, "without bringing other people into it". This turns an abstract moral puzzle into a set of enforceable, law-anchored obligations that engineers can implement and regulators can audit.

Cultural and Contextual Considerations

The Moral Machine experiment undercut the assumption of universal moral principles by exposing clear cultural differences in ethical preference. Western respondents leaned towards utilitarian choices that save the greatest number, whereas respondents in several Eastern cultures placed comparatively greater weight on respect for the elderly and for authority. These patterns suggest that ethical programming may need to be sensitive to local context rather than fixed to a single global standard.

The same research identified three broad clusters of countries whose moral preferences correlate with the strength of national institutions and with deeper cultural traits. Nations with stronger institutions tended to favour rule-based approaches, while those with weaker institutions leaned towards consequentialist reasoning. For global manufacturers and regulators seeking uniform standards across very different markets, these discrepancies are not a detail. They are a central design constraint, because a rule that reads as fair in one jurisdiction may read as unjust in another.

Technical Implementation of Ethical AI Systems

Machine Learning and Moral Decision-Making

Putting ethical decision-making into an AV requires machine learning systems that can process large volumes of sensor data and act quickly under uncertainty. Current approaches combine supervised learning, deep learning, and reinforcement learning so that a vehicle can recognise traffic patterns, predict the behaviour of other road users, and choose an appropriate action.

Convolutional Neural Networks (CNNs) remain the primary technology for environmental perception, allowing an AV to detect and classify pedestrians, cyclists, and other vehicles. The ethical performance of these models depends directly on the data used to train them. If early systems were more likely to fail to recognise people with darker skin, then equitable perception requires diverse and representative training data. Data quality is not merely a technical concern here, because a gap in the training set becomes a gap in the vehicle's duty of care.

Deep learning architectures, particularly attention-based networks, let a vehicle concentrate on the most relevant parts of its environment. Their strength is also their weakness. The opacity of many deep models, often called the black box problem, raises real concerns about transparency and accountability when the question is not only what the vehicle did but why it was entitled to do it.

A Real-Time Ethical Processing Framework

Research has proposed a five-step ethical decision model that can be implemented in real time. In sequence, the vehicle performs a comprehensive risk assessment to identify potential hazards, then legal compliance verification to ensure its options align with traffic law, then a duty of care analysis of its responsibilities to all road users, then identification of the minimal risk condition, meaning the least harmful available path, and finally action selection to execute the chosen manoeuvre.

Each step carries specific ethical principles and technical measures that can be quantified and coded into the control system. Doing this within the milliseconds required for safe operation demands significant computation and capable edge processing, so that decisions are reached at the speed of driving rather than the speed of deliberation. The ordering matters: risk and law are evaluated before the manoeuvre is chosen, which encodes the duty of care as a constraint rather than an afterthought.

Transparency and Explainability

As AVs take on more responsibility, the demand for transparent and explainable AI (XAI) grows. Explainability here is not only a matter of technical performance. It extends to ethical reasoning, requiring systems that can articulate why a particular decision was made in terms of moral principles and legal obligations. A vehicle that cannot explain itself cannot be meaningfully held to account.

Current research explores hybrid systems that combine data-driven machine learning with explicit knowledge structures encoding traffic law and ethical principles. The aim is to keep the performance of learned models while recovering the interpretability of rule-based ones, so that engineers, regulators, and the public can inspect and validate the reasoning behind an AV's behaviour.

Legal and Regulatory Frameworks

Evolving Liability

Autonomous vehicles force a reconsideration of liability rules that have historically placed primary responsibility on the human driver. As a vehicle becomes more autonomous, responsibility shifts towards manufacturers, software developers, and others in the design and deployment chain. Legal approaches to this shift differ sharply by jurisdiction. The European Union has moved towards harmonised frameworks that distinguish between levels of automation and assign responsibility accordingly, while the United States has taken a more fragmented, state-by-state path, which is one reason a single manufacturer can face very different rules across neighbouring states.

Product Liability and Software Defects

As AVs become software-defined products, product liability law is being extended to cover software defects and algorithmic failure. This is difficult for manufacturers, because the complexity of modern AI makes it close to impossible to anticipate every failure mode across the near-infinite range of real traffic. The response emerging in policy debate is procedural rather than perfectionist: mandatory testing standards, certification processes, and ongoing algorithmic auditing, so that safety is demonstrated continuously rather than assumed once at launch.

International Standardisation

Global bodies are working towards consistent frameworks. IEEE 2846-2022, the Standard for Assumptions in Safety-Related Models for Automated Driving Systems, defines a minimum set of reasonable assumptions and foreseeable scenarios that a safety-related model should consider, and does so in a technology-neutral way. In parallel, under the United Nations Economic Commission for Europe (UNECE), UN Regulation No. 157 on Automated Lane Keeping Systems became, on its entry into force on 22 January 2021, the first binding international regulation for Level 3 automation. Its later amendments raised the permitted operating speed and added provision for automated lane changes. Together these instruments show that the vulnerable road user and the ethics of automated decision-making are moving from academic discussion into enforceable type approval.

Case Studies and Real-World Applications

Waymo and the Case for Measured Safety

Waymo, which began as Google's self-driving car project, offers the clearest evidence that a conservative, all-road-users approach can be measured rather than merely asserted. Early patents suggested a focus on protecting vehicle occupants, but the company's operating philosophy has shifted towards defensive driving and caution that prioritises the safety of everyone on the road. According to Waymo's published safety data, and corroborated by peer-reviewed analysis in the journal Traffic Injury Prevention in 2025, across more than 200 million fully driverless miles the Waymo Driver was involved in far fewer injury crashes than human drivers over comparable driving: on the order of 94 per cent fewer crashes causing serious or fatal injury, around 93 per cent fewer pedestrian injury crashes, and around 84 per cent fewer cyclist injury crashes, with a serious-injury rate close to 0.01 per million miles against a human benchmark near 0.23. These are company-reported figures, and independent scrutiny remains essential, but they matter for the ethical argument. If a duty-of-care framework can be shown to reduce the harm that falls on the most vulnerable road users, then the ethical case and the safety case converge.

Tesla, Autopilot, and the Supervision Problem

Tesla represents a contrasting approach. Its Autopilot and Full Self-Driving features remain SAE Level 2 systems that keep the human driver responsible for safety-critical decisions. This sidesteps some direct programming dilemmas but introduces others, principally around how system limitations are communicated and how driver overreliance is prevented. The tension became more visible in 2025. Tesla launched a limited robotaxi pilot in Austin in June 2025, initially with a safety monitor on board, while its consumer systems stayed at Level 2, and later in 2025 the United States National Highway Traffic Safety Administration opened an investigation covering roughly 2.9 million vehicles over the behaviour of Full Self-Driving. The episode underlines a point the ethical literature makes repeatedly: naming and marketing shape driver expectations, and a manufacturer's duty of care includes managing those expectations honestly.

European Union Regulatory Pilots

The European Union has run pilot programmes to test different regulatory approaches. These commonly require transparent algorithms, regular safety audits, and reporting of ethically significant decision scenarios. They demonstrate that a comprehensive regulatory framework can balance innovation against public safety rather than trading one off against the other.

Challenges and Future Directions

Key Challenges

Computational constraints. Making an ethically complex decision within the milliseconds available for safe operation remains a hard technical limit, and it forces a trade-off between ethical sophistication and real-time performance.

Cultural adaptation. The cultural variation in moral preference documented by the Moral Machine experiment makes a globally consistent yet locally acceptable product genuinely difficult to design. Balancing cultural sensitivity against universal ethical principles is unresolved.

Verification and validation. The near-infinite space of possible traffic situations makes exhaustive testing impossible, so assuring ethical behaviour in unforeseen edge cases is a first-order safety problem, not a corner case of one.

Emerging Solutions

Hybrid human and AI decision-making. Researchers are examining systems in which remote human operators can intervene in complex situations, or in which the AI escalates unusually difficult decisions to human supervisors.

Adaptive learning. Future AVs may use techniques such as Reinforcement Learning from Human Feedback to refine their behaviour against real-world experience and evolving social norms, provided that such adaptation is itself auditable.

Distributed governance. Some proposals suggest distributed and cryptographically transparent systems for setting ethical standards, recording decisions, and supporting international cooperation, though these remain early and unproven.

Recommendations and Policy Implications

For Regulators

Base frameworks on realistic technological capabilities, focusing on overall risk minimisation and adherence to legal duties of care rather than on hypothetical dilemmas.
Mandate transparency and explainability through requirements for XAI systems, regular algorithmic audits, and public reporting, so that trust is earned with evidence.
Promote international coordination through bodies such as IEEE and UNECE to prevent a patchwork of incompatible rules and to secure consistent global safety standards.

For Industry

Adopt comprehensive ethical frameworks that incorporate diverse stakeholder perspectives, including ethicists, legal experts, and advocates for vulnerable road users.
Invest in robust testing and validation that assesses ethical performance across diverse scenarios and cultural contexts, using both extensive simulation and real-world pilots, and publish the results.

For Research

Prioritise practical ethical frameworks that can run within real-time computational constraints.
Develop better ways to measure and validate ethical performance, including standardised test scenarios and metrics for fairness and transparency.

Conclusion

The ethical programming of autonomous vehicles is one of the most consequential challenges facing modern technology. Popular discussion often reduces it to a stylised choice like the trolley problem, but the reality is more demanding and, in an important sense, more tractable. Real progress comes from frameworks that address dynamic risk assessment, cultural variation, legal liability, and the technical limits of AI, and that treat ethics as a constraint on ordinary driving rather than a rare emergency subroutine.

As accountability shifts from human drivers to manufacturers and algorithms, society must rethink regulation and oversight, and the recent record shows this is already under way, from the first certified Level 3 systems to binding international regulation and published safety data. The evidence of cross-cultural difference in ethical standards underlines the need for adaptable systems that balance diverse values against shared principles of safety and justice. Meeting that need will require sustained collaboration between technologists, policymakers, ethicists, and the public.

Designing ethical autonomous vehicles is not only technically feasible. It is a precondition for their widespread adoption and for public trust. The task is complex, but the potential gains, in the form of fewer collisions, better accessibility, and more efficient transport, make it an essential undertaking. The choices made today in programming ethical AI will shape not only the future of mobility but the wider relationship between people and intelligent machines.

Sources

Cite this paper

Bhavya Sharma, Jayshree Periwal High School (2025). The Ethical Framework for Autonomous Driving. The OYI Review, One Young India Press. https://www.oneyoungindia.com/white-papers/the-ethical-framework-for-autonomous-driving

← All student work ⤓ Download PDF Submit your paper →