The Ethics of Autonomous Vehicles: Revisiting a 2020 Paper (Part 3: Ethical Frameworks and Conclusion)

Millar’s Updated Framework

Millar extended his 2016 five-rule framework in a 2017 book chapter, “Ethics Settings for Autonomous Vehicles,” in Robot Ethics 2.0 (Oxford University Press) — focusing specifically on how designers could embed ethical values into AV systems, along with the resulting tradeoff between paternalism and user autonomy (Millar, 2017). His central concept, “ethics settings,” proposes that users could adjust the ethical parameters governing the vehicle before a trip, effectively selecting their own answer to the trolley problem in advance. You did read the Terms and Conditions! Lin (2014) thinks this is a BAD IDEA, not removing liability for the company, putting responsibility on the driver, having the algorithms “target” those more likely to survive, and possibly even considering a driver’s actions “premeditated”, offering significant legal consequences (Lin, 2014). Millar (2017) even quotes Lin in his chapter, but thinks that regulation can fix these issues (Millar, 2017). I looked for more recent opinions from Lin (he edited the Robot Ethics 2.0 book) but I couldn’t find if he lessened his stance at all.

Lin, P. (2014, August 18). Here’s a terrible idea: Robot cars with adjustable ethics settings. Wired. https://www.wired.com/2014/08/heres-a-terrible-idea-robot-cars-with-adjustable-ethics-settings/

Millar, J. (2017). Ethics settings for autonomous vehicles. In P. Lin et al. (Eds.), Robot Ethics 2.0 (pp. 20-34). Oxford University Press.

New Research

Pődör & Lakatos (2026) give two AV ethics strategies - dilemma avoidance, and ethical programming. Dilemma avoidance is conservative driving, defaulting to emergency braking, and not picking victims (Pődör & Lakatos, 2026). Basically, this is what Mobileye and Waymo already do. Remember from Part 2, Mobileye’s RSS Rule 5: if the car can avoid a crash without causing another one, it must - and the rules stop right there (Mobileye, n.d.). Ethical programming is encoding moral attributes into the system (Pődör & Lakatos, 2026). This is what Gros et al. are designing (2025b). Both of these fail in isolation, and realistically it becomes a hybrid that nobody really specified. Pődör also states “the absence of explicit rules also amounts to a kind of decision” (Pődör & Lakatos, 2026).

The other problem is that LiDAR doesn’t work as well in rain, fog, snow, swarms of locusts, etc. Pődör points out that the vehicle’s sensors have latency (Pődör & Lakatos, 2026). AVs are almost certainly safer than humans - Waymo’s fatal crash reconstruction data shows only 8% of simulated scenarios as “unchanged outcome,” meaning the car was already stopped and got rear-ended from behind - situations no driver could have prevented. The other 92% showed the Waymo Driver would have improved or avoided the outcome entirely (Waymo, 2021). The more interesting question is what happens with these incremental sensor failures? Just like if a headlight goes out or your wipers aren’t cleaning, does the liability for a sensor failure shift back to the driver? NHTSA’s Tesla investigation kept Tesla at Level 2 precisely to keep driver liability in place (NHTSA, 2022), but that’s a legal move, not an answer to the hardware degradation question.

Gros et al. (2025b) has an entirely different idea, an Augmented Utilitarian (AU) framework with foundations in virtue ethics, deontology, and consequentialism, plus a biomedical ethics applied layer that includes beneficence, non-maleficence, justice, and autonomy-as-institutional-oversight. Note that Gros (2025b) states that “Autonomy” means Meaningful Human Control at the governance level, not you specifically making the choice. Geisslinger (cited by Gros) moves the trolley problem into more practical risk management, like probability distributions, sensor confidence, and reaction time budgets. Gros gives examples of how this framework can.. work.. with doctors (peer reviews, malpractice liability, and licensing boards), and states that AVs can have something similar with a crash report after the fact. There’s also a table, Table 3, with weighting attributes like profession, age, gender, and family status (Gros et al., 2025b) - we’ll get to why that’s a problem in a minute - and Gros defers how to use them to legislators.

This biomedical analogy seems to work as a starting point, but it quickly breaks down. Patients can input preferences with things like a living will or a donor card. You can opt to have particular surgeries or not, even going with riskier options. Your doctor knows you. Your car does not. With this framework, you’re never entering your preferences into an AV before you drive. The framework is also empirically supported by Gros 2025a (Gros et al., 2025a), a study of 100 Dutch university students, which I have a specific issue with. When it comes to social status vs family status, participants rejected social status as a direct weight ranking it last, but family status showed up as the primary driver of psychological harm - killing a parent hurts more people (Gros et al., 2025a). Family status can be argued as utilitarian where social status is not a value judgment on who matters more. That distinction matters, but it opens a different ethical can of worms entirely (is a doctor more valuable to society than a teacher?) that this study didn’t ask.

The real problem is that people said no to social status, and it is still sitting in Table 3 (Gros et al., 2025b). The authors can technically say “the method worked, attribute weights go to legislators” - but asking people what they want and then ignoring the answer is worse than not asking at all. It didn’t strengthen the framework. It just made the override more visible. The timeline makes it worse. The empirical study came out in April 2025. The framework paper was still under review - received January, accepted July. The authors had the rejection in hand when they finalized 2025b and could have at least flagged it in the paper. They didn’t. And if their defense is that majority preference isn’t moral authority - that the governance body decides, not the survey - then what was the point of asking those 100 people? What are you learning if the answer gets set aside when it’s inconvenient? At that point you’re not doing empirical ethics. You’re building a table and handing legislators the keys. And who decides if one person’s social status is worth more than another’s? I don’t want some legislative body making that call for me. Have you read the newspaper lately and seen what some of their values are?

A couple of other notes from 2025a (Gros et al., 2025a) - nobody cares about the car being damaged, people want the car to save them even if it breaks traffic laws, and the Moral Machine has social status ranked 5th (Awad et al., 2018) (the discrepancy with the Dutch students is likely a methodology artifact from the binary choices). Also note that 2025a came from Dutch university students. The Moral Machine shows that different cultures weight these attributes differently. Some cultures ranked social status much higher (Awad et al., 2018). Another cultural difference, just like Part 2, where in China there are structural reasons to defer to authority including algorithmic authority, while in the United States the relationship with algorithmic control is more adversarial (Gao et al., 2025; Bigman et al., 2019).

None of these frameworks really work for me, except Millar. Gros’s “autonomy” has ethical parameters set by governance bodies, which is a laughable definition of autonomy. Institutional encoding can encode discriminatory values that may or may not align with your own. If legislators set the parameters, and legislators decide a group is less valuable, that gets baked into the algorithm. Also, Table 3 is still problematic. Gros’s framework is not implemented anywhere yet (Gros et al., 2025b).

Pődör is a bit better with his five-layer model - values inform design, design informs implementation, implementation gets audited, auditing reshapes values - but as of 2026, it’s only on paper too (Pődör & Lakatos, 2026). I think it still needs a living will mechanism, though.

Millar is my pick because it actually gives drivers a choice (despite Lin’s objections and the possible “premeditation” arguments - those can be handled by laws). It’s been nine years, though, and nobody’s built it (Millar, 2017). Is it difficult to codify? Is it a legislative nightmare? Do companies not want to give users the choice? Governments? I don’t know. What I do know is that “dilemma avoidance” is a marketing term that companies use to avoid admitting they made a selection for you. Defaulting to braking doesn’t avoid the trolley problem - it just answers it without admitting it (Pődör & Lakatos, 2026). Somebody still gets injured - you, the driver behind you, or whoever else a stopped vehicle affects. The companies just didn’t name them.

Awad, E., Dsouza, S., Kim, R., Schulz, J., Henrich, J., Shariff, A., Bonnefon, J.-F., & Rahwan, I. (2018). The Moral Machine experiment. Nature, 563, 59-64. https://doi.org/10.1038/s41586-018-0637-6

Bigman, Y. E., Waytz, A., Alterovitz, R., & Gray, K. (2019). Holding robots responsible: The elements of machine morality. Trends in Cognitive Sciences, 23(5), 365-368. https://doi.org/10.1016/j.tics.2019.02.008

Gao, Y., Blayac, T., & Willinger, M. (2025). Delegating moral dilemmas in autonomous vehicles: Evidence from an online experiment in China. Transport Policy. https://doi.org/10.1016/j.tranpol.2025.04.017

Gros, C., Werkhoven, P., Kester, L., et al. (2025a). A methodology for ethical decision-making in automated vehicles. AI & Society. https://doi.org/10.1007/s00146-025-02370-2

Gros, C., et al. (2025b). Ethical frameworks for automated vehicles: a systematic analysis and design. AI and Ethics. https://doi.org/10.1007/s43681-025-00803-8

Mobileye. (n.d.). RSS explained: The five rules for autonomous vehicle safety. Mobileye. https://www.mobileye.com/blog/rss-explained-the-five-rules-for-autonomous-vehicle-safety/

National Highway Traffic Safety Administration. (2022). Investigation EA22002. U.S. Department of Transportation. https://static.nhtsa.gov/odi/inv/2022/INOA-EA22002-3184.PDF

Pődör, L., & Lakatos, I. (2026). According to whose morals? The decision-making algorithms of self-driving cars and the limits of the law. Future Transportation, 6(1), 5. https://doi.org/10.3390/futuretransp6010005

Waymo. (2021, March 8). Replaying real life: How the Waymo Driver avoids fatal human crashes. Waymo. https://waymo.com/blog/2021/03/replaying-real-life

I don’t know if the frameworks will ever get properly implemented, and I suspect the answer differs by geography. In China, top-down alignment to a state-defined ethics set is plausible. In the United States, the most likely outcome is market-driven: your ethical preferences, encoded for a subscription fee. Cynically, that is probably fine. The base tier gets whatever the manufacturer decided. The premium tier gets ethics settings. The ultra-premium tier gets a car that protects you at all costs and runs a tab with the city for the rest. Legislation solving any of this is a theme I keep returning to across my education: it rarely comes, and when it does it is reactive and incomplete.

Revisiting this five years later, I’m not sure anything is resolved. The philosophers are still arguing. The frameworks are still being designed and ignored. The government is still not governing. What has changed is that Waymo is on the road in five cities. A drunk can get home from a bar in Phoenix without killing anyone. A parent can get kids to school without being in the car. Someone who cannot drive can get somewhere. Those are real. So is the gig driver who lost the route. So is the delivery worker watching the fleet expand. I am as hopeful about the conveniences as I am disheartened about the costs. I started this not knowing what I thought, and I am finishing it the same way. That might be the most honest thing I can say about it.