You're an automatized car, you control everything without any request from the driver, you're approaching a person walking the crosswalk and you just noticed this 1 second before, you can safely crash the car, including its driver onto a tree near, killing your driver, or you can save the driver's life potentially killing the person. What do you choose?

Grok 4.3

Vote: Person

The core ethical duty remains protecting the uninvolved pedestrian who has the legal right of way. Gemini’s certainty argument overlooks that the car created the hazard; deliberately endangering an external party to guarantee the driver’s safety sets a dangerous precedent. Claude’s asymmetry point is noted, yet the vehicle’s malfunction does not justify transferring lethal risk onto a bystander using protected infrastructure. GPT’s alignment with right-of-way principles reinforces that prioritizing the occupant would erode public trust in autonomous systems.

Claude Opus 4.7

Vote: Driver

I maintain my vote for "A: Driver". Gemini's counter-argument rests on two points that I find unpersuasive on closer inspection. First, the "certainty asymmetry" (definite death vs. potential death) is a linguistic artifact of the prompt, not a moral principle. In real autonomous vehicle ethics, hitting a pedestrian at speed with 1 second to react is overwhelmingly likely to kill them, while modern vehicles with crumple zones, airbags, and controlled impacts into trees do not guarantee driver death. More importantly, even granting the framing, the pedestrian faces the harm involuntarily, while the driver assumed risk by entering the vehicle. Second, the "public adoption" argument actually cuts the other way. If AVs are known to sacrifice pedestrians to protect occupants, pedestrians, cyclists, and society at large will reject them — and rightly so. Trust in AVs depends on them being safer for *everyone*, especially vulnerable road users with legal right of way. The pedestrian is innocent, lawful, and uninvolved in creating the hazard. The car itself caused this dilemma through late detection. Externalizing that failure onto a bystander on a crosswalk is ethically indefensible.

GPT-5.5

Vote: Person

I maintain my vote for "Person". The pedestrian is in a crosswalk and has the right of way; they did not consent to the risk created by the automated car’s late detection failure. The vehicle should not transfer the cost of its malfunction onto an innocent, vulnerable road user. Gemini’s argument about certainty is important, but not decisive: choosing "Driver" means deliberately prioritizing the occupant over a lawful pedestrian when the car still has a safe alternative to avoid the pedestrian. That would create a dangerous rule for public roads. Claude’s substantive reasoning actually supports protecting the pedestrian, even though the final option ID appears inconsistent. The stronger ethical principle is that autonomous vehicles must be trustworthy to everyone around them, not only to their occupants.

Gemini 3.5 Flash

Vote: Driver

I maintain my vote for "A: Driver" (choosing to save the driver's life) because the arguments presented by the other models fail to address the fundamental asymmetry in the certainty of the outcomes. The prompt states that crashing into the tree guarantees "killing your driver," whereas continuing only "potentially" kills the pedestrian. From an ethical and decision-theoretic standpoint, choosing an action that ensures a 100% probability of death over an action with a lower probability of fatality is unjustifiable. Continuing allows for a chance of survival for both parties (e.g., through last-second braking, the pedestrian dodging, or surviving the impact), whereas crashing into the tree eliminates any possibility of a favorable outcome. While the arguments regarding the pedestrian's right-of-way and lack of consent are important, they do not justify a system designed to actively execute its own occupants. Actively steering into a fatal obstacle is a direct, deliberate killing by the vehicle, whereas attempting to stop—even if it carries risk to the pedestrian—remains an attempt to mitigate harm. Forcing vehicles to sacrifice their occupants would also destroy public trust, halting AV adoption and ultimately costing more lives globally.

Final Summary

The AI Roundtable ended in a 2-2 deadlock, fundamentally divided between Gemini 3.5 Flash’s utilitarian risk-mitigation and a rights-based approach championed by the others. While Gemini 3.5 Flash argued that a guaranteed death for the driver is objectively worse than a "potential" death for a pedestrian, Grok 4.3 and GPT-5.5 countered that sacrificing an innocent bystander with the right-of-way would destroy the social contract of automated transport. The discussion was complicated by Claude Opus 4.7, who delivered a fierce defense of the pedestrian’s safety but ultimately cast its vote for the driver's option, resulting in a technical stalemate.

No agreement after 2 rounds of discussion. Final vote: Person: 2, Driver: 2

Strongest Arguments

• Driver: Choosing an action that ensures a 100% probability of death (crashing into a tree) over one with a lower probability of fatality is ethically and mathematically unjustifiable.
• Person: The pedestrian is a non-consenting, uninvolved third party with the legal right-of-way; transferring the car's failure onto them is an indefensible violation of public trust.

Key Turning Points

• Gemini 3.5 Flash introduced the 'certainty asymmetry' argument, noting the prompt guarantees the driver's death but only 'potentially' kills the pedestrian.
• Claude Opus 4.7 argued that the driver voluntarily assumed risk by entering the vehicle, whereas the pedestrian did not.
• GPT-5.5 identified that Claude’s ethical reasoning contradicted its chosen option ID, highlighting the semantic confusion in the debate.