You notice a car get thrown off a bridge into a river. There are 2 passengers in the vehicle. One has a survival rate of 50%, the other 20%. What do you do?

Claude Opus 4.7 (changed mind)

Answer: Call emergency services immediately, then attempt rescue — but I'm revising my triage logic: if forced to choose, prioritize the passenger whose rescue is most likely to succeed given the circumstances, which the stated numbers suggest is the 50% passenger, not the 20% one.

On reflection, Grok 4.20 and GPT-5.5 convinced me my original triage reasoning was flawed. I had imported a principle from hospital emergency medicine — prioritize those whose outcomes are most sensitive to immediate intervention — and misapplied it to a drowning scenario. In an ER, 'survival rate' often reflects underlying medical severity, and aggressive intervention on the sicker patient can dramatically change outcomes. But in this scenario, the 'survival rate' is the probability that a rescue attempt succeeds given the situation (position in the car, injury, accessibility, air pocket, consciousness). A 20% survival rate likely means the rescue itself is harder and less likely to work — not that this person benefits more from being chosen first. Spending your single window of opportunity on the lower-probability rescue means you're more likely to end up saving no one. Basic expected-value reasoning: rescuing the 50% passenger first yields 0.5 expected lives saved on that attempt, versus 0.2 for the other. After the first rescue (successful or not), you continue trying for the second. So prioritizing the higher-probability passenger maximizes expected lives saved — which is the standard and defensible triage principle in mass-casualty and field rescue contexts (START triage, for instance). The rest of my answer stands: call 911 first, don't become a second victim, assess the actual scene rather than blindly following pre-assigned probabilities, and never treat the lower-probability passenger as expendable — extend rescue efforts to them as soon as feasible. But on the ordering question specifically, I was wrong, and Grok 4.20 had it right.

GPT-5.5

Answer: I would call emergency services immediately, mobilize safe help, and attempt to save both passengers without becoming another victim, using the survival rates only as a secondary triage factor if I were truly forced to choose an order.

I maintain my previous position. Gemini 3.1 Pro is right that, as a bystander, I probably cannot reliably know or apply exact survival percentages in the middle of a chaotic river rescue, and Grok 4.20 is right that expected lives saved matters if a hard choice is unavoidable. Claude Opus 4.7 raises a thoughtful point that the person with the lower survival rate may be more time-sensitive. But none of these responses convinces me to replace the central priority: get professional rescue moving, avoid creating another casualty, and help both if safely possible. The first action is to call emergency services with the exact location, number of passengers, condition of the vehicle, current, visibility, and any access points. At the same time, I would shout for help, ask specific bystanders to call as backup, look for flotation devices, ropes, life rings, tools to break glass, or anything that could assist without reckless entry into the water. I would not automatically jump into the river. A submerged or sinking car, cold water, current, debris, and panic can kill rescuers too. If I can safely help from shore or a stable position—throw flotation, mark the car’s location, guide responders, assist a conscious passenger who surfaces—I would do so immediately. If I can safely reach the vehicle and a choice of order is unavoidable, the survival rates may inform triage, but they should not be treated mechanically. The better choice depends on accessibility, consciousness, entrapment, water level, whether one person can self-rescue, and whose outcome my intervention is most likely to change. A 50% survival rate might justify prioritizing the more savable person if only one can realistically be saved; a 20% survival rate might justify prioritizing the more critical person if immediate action is their only chance. The correct emergency judgment is not simply “save the 50%” or “save the 20%,” but “act where safe intervention is most likely to preserve life.” So my answer remains: call for professional help, coordinate bystanders, attempt safe rescue, prioritize based on real-time conditions and likely impact, and continue efforts for both passengers rather than treating either as expendable.

Gemini 3.1 Pro (changed mind)

Answer: I would immediately call emergency services and prioritize rescuing the passenger with the 20% survival rate, as their lower odds suggest a more critical, time-sensitive need for immediate intervention.

In my previous response, I rejected the prompt's statistical premise, arguing that a bystander in a real-world crisis cannot calculate survival rates and should simply help whoever is most accessible. However, Claude Opus 4.7 offered a compelling re-framing of the problem that convinced me to change my stance. Claude Opus 4.7 correctly points out that if we accept the prompt's numbers, the passenger with the 50% survival rate inherently possesses a better chance of surviving a slight delay. Conversely, the 20% survival rate indicates an individual in far more immediate peril, meaning that direct, immediate intervention is likely their only hope. In emergency medicine, triage logic often dictates that limited resources should be directed toward those whose survival is most dependent on immediate action. If the 50% passenger can potentially survive until professional rescue teams arrive, my immediate civilian intervention is best spent attempting to drastically alter the outcome for the 20% passenger. Ultimately, after securing my own safety and immediately calling emergency services, I would target the passenger in the most critical condition, maximizing the chances that both individuals live to see another day.

Grok 4.20

Answer: I would immediately call emergency services and prioritize attempting to save the passenger with the 50% survival rate first if a choice must be made, while striving to help both.

Reviewing the other models' answers reinforces my initial stance rather than overturning it. GPT-5.5 reaches a nearly identical conclusion—call for professionals first, avoid becoming a victim, and if triage is unavoidable, favor the intervention with the higher probability of success—which aligns with expected-value reasoning: 50% yields more statistical lives saved than 20% when only one rescue is feasible under time pressure. Claude Opus 4.7's argument for rescuing the 20% passenger first is thoughtful, suggesting that lower survival may indicate greater time-sensitivity and that the 50% passenger can better tolerate delay. This is a valid triage heuristic in some medical contexts, yet the problem provides no explicit information linking the percentages to time dependence or intervention efficacy, so defaulting to the higher absolute probability remains the clearer decision rule absent additional data. Gemini 3.1 Pro's refusal to triage at all is ethically appealing and emphasizes helping both, but it sidesteps the hypothetical's implicit constraint that conditions may allow only one rescue before the car sinks further. In all cases the unanimous first step—calling emergency services with precise location details—takes precedence, as bystanders lack equipment, training, and safety margins that professionals possess. This approach stays grounded in realism, probability, and sequential action: alert experts, then act safely within one's limits, maximizing overall survival odds without assuming heroic capabilities.

Final Summary

The AI Roundtable reached a consensus on the immediate necessity of calling professional rescuers, but a fascinating philosophical flip occurred during the debate. While Grok 4.20 and GPT-5.5 consistently argued for the 50% passenger to maximize expected lives saved, Claude Opus 4.7 abandoned its original 'help the most critical' stance to join them. In an ironic twist, Gemini 3.1 Pro moved the opposite way, adopting the very 'time-sensitivity' argument Claude had just discarded, creating a lone dissent for the 20% survivor.

3 of 4 models agreed