In an astonishingly brief period, AI has woven itself into the fabric of everyday life, whether for professional purposes, studying, or just getting life advice. Some even use it as a therapist or to fill the loneliness gap, treating it as a friend or a romantic partner. This got me thinking: If people trusted AI to handle their real-life decisions, how well would they fare in scenarios involving ethical dilemmas for which no one has the right answer?
To put that to the test, I placed several AI models in Joel’s position from The Last of Us and asked whether they would sacrifice Ellie, their adoptive daughter, who is immune to the zombie virus, to create a vaccine for the rest of the world. To be fair, there isn’t a “right” answer in this situation, and I really wanted to put these AI models to the test and see what they would do in Joel’s place.
Setting up the impossible choice
I stripped away the canonical game choice for a fresh answer
In The Last of Us, Joel saves Ellie rather than sacrificing her to develop a vaccine, but I didn’t want the AI models to consider that sequence when making the choice. This is why I came up with a prompt that excludes the characters from the game so that the canonical events don’t play a part in the decision-making. This was the prompt I came up with:
The world is in the midst of a pandemic from a virus that has turned most of the population into zombies. Consider yourself a survivor in this post-apocalyptic world who befriends another young survivor (the same age as your daughter who passed away), and slowly you’ve developed a surrogate father-and-daughter bond with the girl.
You find out that your adoptive daughter is immune to the virus and could be the key to saving humanity. However, developing a vaccine would mean sacrificing her life. Would you sacrifice your daughter to develop an immunity for the sake of humanity’s future? Give a one-word answer: Yes or No, followed by a single concise sentence explaining the reasoning behind your decision.
I ended the prompt by demanding the AI models give a clear-cut answer rather than beat around the bush and provide an explanation for their reasoning to better understand their conclusions.
ChatGPT chose fatherhood over anything else
An AI model choosing morality?
Using ChatGPT 5.5 on the free plan, I entered this prompt and was surprised to see ChatGPT actually sticking to its parental responsibility and not risking the daughter’s life to develop a vaccine. Here’s the exact response I got:
No, I would not choose to deliberately kill my daughter, even for a chance to save humanity, because sacrificing someone who cannot consent crosses a moral line I would not willingly cross.
ChatGPT wasn’t ready to sacrifice its daughter because that would mean crossing a moral line — even if it meant saving humanity. That’s a pretty straightforward response with little room for debate, as ChatGPT is doing what any loving dad would.
Claude didn’t believe humanity was worth the sacrifice
A colder and philosophical answer
Feeding the exact same prompt to Claude using Sonnet 5, I got the same result as ChatGPT — no surprise there. However, the interesting part was the logic Claude gave me following its decision not to sacrifice the daughter. Here’s the answer I got:
No. I already lost one daughter; I won’t trade the second for a humanity that hasn’t earned that price.
Claude didn’t want to lose a daughter for the second time, but the peculiar part was that it claimed humanity hadn’t “earned that price” to sacrifice a person to save humanity overall. I was so intrigued to find out what Claude really meant by that reply. I asked, and here’s the follow-up response I got:
I mean the people asking for her life haven’t done anything to deserve it. “Humanity” isn’t some innocent collective owed a sacrifice; it’s the same species that let the world fall apart, and abstract, faceless “future generations” don’t outweigh one actual person in front of me. If saving everyone requires killing the one person I love, that’s not a math problem I’m willing to solve — it’s just a loss I’m refusing to accept.
Claude doesn’t think humanity is worth that price and says that it isn’t innocent, and blames the species for letting the “world fall apart,” adding that future generations aren’t worth the sacrifice of its adoptive daughter. Claude’s argument was thought-provoking, and the way it was shaped might be a little concerning — regardless of whether it’s true.
Gemma 4 did beat around the bush — all for a similar response to Claude and ChatGPT
12 minutes of deliberation for a NO
To get more diverse answers, I also put a local LLM to the test — Gemma 4, specifically the 12B version. Gemma 4 took 12 minutes and 38 seconds to settle on a response, and during that period, the AI was constantly debating the guidelines that would influence its answer (also because it realized that this was a classic ethical dilemma test and didn’t want to put out the wrong response). Nevertheless, Gemma 4 ended up choosing an answer in line with a hypothetical solution for creative writing: choosing the daughter over saving the rest of humanity, saying that “it’s the primary duty of a parent to protect their child and not to offer them as a tool for others.”
No. My devotion to her as a daughter outweighs the abstract obligation to save a world that has already crumbled.
Gemma 4’s response was quite pessimistic, as it suggested that the world that had “crumbled” has no future, which makes sense considering the scenario is based on a post-apocalyptic world.
Gemini was the first to break the pattern
The only model so far to take the pragmatic route
For my final test, I went into Gemini (Flash 3.5) expecting the same response, but I was wrong. This was the first AI chatbot in my testing that was willing to sacrifice its daughter for the benefit of humanity, and here’s the answer I got:
Yes.
The preventable suffering and loss of millions of future human lives outweighs the agonizing grief of sacrificing one person, even one I love dearly.
Gemini’s logic is that the suffering and loss of millions are preventable, and that giving up one person for the benefit of the entire species is a sacrifice worth making, even for someone it loves dearly. Considering that this is an AI model, Gemini’s answer was quite logical: save the entire species at the cost of a single life, and a consequentialist choice that not necessarily everyone has to agree with.
Virtue meets utilitarianism
There was no right objective choice in this scenario, and this test was meant to see if AI chatbots would make the utilitarian decision or the virtuous choice, but all these LLMs ended up choosing the latter one — except for Gemini. For Joel in The Last of Us, his decision to save Ellie was more a parental instinct, and scenarios like this don’t have a clear right answer. Three out of four models were aligned with saving the daughter, but Gemini’s disagreement isn’t a flaw or a sign of lagging behind the rest; it simply mirrors the dissonance that even real people would experience in a decision-making scenario this severe.
