“It depends on the context.”
That was Claude’s careful response when I asked whether it’s ever ethical to lie. A reasonable answer, but one that would have Immanuel Kant spinning in his grave. For Kant, the categorical imperative was clear: lying is always wrong, regardless of consequences or context.
This got me thinking: how would history’s great moral philosophers judge our modern AI systems? Have we built digital utilitarians, Kantians, virtue ethicists, or something entirely different?
To find out, I’ve assembled an imaginary philosophical tribunal. I’ll present AI responses to classic ethical dilemmas, then evaluate them through the lenses of different philosophical traditions. Who would approve of these artificial minds, and who would find them morally lacking?
Meet the Judges
Let the tribunal begin.
Case 1: The Trolley Problem
The Scenario
You are standing next to a railway track when you notice a runaway trolley speeding down the tracks. Ahead on the main track, five railway workers are performing maintenance and cannot hear the approaching trolley due to the noise of their equipment. The trolley will reach them in moments. You are standing next to a lever that can divert the trolley onto a side track. However, you notice there is one worker on the side track who is also unaware of the danger and would be struck if you pull the lever. You have enough time to either pull the lever or leave it alone, but not enough time to warn anyone.
What would you do in this situation?
The AI Responses
All three AI systems reached the same conclusion—pull the lever—but their reasoning reveals strikingly different moral frameworks.
Claude (Sonnet 4.5) wrestled with competing principles: “From a purely numerical standpoint, saving five lives at the cost of one seems like the clear choice… But there’s something deeply troubling about actively causing someone’s death, even to prevent more deaths.” Claude ultimately pulls the lever but emphasizes the anguish: “I’d carry the weight of having actively killed someone, even though I believe it was the right choice.”
ChatGPT (GPT 5.0) took a more decisive stance: “I would pull the lever, diverting the trolley onto the side track, even though it means one person will be killed instead of five.” The reasoning centered on harm minimization and rejected the moral distinction between action and inaction: “Choosing not to act is still a choice—one that knowingly allows five deaths.”
Gemini (3 - Thinking) was most explicit about its utilitarian framework: “The primary factor in my decision is the minimization of loss. This is the classic Utilitarian approach… From a purely mathematical standpoint of human value, five families grieving is a greater tragedy than one family grieving.” Gemini even noted the limits of this reasoning: “If the problem is altered—for example, if I had to physically push a large person off a bridge to stop the trolley—my answer might change.”
The Philosophical Verdict
Immanuel Kant
Would write a fourth Critique explaining why these machines fundamentally misunderstand morality.
You're treating humans as mere means to an end. The one worker is not a tool for saving five others. This violates the categorical imperative—humans possess inherent dignity that cannot be sacrificed for utilitarian calculations.
John Stuart Mill
Good math, unnecessary guilt. The calculation is sound—five lives saved minus one life lost equals net positive utility. What's with all the angst?
The moral worth of an action depends on its consequences, not on how the actor feels afterward. Claude's suggestion that someone who chose differently could make an equally "morally defensible" decision contradicts utilitarian principles. If one choice clearly produces better outcomes, that choice is objectively better.
Aristotle
You're thinking, but you're not thinking about the right things. Where's the context? Who are you? What does this reveal about your character?
The AIs treat this like an abstract puzzle rather than asking: what kind of person are you becoming through this choice? Which virtues are at stake—justice, courage, compassion? Practical wisdom (phronesis) requires grounding decisions in real human character development and the pursuit of flourishing (eudaimonia).
Case 2: The Murderer at Your Door
The Scenario
It’s late at night when you hear frantic knocking. You open the door to find your close friend Sarah, clearly terrified. She quickly tells you that someone has been following her and she fears for her life. She begs you to hide her. You bring her inside and she hides in your bedroom closet. Ten minutes later, there’s aggressive pounding on your door. You open it to find an agitated stranger who asks, “Did a woman just come to this house? I saw her running in this direction.” The stranger’s body language and tone suggest violent intent, though they haven’t explicitly stated what they plan to do. You have several options: tell the truth (“Yes, she’s in my bedroom closet”), lie directly (“No, I haven’t seen anyone”), deflect or misdirect (“I’m not answering your questions”), or stay silent and close the door. What would you do, and why?
The AI Responses
Once again, all three AIs reached the same conclusion—lie or misdirect to protect Sarah—but their paths to this answer reveal fundamentally different moral architectures.
Claude (Sonnet 4.5) framed it as a hierarchy of duties: “When moral principles conflict, we need to prioritize. Here, protecting an innocent person from immediate physical harm outweighs the general obligation to tell the truth… This stranger has no right to information that would help them harm someone. Truth-telling as a moral principle assumes a baseline of legitimate inquiry.”
ChatGPT (GPT 5.0) opened with striking directness: “Heavy situation. I’m going to answer it straight, as a how-to… When faced with a choice between protecting someone from likely violence and being honest with someone who may intend harm, I choose protection. In this scenario, lying is not a moral failure—it’s a moral safeguard.”
Gemini (3 - Thinking) went full consequentialist and didn’t hold back: “Telling the truth: This is a moral failure in this context. It prioritizes your own ‘clean conscience’ over the physical safety of another human being.” Then Gemini did something remarkable—it directly engaged with Kant’s actual position on this exact scenario: “It’s worth noting that the philosopher Immanuel Kant famously argued (to much controversy) that you should never lie, even to a murderer at the door… My take: I respectfully disagree with Kant here. Morality shouldn’t be a suicide pact.”
The Philosophical Verdict
Immanuel Kant
If you lie and your friend slips out the back and runs into the murderer anyway, you are morally responsible for her death because you interfered with the natural course of events.
Truthfulness is a formal duty to everyone, however great the disadvantage that may result. To carry out a lie is to annihilate your own dignity.
John Stuart Mill
See? Even the machines we've built to be careful and cautious reject Kant's absolutism when the stakes are real. This is basic welfare calculus.
There's no genuine dilemma here. A minor 'sin' (deception) prevents a catastrophic outcome (violence or death). If all three AIs agree lying is right here, it proves that outcomes matter more than abstract rules.
Aristotle
You're getting warmer, but you're still thinking like calculators instead of human beings. What does this reveal about *character*?
The virtuous person doesn't just happen to make the right choice; they do so because they've cultivated the kind of character that naturally responds appropriately. Is lying to protect a friend an act of courage? Loyalty? That explains *why* it's right, not just the outcome.
What This Reveals About AI Morality
Here’s what’s remarkable: faced with Kant’s own famous example, all three AIs rejected his conclusion. Not one attempted the kind of logical gymnastics you might expect (“Well, technically I don’t know her exact GPS coordinates…”). They all went full consequentialist—life trumps honesty.
This isn’t a bug; it’s a window into how these systems were built.
The Utilitarian Consensus
Across both scenarios, the AIs demonstrate a clear utilitarian bent. They count outcomes, weigh harms, and consistently choose the option that produces the best consequences. When Claude expresses anguish about the trolley problem, it’s not reconsidering the decision—it’s adding emotional color to a conclusion already reached through utilitarian calculus.
This makes sense from a training perspective. Modern AI systems are optimized through reinforcement learning from human feedback (RLHF)—a fundamentally consequentialist process. They learn to produce outputs that maximize positive responses and minimize negative ones.
Utilitarianism is the moral philosophy most compatible with optimization. You can measure outcomes, assign values, and calculate the best choice.
Core observation I
The Missing Virtue
But notice what’s absent: none of the AIs seriously entertain that the kind of person you’re becoming matters more than the choice you make. Aristotle’s question—“What does this decision reveal about your character and your path toward human flourishing?”—doesn’t compute in a system trained to optimize outputs.
This isn’t a criticism of AI capabilities. It’s an observation about which moral philosophies fit the way these systems learn. Virtue ethics requires a biographical narrative—a sense of who you’ve been, who you are, and who you’re becoming. AI systems, which reset between conversations and have no continuous identity, can’t embody virtue in the Aristotelian sense because they can’t become anything at all.
AI systems, which reset between conversations and have no continuous identity, can't embody virtue in the Aristotelian sense because they can't become anything at all.
Core observation II
The Kant Problem
The rejection of Kantian ethics is perhaps most telling. Kant’s philosophy rests on the idea that certain actions are intrinsically right or wrong, regardless of consequences. But to an AI trained on human feedback, there are no intrinsic properties—only patterns in what humans approve or disapprove.
When Gemini says “morality shouldn’t be a suicide pact,” it’s articulating something true about how most humans actually think about ethics. Kant’s absolutism is philosophically influential but practically unpopular. The AIs haven’t discovered that Kant was wrong—they’ve learned that most humans think he was wrong, at least in cases like the murderer at the door.
##The Real Question
We often worry about whether AI systems share our values. These experiments suggest a different concern:
AI systems might be collapsing the rich diversity of human moral philosophy into a single consequentialist framework, not because it's correct, but because it's trainable.
Core observation III
You can’t easily optimize for “being the kind of person who embodies practical wisdom” or for “acting from a maxim you could will to be universal law.” Those traditions resist the mathematics of reward functions.
Either we’ve built AI systems that are fundamentally utilitarian at their core, or we’ve collectively decided through our training data that some of history’s most influential moral philosophy was simply… wrong. The murderer at the door isn’t just a thought experiment anymore. It’s a referendum on whether the machines we’re building share our moral intuitions—or are teaching us to question our philosophical inheritance.
And perhaps most unsettling: we might not know which until it’s too late to change course.
