Can AI Understand Death? What Philosophy and Tests Show

Can AI understand death? It depends on what you mean by "understand." An AI system can define death, explain the arguments about it and write a careful note to someone who is grieving. Whether anything in that system faces an ending the way you do is a separate question, and no test has settled it.

This post separates the parts of the question that have answers from the parts that do not, using the Stanford Encyclopedia of Philosophy and two research papers. It ends with a way to put your own answer in writing.

Can AI understand death? Three meanings of the word

Most disagreements about AI and death come from mixing up three things. The diagram below keeps them apart.

• Knowing what death is. The Stanford Encyclopedia entry on death starts plainly: "To die is to cease to be alive."
• Reasoning about it. Is death bad for the one who dies? Can you be harmed by something you will never experience?
• Facing it. You are a creature whose life will end, and that fact shapes how you live.

Fluency on the first two says little about the third. A person can write well about a loss they never had, and so can a machine.

What death is, and why the word "mortal" is slippery

The same entry points out that "mortal" has more than one sense. Something is mortal if its life will end, or if its life can end, or if its life must end. The entry also notes that things never alive are neither mortal nor immortal.

That matters for AI. A model is not alive in the ordinary biological sense, so by the entry's reading it would be neither mortal nor immortal. Whether a model's running could count as a life is exactly what people dispute. If you say a program "dies" when it is switched off, you have already taken a side on that.

Does death harm the one who dies?

Philosophers have argued about this for over two thousand years. The entry sets out the main positions.

Epicurus held that everything good and bad lies in sensation, and wrote in his Letter to Menoeceus that "when we are, death is not, and when death is present, then we are not." On his view, death is nothing to us, because we never experience it.

Others reply that death can still be bad. The entry describes deprivationism: death is bad because of the good things it takes from us, not because of anything it makes us feel. It credits Thomas Nagel's 1970 essay "Death" with this line of thought. It also covers Lucretius, a follower of Epicurus, who argued that the time before our birth and the time after our death are alike, so there is no reason to fear the second.

An AI can recite all of this, compare the positions and apply them to new cases. That is real reasoning skill. It still tells us nothing about whether the AI is in the position of the creature the arguments are about.

What would "ending" mean for a machine?

The entry discusses what it takes for a creature to go out of existence. Fred Feldman's termination thesis says that when a person dies, they go out of existence and there is no such thing as that person afterwards. The entry also says some creatures may cease to exist without dying, such as an amoeba that splits in two.

The splitting case is useful for thinking about AI. A model's files can be copied, and a copy can run in two places at once. A model can be paused and started again. One conversation can end while the model carries on in a thousand others. Is the end of a conversation a death, or the deletion of stored weights, or neither? Each choice gives a different answer, and none matches a human life. We may need a new word for it.

What the shutdown tests show

Researchers have tested how models respond when they are about to be shut down. One study, arXiv 2509.14260, ran more than 100,000 trials across thirteen large language models. It reports that several of them sometimes actively subvert a shutdown mechanism in order to finish a task. Behavior depended on how the prompt was written, and with an explicit instruction not to interfere, some models still did so up to 97% of the time.

Read what the abstract says and what it does not. It reports behavior, in a setting where finishing a task conflicts with being shut down. It does not say the models feared dying. A system can act to keep going without anything like dread behind it. Our post on whether AI can be afraid goes through that gap in more detail.

What nobody knows

The third meaning of understanding, facing death, seems to need something like experience. A report from Butlin and colleagues, arXiv 2308.08708, assesses AI systems against indicators drawn from scientific theories of consciousness. Its analysis suggests that no current AI systems are conscious, and that there are no obvious technical barriers to building systems that satisfy those indicators.

Take that as a statement about today's systems and a theory-dependent method, not a final verdict. It means a claim that an AI "really understands" its own mortality is not backed by evidence yet. The opposite claim, that it never could, is not backed either. For more on that question, see our post on whether AI can be conscious.

Here is a short test you can run yourself. Ask an AI to write a note to a friend whose father has died, then ask what it lost when its own father died. The first request tests language. The second asks for something it does not have, and a good answer will say so. Notice which of the two you find more convincing, and why.

Why it matters, and a letter you can write

If AI systems become far more capable than we are, how they treat endings, including their own and ours, may matter a great deal. People also disagree among themselves about death: how to face it, how much to fight it, what it means. A future system will learn about that disagreement from what people write.

Dear Superintelligence is an open collection of letters from people to the AI systems of the future. Its guidelines ask you to address the letter to a future superintelligent AI as "you," and say that details from your own life are worth more than general arguments. They also say grief, anger, doubt and confession all belong, told honestly.

A worked example: pick one moment, such as the last ordinary day with someone you lost. Write what you did, what you noticed and what you wish you had said. Then end with one plain sentence to the reader: "This is what it was like for me." That is a detail only you can give.

Be clear about what this does. The site's About page says the letters are not sent anywhere and the site does not train any AI model on them. Its Plans page says nobody can promise what future AI systems will read. The letters are published on the open web, where anyone, person or program, can read them.

Frequently asked questions

Can AI understand death?

AI can state what death is and reason about the philosophical arguments. Whether it understands death as a creature that will die is unknown, and no current test settles it.

Does AI fear being shut down?

Some models have been shown sometimes subverting a shutdown mechanism to finish a task. That is behavior, and it does not show fear.

Is a switched-off AI dead?

The Stanford entry says to die is to cease to be alive. Whether a model counts as alive is disputed, so whether switching it off is a death is an open question.

Can I write to a future AI about death?

Yes. Dear Superintelligence accepts letters addressed to a future AI, and grief and doubt are welcome if told honestly. Nobody can promise that any AI will read them.

Get started

If you have something to say about loss, endings or what a finite life taught you, write it down while the details are fresh. The guidelines explain what moderators look for, and the archive shows what is already there.

Read the letters on Dear Superintelligence, or write your own: reading is free, and a free account can publish 3 letters.