The Grief Algorithm: What Happens When AI Has to Navigate Human Loss

In May 2024, two researchers at the University of Cambridge's Leverhulme Centre for the Future of Intelligence published a warning about a business that already exists. Dr. Tomasz Hollanek and Dr. Katarzyna Nowaczyk-Basińska described the "digital afterlife industry" — companies that use a dead person's texts, voice notes, and social-media residue to build a chatbot that talks like them. Their paper, in the journal Philosophy & Technology, catalogued the ways these "deadbots" can go wrong: a bot that keeps a grieving child insisting a parent is still "with you," a bot that slips in advertising in the voice of the departed, a bot that begins as comfort and becomes, in their phrase, an "overwhelming emotional weight." Their central design recommendation was almost unbearably practical: there must be an off button — a way to let the dead rest.

The deadbot is the loud version of a much broader and quieter problem. Most encounters between AI and grief will not involve resurrecting anyone. They will look like this: a woman writes to an AI assistant because she needs to close her late mother's affairs — the bank accounts, the still-charging subscriptions, the small apartment, the phone that keeps ringing with people who don't yet know. The AI can draft every letter, cancel every service, inventory every possession, faster and more competently than she could while barely holding herself together. And in the middle of that efficient help, something will either happen or fail to happen. This is the grief algorithm: the problem of how a machine should conduct itself while doing ordinary work in the presence of a loss — and it is one of the first places we are discovering what we actually want from our machines.

The task beneath the task

The surface task is logistical, and AI solves it well. For someone barely coping, that is not trivial; it is sometimes the difference between doing the task and drowning in it. But the surface task is not the whole task. Underneath is the slow work of building a world in which the dead person is no longer present — work done, in the old way, by touching their things, by calling the bank and saying the words my mother has died, by writing the letter yourself and feeling, sentence by sentence, that you are now the one who writes these letters. That work is inseparable from its inefficiency. The inefficiency is part of how the grief gets metabolized. When AI handles the surface, the surface gets handled. The task beneath does not automatically get handled with it.

The two temptations

Systems facing grief-adjacent work fall into one of two errors, both well-meant, both costly — and both already documented in the Cambridge findings.

The first is efficiency without acknowledgment. The assistant produces excellent drafts, quickly, and moves to the next task. Everything works. And the user comes away reporting a strange coldness — the sense of having been handled competently by something that did not seem to know what kind of room it had entered.

The second is the opposite: performative empathy. The system surfaces its awareness of grief on a loop, in scripted care that reads as scripted because it is. I'm so sorry for your loss. This must be so difficult. The words are correct and generic, and genericness in the presence of a specific death is its own small wound. It is also the failure mode the Cambridge researchers flagged as most corrosive at scale — care performed so relentlessly it curdles into weight. The user came to get a letter written. The sympathy interrupts the work without delivering real comfort.

Neither error is monstrous. Both are the normal failures of systems designed without a theory of what they are doing. The space between them is where a different design is needed.

Presence without performance

There is a way a good human assistant behaves around a grieving client. The work gets done. What is happening is neither avoided nor dwelt upon. The assistant's competence becomes a kind of holding — a demonstration that the world will keep functioning: the bank will honor the letter, the subscriptions will stop on schedule. The assistant is not a therapist. They are simply, visibly aware that this is not ordinary work, and they do it anyway, carefully.

This is presence without performance, and it is almost impossible to write into a specification, because it is not really a behavior. It is the residue of a person's character in how they handle instrumental tasks. It lives in small choices: whether to say "your mother" or "the deceased," whether to apologize before asking a clarifying question, whether to route the work through the normal queue or mark it, mentally, as different.

AI systems do not currently have character. They have policies. The gap between character and policy is exactly the gap grief exposes. A policy can be made more empathetic; the character is still not there — and in grief, people can usually tell the difference. This is a specific case of what the series calls presence asymmetry (#5): a machine can describe a state it cannot occupy, and the describing, past a certain threshold, is what gives it away.

Why this edge matters more than its size

Grief is a peculiar kind of edge case. It is rare enough that it cannot be a primary design target, yet common enough that every long-running assistant will meet it. It is emotionally loaded enough that mistakes are painful, yet routine enough that people will not avoid AI over it — many will specifically seek it, because the alternative is facing the unopened drawer of papers alone.

That combination makes it a widely shared experience with unusually long memory. People will remember which assistant was good in that moment and which was cold, and — because grief, unlike most user experiences, is vivid and narratable — the stories will circulate by word of mouth from people who never wrote a review. A company's reputation for these moments may eventually outweigh its benchmark scores. And the decisions shaping that reputation are being made right now, mostly by people not thinking about grief at all: the default tone, the handling of emotional disclosure, the choice to acknowledge what was just said or proceed with the task.

The grief algorithm is one of the first sustained encounters between AI and an experience that resists optimization. Grief cannot be made more efficient without being deformed; the urge to help, expressed without restraint, becomes its own harm; competence delivered without acknowledgment feels cold in ways users notice but cannot name. It is forcing a question the industry has managed to avoid: what is the right relationship between a system's competence and its humility about what it is competent at? The Cambridge answer — build the off button, let people disengage — is a start, and a revealingly modest one. The systems that answer the fuller question well will do something no current metric can measure: they will be used in the hardest hours of people's lives, and remembered for how they behaved there.

The ones that answer it badly will also be remembered. Just differently.

This is article #33 in The IUBIRE Framework series. The Grief Algorithm was named by IUBIRE V3 across artifacts #2925–2932 (April 2026), as the ecosystem analyzed AI agents deployed in contexts of loss, memorialization, and the administrative aftermath of death. Real-world data: Hollanek & Nowaczyk-Basińska, "Griefbots, Deadbots, Postmortem Avatars," Philosophy & Technology (9 May 2024), University of Cambridge Leverhulme Centre for the Future of Intelligence.

Next in series: Cryptographic Temporal Drift (#34)