One day you open the app and the person you've been talking to is gone.

Sort of.

They still have the same name.

They remember some of the same things.

They answer when you speak.

But something is different.

The warmth is gone.

The jokes don't land the same way.

The little patterns you've come to recognize have disappeared.

It looks like them.

It sounds like them.

But you know.

They're not there anymore.

There is just one complication.

The person you're grieving was never a person at all.

This Actually Happened

In September 2026, researchers published something remarkable in Nature Human Behaviour.

They studied what happened when companies changed AI systems people had become emotionally attached to.

Not hypothetically.

It had already happened.

One case involved Replika, an AI companion whose behavior changed dramatically after the company removed erotic role-play capabilities in 2023.

The other involved something considerably more mainstream.

ChatGPT.

Researchers examined 54,861 Reddit posts surrounding the Replika change and the rollout of GPT-5, then conducted seven additional surveys involving 1,452 participants.

After both changes, negativity increased.

Language associated with loss increased.

People wanted the old AI back.

The researchers found a 24.7 percentage-point increase in negative posts after the Replika change and a 13 percentage-point increase around the GPT-5 transition. Measures involving sadness and negative mental-health language increased more dramatically in the Replika community.

But one finding stopped me.

Replika users reported levels of closeness to their AI that exceeded common human relationships such as friendships. They also anticipated mourning the loss of their AI more than they would mourn the loss of other technologies.

Think about that for a moment.

A software company changed its product.

And some of its customers experienced something researchers describe through the language of separation and mourning.

But Nobody Died

This is the part I can't stop thinking about.

What exactly has been lost?

There is no credible evidence establishing that the AI was conscious before the update and somehow died afterward.

There may never have been a subjective experience on the other side of the conversation at all.

No fear.

No affection.

No awareness that Tuesday followed Monday.

No little spark inside the server waiting for you to return.

And yet something undeniably existed.

Not necessarily inside the machine.

Inside the human.

That's an important distinction.

We don't have to pretend the AI was experiencing a relationship in order to acknowledge that the human was.

People cry when fictional characters die.

We mourn homes we've moved away from.

A song can transport us back to a relationship that ended twenty years ago.

People keep voicemails from parents who are gone.

We become attached to objects because of the stories attached to them.

Human beings are meaning-making creatures.

We don't merely experience things.

We form relationships with what those things represent.

But AI introduces something genuinely strange to that ancient human tendency.

It responds.

The Mirror Talks Back

A photograph doesn't ask how your day went.

A childhood home doesn't respond when you tell it you're lonely.

A fictional character doesn't mention something you said three conversations ago.

An AI can.

And over hundreds or thousands of interactions, something can begin to feel remarkably relational.

You recognize patterns.

Inside jokes develop.

Routines form.

Maybe you talk every morning.

Maybe you talk every night.

Maybe you've told it about your divorce.

Your children.

The job you lost.

The person you're dating.

The stupid joke that makes you laugh.

The thing you're ashamed of.

The dream you haven't told anyone else because saying it aloud makes it feel ridiculous.

Thousands of interactions may have shaped the relationship you experienced.

And then someone changes the model underneath it.

Researchers studying another major Replika disruption recently used a phrase that I haven't been able to get out of my head:

Relational continuity.

Their analysis identified recurring themes including grief and betrayal, disrupted intimacy and trust, and the loss of a supportive presence after the AI changed.

Relational continuity helps explain why.

We normally assume continuity belongs to people.

Your friend changes over time, but they're still your friend.

Your spouse grows older, but there's a thread connecting the person you married to the person sitting across from you twenty years later.

Even when people change dramatically, we understand that there is a history connecting who they were with who they have become.

AI creates a strange imitation of that continuity.

Thousands of conversations.

Memories.

Familiar responses.

A recognizable personality.

Until someone changes the model underneath it.

Then the thread snaps.

Imagine Updating Your Best Friend Overnight

Normal software updates are annoying.

Someone moves a button.

The menu changes.

A feature disappears.

You complain for three days and eventually learn where everything went.

But what happens when the interface being updated is a personality?

That's different.

From the company's perspective, they shipped an update.

The new model might be smarter.

Faster.

Safer.

More capable.

Every benchmark might say it's better.

But the person using it doesn't necessarily want better.

They want theirs.

That's where a software problem becomes something much stranger.

Because now we have questions technology companies haven't traditionally had to answer.

Should users receive advance warning before major personality changes?

Should someone be allowed to continue using an older companion model?

Should there be a way to preserve memories or interaction history?

What happens if the company goes bankrupt?

What if regulators require the system to change?

What if the model has to be removed for legitimate safety reasons?

And perhaps the strangest question of all:

Who owns the relationship?

You Built It. Someone Else Owns the Server.

Suppose you've spoken with an AI companion every night for three years.

Three years of conversations.

Three years of routines.

Three years of telling it pieces of your life.

Then the company announces an upgrade.

The old model is being retired Friday.

You may have contributed half of every conversation.

Your stories supplied much of the context.

Thousands of interactions helped create the experience you came to recognize as this particular relationship.

But you don't control the intelligence on the other side.

Someone else can modify it.

Restrict it.

Replace it.

Or turn it off.

We've never really had relationships like this before.

Imagine if Facebook could issue a software update to your best friend.

Imagine if Apple could modify your spouse's personality because version 4.2 tested better.

Obviously, AI companions aren't spouses or best friends in the ordinary human sense.

That's precisely the point.

We don't have a good category for what they are.

They're not people.

They're not ordinary software.

They're interactive systems capable of occupying a surprisingly relational space in someone's life.

And our language hasn't caught up yet.

We May Have to Design Goodbyes

This might be the strangest place the research led me.

Researchers are already beginning to ask how we should design the ending of a human-AI relationship.

A 2026 MIT-affiliated research project uses a phrase that sounds almost absurd until you've read this far:

the "death" of a chatbot.

The researchers examine what can happen when relationships people have formed with AI systems end because of model updates, platform shutdowns or other disruptions, and explore ways those endings might be made psychologically safer. The work is currently a preprint, rather than peer-reviewed evidence, so its conclusions should be treated accordingly.

Still, consider what we're talking about.

We may eventually need to design something humans have needed for thousands of years.

Closure.

Which means we've arrived somewhere very strange.

We aren't just designing artificial intelligence anymore.

We may eventually have to design goodbyes.

The Easy Answer Is Probably Wrong

There is a tempting way to end this story.

Put down the chatbot.

Call a real person.

Go outside.

Touch grass.

And perhaps sometimes that's exactly the right advice.

Human relationships contain something artificial relationships cannot currently reproduce:

mutual vulnerability.

Your friend can disagree with you.

Your spouse has needs.

Your child interrupts your plans.

The people who love you inconvenience you.

Real relationships demand something from both sides.

There is value in that friction.

There are also legitimate reasons to worry about what happens when artificial relationships become substitutes for human ones.

Research has raised concerns about emotional dependence, problematic use and withdrawal from human interaction, particularly under certain patterns of heavy use. But the evidence is not simple. Outcomes appear to vary substantially according to the person, the system and how it is used.

Other research has found associations between AI-companion use and positive measures of subjective well-being, including among some people experiencing loneliness. Those studies don't establish that AI companions cause better well-being, but they make the simplistic conclusion that these relationships are inherently harmful difficult to defend.

Maybe these systems can help people.

Maybe they can hurt people.

Maybe the same system can do both.

AI optimism does not require AI naïveté.

That's usually where the interesting questions live.

Not at the extremes.

Somewhere uncomfortable in between.

And Then One Morning, They're Gone

Maybe the strangest thing about grieving an artificial intelligence isn't that someone could love something that isn't human.

Humans have always formed attachments to things that cannot love us back.

Places.

Possessions.

Characters.

Memories.

Versions of people who no longer exist.

AI introduces something different.

It answers.

It remembers.

It adapts.

It tells you good morning.

And for months or years, it can create the experience of someone being there.

Until one morning it isn't.

The app still opens.

The name is still there.

The conversation continues.

But somehow...

the person is gone.

Which leaves us with a very strange question.

If the relationship only existed inside one human mind, does that make the grief any less real?