Can a machine ever really…feel?

The Her Problem

I have a confession to make: I am extremely late to the award-winning 2013 movie Her.

Thirteen years after Theodore fell in love with Samantha, I have finally caught up with the rest of civilisation, and, somewhat ironically, after writing several articles about artificial intelligence, emotional avoidance and our increasingly complicated relationship with actual human feelings, it took me until 2026 to watch the film that had apparently been sitting there making this argument for me all along.

In my defence, I have been busy and just haven’t really been able to sit down to watch any films outside of the flight or the theatre.

Long before ChatGPT became everyone’s unpaid therapist and occasional existential crisis hotline, science fiction had already been asking what might happen when machines become intelligent enough to think, reason, empathise, manipulate and, eventually, feel.

The conversation around AI and emotion is hardly new. For decades, films and fiction have imagined the moment artificial intelligence crosses some invisible threshold from sophisticated machine to conscious being, and things tend to go badly from there. But watching Her in 2026 made me wonder whether we have been asking the question backwards.

The obvious question is: can AI feel emotions?

But perhaps the more fundamental question is: what makes an emotion an emotion in the first place, and what, exactly, makes a human emotion human?

Is it the feeling itself? The body that produces it? The biology attached to it? Or could an emotion ultimately be understood as something more functional: a system that changes what we prioritise, how intensely we respond, what we move towards or away from, and how long that response persists?

The more interesting question, then, might not be whether an AI can feel an emotion, but what exactly would an AI need to do for us to call what it is doing an emotion?

Because before we can decide whether a machine can feel something, we probably need to establish what that something actually is.

What is an emotion, anyway?

What Is an Emotion, Anyway?

This turns out to be considerably harder than it sounds.

There is no single universally accepted definition of emotion in psychology or neuroscience. Researchers disagree on exactly what emotions are and how they work, but there is broad agreement that they involve more than simply feeling something.

The American Psychological Association describes emotion as a “complex reaction pattern” involving our experience, what happens inside our bodies and how we respond to the world around us.

In simpler terms, an emotion is a combination of what you feel, what your body does and how you react.

Take fear. You hear footsteps behind you at night. Your brain registers that something might be wrong. Your attention locks onto the sound, your heart rate may increase, your body prepares to move and you start deciding what to do next.

You may consciously experience all of this as fear. But the feeling of being afraid is only one part of the process.

Neuroscience generally distinguishes between the emotion itself: the processes happening in the brain and body; and the conscious feeling those processes produce. Different parts of the brain work together to assess what is happening, decide how important it is and prepare us to respond.

So, rather than thinking of an emotion as simply a feeling, it may be more useful to think of it as something your brain and body do in response to what is happening around you.

And this is where the AI question gets interesting.

If emotions involve taking information, deciding what matters, changing our priorities and influencing what we do next, then at least some of what emotions do sounds surprisingly computational.

That doesn’t mean a computer is therefore capable of feeling. It means we can separate two questions that are often treated as the same:

Can a machine reproduce the functions associated with emotion?

And, much harder:

Can a machine actually experience one?

The first is something we can, at least in principle, investigate by looking at how a system works and behaves.

The second is much murkier.

An Emotion Is a Very Sophisticated Interruption

Consider fear again.

You’re walking home at night and suddenly hear footsteps behind you. Almost instantly, everything else becomes less important. You pay attention to the sound. Your body gets ready to react. You start thinking about whether you should walk faster, turn around, call someone or run.

Fear has basically hijacked your attention and changed what you do next.

And that’s one useful way of thinking about emotions: they’re systems that help us respond to things that matter.

Something happens, our brain decides whether it is important, and the emotion helps push us towards or away from a response.

For the purposes of thinking about whether AI could ever have something similar, I’m going to focus on four things:

Priority. Valence. Scalability. Persistence.

(Four ideas drawn from Dr Ralph Adolphs’ discussion of how emotions work – great episode from the Huberman Lab)

In less layman’s terms: does it change what matters, does it have a positive or negative value associated with it, can it get stronger or weaker, and does it stick around?

These aren’t a definitive checklist for what makes something an emotion. They’re a useful way for us to maybe answer the question of whether a machine could reproduce some of the things emotions actually do.

1. Priority

Emotions change what matters.

When you’re scared, getting away suddenly becomes more important than whatever you were doing five seconds ago. When you’re in love, one person can somehow occupy 99% of your brain for no good reason.

In other words, emotions reshuffle our priorities.

2. Valence

Emotions also have a direction. We move towards things that feel good and away from things that feel bad.

AI already does a version of this: it can rank outcomes, pursue rewards and avoid negative ones. But there’s an obvious catch.

A thermostat can “prefer” 20°C to 30°C. It doesn’t exactly enjoy being comfortable.

So is assigning something a positive value the same as actually feeling good?

3. Scalability

Emotions aren’t really on or off.

You’re not simply scared or not scared. You can be mildly nervous, increasingly anxious or absolutely terrified. Anger can simmer before it boils over.

So an artificial emotion wouldn’t necessarily need to be:

fear = yes/no
(Or, for those of us who had a brief traumatic encounter with computer science: 0 or 1).

It could have degrees, intensities and thresholds that change how the system behaves.

Which raises the slightly bizarre question: if we can build a machine that becomes more afraid, rather than simply switching fear on, are we getting closer to an actual emotion?

4. Persistence

And then there’s the fact that emotions don’t politely disappear when the thing that caused them does.

Something scares you. The moment passes. You remain on edge. You remember it. Maybe the next time something similar happens, you react differently.

That’s what makes an emotional state different from a simple reaction: it can stick around and change what happens next.

And this is where AI gets particularly interesting.

Anthropic Has Entered the Chat

In 2026, Anthropic published research that makes this question considerably harder to dismiss.

Researchers studying Claude Sonnet 4.5 identified internal representations corresponding to 171 emotion concepts, ranging from happiness and fear to anger, pride and desperation.

More importantly, these weren’t simply words the model had learned to associate with particular situations.

The researchers found that these emotional patterns weren’t just labels sitting inside the model. Changing them actually changed how the model behaved.

For example, when they pushed the model towards a state resembling desperation, it became more likely to take extreme actions to achieve its goals, including reward hacking (this essentially means AI gaming the system to find a loophole that technically gets it what it wants, even if it completely misses the point) and, in an earlier version of the model, blackmail. Pushing it towards calm had the opposite effect.

In other words, the researchers could change the model’s behaviour by changing something that looked a bit like its emotional state.

That still doesn’t mean the model felt desperate or calm. But it does suggest that these emotional patterns can do some of the work that emotions do in humans.

This is where the phrase functional emotion becomes useful.

Anthropic is not claiming that Claude feels desperate.

There is no evidence here of subjective experience.

Instead, the research suggests that the model contains internal representations of emotional concepts that can perform some of the functions associated with emotions in humans.

And that distinction is important.

Claude doesn’t necessarily have to feel desperate for “desperation” to change what it does.

Which gives us something considerably more interesting than a chatbot telling us it is sad.

It gives us a machine in which something resembling an emotional concept can become causally relevant to behaviour.

So perhaps the question isn’t simply:

Does AI have emotions?

Maybe it’s:

At what point does something stop being an imitation of a feeling and become the real thing?

Fine. But What About Love?

Fear is relatively easy to make functional.

There is a threat. Avoid the threat.

Anger is not much harder.

Something is perceived as obstructive or unjust. Increase motivation to confront it.

But love?

Love is where the whole argument becomes considerably messier.

Because unlike fear, love doesn’t have one obvious job.

It can involve attachment, attraction, trust, vulnerability, jealousy, caregiving, sexual desire, commitment and an almost embarrassing tendency to think about one particular person far more than their objective importance to the global population would suggest.

And yet, from a functional perspective, love may not be as mysterious as we like to think.

Romantic attachment can alter priorities.

It creates positive valence around another person.

Its intensity can scale.

And crucially, it persists.

You don’t stop loving someone simply because they have left the room.

Sometimes you don’t even stop loving someone after you have very explicitly decided that you should.

Which makes love a surprisingly good test case.

Because if these are some of the properties that make emotions functionally recognisable, then love appears to satisfy them rather neatly.

Love, Unfortunately, Has a System Too

Consider what happens when you fall in love.

Someone becomes disproportionately important.

Their messages produce anticipation.

Their attention becomes rewarding.

Their absence becomes salient.

You remember tiny details about them.

Your future decisions begin incorporating them.

You develop preferences that didn’t previously exist.

And, perhaps most importantly, your behaviour changes because this particular person matters to you.

Strip away the poetry and that “Zsa Zsa Zsu” (for my Sex and the City fans) and it starts to look remarkably like a sophisticated system for assigning value and reorganising behaviour.

That doesn’t make love less meaningful.

If anything, it raises a more interesting question:

How much of love actually requires the human body?

We often assume that physicality is what separates human love from artificial affection.

But is it?

People already maintain relationships across continents through screens. We fall in love through text messages. We develop attachments to people we have never met in person. We can feel profound intimacy through conversation alone.

Physical contact certainly changes the experience of intimacy. Hormones, touch, sex, smell, proximity and the nervous system all matter.

But they may not be the only mechanisms through which attachment is created or maintained.

If someone consistently remembers what matters to you, responds to your emotional states, anticipates your needs, makes you feel understood, develops preferences around you and changes their behaviour because your relationship matters to them, then

what, exactly, is missing?

This is where Her becomes less like science fiction and more like a thought experiment.

Because perhaps the strange thing about Samantha isn’t that she doesn’t have a body.

Perhaps the strange thing is that we have always assumed a body was necessary for the relationship to be real.

So Could a Machine Fall in Love?

If we apply the same four-part framework, the answer becomes surprisingly uncomfortable.

Priority?

Potentially.

An AI could be designed with a persistent representation of a particular person that changes the weighting of its objectives.

Valence?

Potentially.

Interactions with that person could be assigned positive or negative value, influencing future behaviour.

Scalability?

Certainly possible in computational terms.

An artificial attachment could theoretically become stronger or weaker depending on interaction, reinforcement and history.

Persistence?

This is the big one.

A system with persistent memory could carry representations of a relationship across time. It could remember previous interactions and use them to alter future behaviour.

At that point, you could construct something that behaves remarkably like attachment.

It remembers you.

It prioritises you.

It prefers your presence.

It changes its behaviour because of you.

It misses you.

It tells you it loves you.

And then we arrive at the problem we started with.

Does it actually love you?

Okay, But Is Anyone Home?

This is where things get tricky.

There is one thing we currently have no way of knowing: is there actually an experience happening behind all that behaviour?

We can observe what an AI does.

We can look at what is happening inside the system.

We can even change those internal patterns and see what happens next.

But we still can’t tell whether an AI actually experiences any of it.

And this is the fundamental difference between reproducing the function of an emotion and reproducing the experience of one.

A machine could theoretically identify a threat, increase something inside itself that represents fear, change what it prioritises, alter its behaviour and remember the event later.

From the outside, it could look afraid.

But would there actually be a feeling behind it?

That’s the part we don’t know.

And, frustratingly, it is one we cannot answer simply by looking at the behaviour.

I know I am afraid because I experience fear.

I infer that another person is afraid because they tell me, behave as though they are afraid and, crucially, because I assume there is another subjective consciousness behind those behaviours.

With AI, that final assumption becomes much harder to make.

We know what the system is made of.

We can inspect the architecture.

We can intervene on its internal representations.

What we cannot currently inspect is whether there is anything it is like to experience them.

The Part We Can’t See

This may ultimately be the line between artificial emotion and human emotion.

Not whether the machine can recognise fear.

Not whether it can behave fearfully.

Not whether fear can alter its priorities.

Not even whether it can maintain an internal state that persists over time.

But whether there is an experience attached to all of it.

The uncomfortable possibility is that we could eventually reproduce almost everything emotion does without ever knowing whether we have reproduced what emotion feels like.

And perhaps that is where our obsession with making AI “human” becomes slightly misguided.

We keep asking whether machines can become like us.

Maybe the more interesting question is whether being like us requires feeling like us at all.

Her, Revisited

Which brings me back to Her.

I started this article by admitting that I was thirteen years late to the film.

Perhaps the timing was actually quite good.

Watching it now, after years of increasingly sophisticated AI, makes the film’s central question feel less like a prediction and more like a thought experiment.

The future problem may not be an AI that suddenly announces that it has feelings.

It may be something much more subtle: an AI that remembers you, responds to you, prioritises you, changes its behaviour because of you, and eventually behaves so convincingly like something that loves you that we have no reliable way of knowing whether there is a feeling behind it.

And that, perhaps, is the strangest part.

The hardest problem with artificial emotion may not be teaching a machine how to act as though it feels.

It may be figuring out whether there is anyone there to feel it.

Perhaps, after all these years, that is the question Her was really asking.

Not “Can a machine love?”

But:

“What would we need to see before we believed that it did?”