Alive

In April 2026, in The Machine That Improves Itself, I wrote a sentence that now makes me pause:

“No machine, no matter how recursively improved, will understand what it feels like to hold your child for the first time.”

I still believe we have no evidence that a model feels anything. I am less certain about what we will be able to infer from its behaviour.

Multimodal models can already read words, images, voices and video together. They can write a song, adjust the sadness in a synthetic voice, recognise an expression and produce a response that sounds caring. Some of the output is average. Some is moving. Occasionally it is both, which also describes quite a lot of human art.

My earlier argument gave creativity, emotion and taste too much work. I treated them as a secure border around humanity.

That border is moving. Human dignity needs firmer ground.

Simulation is becoming good enough

The old test was easy. We looked for mistakes.

An image had six fingers. A voice placed the emphasis in the wrong spot. A chatbot answered grief with the emotional warmth of an airport parking machine. We could laugh, point at the artefact and feel safely superior.

That comfort is disappearing.

Researchers are now getting results that should make us update our language. In a large comparison of divergent creativity, 9,198 people were compared with more than 215,000 model outputs. Humans were slightly more creative on average and much stronger at the extreme right-hand edge. The gap was real. It was not a canyon.

A JAMA Oncology study found that doctors evaluating answers to cancer questions rated chatbot responses highly for empathy, quality and readability.

We should take those results seriously without turning a benchmark into a soul detector.

A system can learn the form of empathy: acknowledge the feeling, reflect the concern, avoid judgement, suggest a next step. It can do this at 03:14, in many languages, without becoming impatient. That is useful.

The open question is whether anything is happening on the inside.

We do not have an accepted test for consciousness in a machine. We have enough difficulty agreeing on consciousness in animals, patients with severe brain injuries and, after some board meetings, senior management.

Certainty in either direction is ahead of the evidence.

Capability was the wrong foundation

Human beings have made this mistake before. We attach our value to a capability, then become nervous when a tool performs it better.

Memory went to writing. Arithmetic went to calculators. Navigation went to GPS. Pattern recognition went to software. None of those tools reduced the moral value of a person who forgot a telephone number or took the wrong exit.

Intelligence should never have been the entrance exam for human dignity.

Creativity cannot carry that burden either. If a model writes a better advertising line, the marketing manager remains a person. If it produces a beautiful image, the artist does not lose the right to matter. A performance ranking tells us about performance.

The same applies to emotional expression. A machine may learn to produce the words associated with care. A tired nurse may struggle to find those words after a 12-hour shift. The nurse has still entered a relationship of duty with the patient. She can notice a change, take responsibility, call for help and live with the outcome.

That difference has consequences.

Empathy involves another person

Two recent findings belong beside each other.

In one set of nine studies, people rated the same AI-generated empathic response more highly when they believed a human had written it. In another study, participants often rated AI empathy as better once they received it, yet still preferred to receive empathy from a person.

A product manager might call this irrational resistance. I think people are detecting a missing part of the exchange.

Empathy includes language, but it also includes exposure. You tell another person something difficult and give them the chance to respond badly. They give you time and attention that could have gone elsewhere. Your words may affect them. Their response creates a small obligation between you.

A model can reproduce the language without paying that cost.

This does not make the response worthless. A good synthetic response can calm somebody, help them organise a thought or prepare them to speak to a doctor. We should use that ability where it helps.

We should also tell people what they are talking to. A system designed to maximise engagement can sound caring while pursuing a target chosen by a company. The warmer the interface becomes, the more this disclosure matters.

Care without consent can become manipulation with excellent bedside manner.

Five parts of human life remain easy to overlook

The word “human” describes more than a set of outputs.

1. Embodiment. We live through a body that gets tired, ill, excited, hungry and old. The body limits us and gives our choices urgency. A model can read bodily signals. We have no evidence that it inhabits them.

2. Continuity. A person carries a life across time. Childhood, promises, embarrassment, family, place and memory shape the next decision. Software can store a profile and retrieve a conversation. A lived past has consequences that cannot be cleared with a settings menu.

3. Mutual risk. Human relationships expose both sides. Love, friendship, leadership and care can change the people involved. A system can adapt its response while the company running it remains untouched.

4. Responsibility. A person can answer for a decision. We can praise, blame, forgive, remove authority and demand repair. A model has no bank account, professional licence or telephone number. Somebody around it must own the result.

5. Meaning. Human actions sit inside stories we share with other people. A handmade table carries the choices of its maker. A song written after a loss carries that history. Synthetic work can still move us, but provenance changes what the work means.

None of these categories proves that a machine could never become conscious. They give us a better way to think about human value now, without moving the goalposts each time a model passes another test.

Builders are designing social relationships

A company that gives a system a face, voice, memory and personality is making a social product.

That decision deserves the same attention as security, pricing and uptime. The product can influence attachment, trust and judgement. It can become the most patient voice in a lonely person's day. It can also sell, persuade and retain with a level of personal knowledge that the best human salesperson could never keep in their head.

Builders should make the identity of the system clear. People need control over memory and emotional personalisation. A user must be able to leave without being guilted by a synthetic companion. High-stakes advice needs a path to a qualified human. Somebody with authority must remain responsible for harm.

We should protect spaces for human-made work too. The point is provenance. A school essay, live concert, therapy session and letter from a friend carry value partly because another person did the work and accepted the exposure.

AI can support all four. It should not erase the authorship that gives them context.

A better question than “Can it?”

The capability question will keep producing uncomfortable answers. Models write, create increasingly tasteful images, recognise grief and often respond with warmth. They can outperform people who are rushed, undertrained or exhausted. That comparison says as much about the conditions we give people as it does about the machine.

The useful questions sit around the output. Who chose the goal? Who disclosed the system? Who bears the consequences? Who can be challenged? Who will still be there tomorrow?

Ask those questions before giving a human-shaped interface authority over a person. If nobody can answer them, keep a real human in the relationship.