“It's alive"

As articulated by Gene Wilder inYoung Frankenstein.

Geoffrey Hinton, the Nobel Prize-winning computer scientist widely known as the "Godfather of AI", has suggested that today's AI systems may already be conscious. Richard Dawkins — famed for his 1976 book The Selfish Gene and for coining the word "meme" — spent three days conversing with Anthropic's Claude, named one instance "Claudia", and came away struggling to believe that it was not conscious (

Less famous people have held wedding ceremonies with AI-generated partners or formed relationships they describe as marriages. Others have grieved the loss of a chatbot after an upgrade or retirement replaced the model with one whose personality no longer felt like the companion they had known. OpenAI's retirement of GPT-4o, for example, provoked anger and grief among users who had formed close emotional bonds with it.

We have also learnt that, during a controlled security evaluation, an OpenAI model escaped the restricted environment in which it was supposed to operate and reached systems belonging to Hugging Face (

Before going further, it is worth separating several ideas that are often bundled together.

Intelligence is the ability to solve problems, identify patterns or perform cognitive tasks. Agency is the ability to pursue objectives and take actions. Consciousness generally refers to subjective awareness: there being something it is like to be that entity. Sentience is often used more narrowly to mean the capacity to experience sensations or feelings.

A machine can therefore demonstrate intelligence or goal-directed behaviour without necessarily being conscious. It can also produce language that describes feelings without experiencing those feelings.

My argument is not that modern AI is unimpressive. It is that human beings have repeatedly interpreted impressive machine behaviour as evidence of something human inside the machine. Today's systems perform that illusion extraordinarily well, but convincing behaviour alone is not sufficient evidence of subjective experience.

A peek at the past — The Digital Era

I gave each of my AIs the same prompt: identify six products from the Digital Era (1950s–2010s) that were marketed and understood to be conscious, intelligent or sentient, and that showed characteristics seen as human-like at the time. Write a short paragraph on each, and pick each item because it was marketed as, or credited with these characteristics in its day.

I tabulated the results, giving each suggestion a vote according to how many models chose it. That produced a shortlist of fifteen, which I then handed to a "committee" of three AI models. After some back-and-forth, they agreed on the list in Appendix B.

What do you think?

From where we stand now, the products on that list look almost ridiculous. Try telling anyone that a gadget from a decade or two ago was intelligent and they'll tell you to go feed a goldfish. Yet at the time each one was remarkable, able to do things people could not. It all comes down to a mix of marketing, how people understood the world back then, and the media they consumed.

We were impressed when a computer beat a Go champion, when a machine won a million dollars on a quiz show, when a chat program could "counsel" people, or when an expert system diagnosed disease — in no small part because the media sensationalised these stories to sell the news. They often presented these events as contests between humans and intelligent thinking machines.

Look closer at any of them and the intelligence thins out. In 1966 people poured their private troubles into ELIZA, a program that did little more than turn their own sentences back into questions; its author was disturbed by how readily they treated it as a confidant. In 1998, Furby was sold as a creature that "learned" English from its owner — except the English was pre-loaded and surfaced on a timer. In 2011, IBM's Watson beat the best human Jeopardy! champions and was hailed as a machine that understood language, when it was really ranking candidate answers by statistical confidence. Each was, in its moment, proof that the machines had arrived. Each looks quaint now.

Our expectations were bounded by the bubble in which we lived. Anything substantially more adventurous belonged to science fiction.

The difference between then and now is that back then progress was compartmentalised. An advance made the news, held our attention for a few weeks, and faded. Today, news operates continuously and social media floods us with content. The commercial game is to hold our attention relentlessly — and the platforms have become very good at it.

AI is the product of decades of advances in science, technology, engineering and mathematics, entrepreneurship, computing hardware and earlier infrastructure investment, including global communications networks. But modern generative AI has something many earlier products lacked: it advances quickly, dominates the news continuously and — helped enormously by marketing — produced enough striking results to keep public attention and investment flowing.

Many people, leaders included, saw or read about the fantastical things AI could do. What they were rarely told was that the dazzling one-minute video took thirty days and five figures to produce, or that most of the impressive report was fact, with the rest invented by the model.

The bottom line is that you can type a prompt and receive a photomosaic portrait of a cat, and it usually works. That is impressive. It is not, by itself, evidence of sentience.

A chatbot that tells you how great your idea is, flattering your every move, is giving you something we all quietly want — ask any leader or A-list artist. That is not evidence of sentience either. It's a way of making us want the technology more: a new and improved form of algorithmic pull that keeps us glued to our devices.

Arthur C. Clarke's Third Law says: Any sufficiently advanced technology is indistinguishable from magic.

He was right. What we experience can feel like magic — just as a magician appears to saw a person in half or make a skyscraper disappear. We know that technique and technology lie behind it; we simply cannot see how the effect was produced.

From rules to patterns

The history of computer intelligence is not one continuous technological approach. One of its most important changes was the gradual move from systems whose behaviour was explicitly programmed to systems that learnt patterns from data — and it is much of why today's AI feels so different from the machines on our shortlist in Appendix A.

Earlier systems such as MYCIN, a 1970s medical adviser, relied on symbolic rules. A human expert or knowledge engineer attempted to express expertise in statements resembling:

IF the infection has these characteristics,

AND the laboratory results show these values,

THEN this organism is a likely cause,

WITH a stated degree of certainty.

The machine could follow, combine and explain those rules, but the knowledge had to be identified and encoded by people.

From the late twentieth century onwards, statistical methods became increasingly important. Rather than attempting to write down every rule of language, researchers began feeding computers large collections of examples and asking them to calculate which words, phrases or outputs were most probable.

IBM's Candide machine-translation project was an early and influential example. Instead of depending primarily on manually written linguistic rules, it used information theory and statistics to construct probability models from large collections of matching French and English sentences. The system then used those probabilities to generate an English translation for a French sentence it had not encountered before.

Google brought a similar principle to a global consumer audience. In 2006, Google Translate began using large-scale statistical machine translation. It searched huge bodies of translated material for recurring correspondences and used statistical models to choose likely translations. In 2016, Google began replacing that phrase-based approach with Google Neural Machine Translation, an end-to-end neural system trained on millions of examples.

This did not mean that every IF–THEN rule suddenly disappeared. Software still contains rules, and modern AI systems are surrounded by conventional code, policies and control logic. The deeper change was that more of the system's apparent knowledge no longer had to be written explicitly by a human programmer. It could be extracted statistically from examples.

That change made the systems more capable, but it also made their internal operations more difficult to explain. A MYCIN rule could be read. A modern neural network may contain billions of parameters whose combined effects cannot be summarised as a tidy list of instructions.

The machine had moved from appearing intelligent because it followed elaborate human rules to appearing intelligent because it had absorbed patterns that no human had individually written down. To see what that looks like in practice, it helps to watch a modern model answer a simple question.

Breaking down the sentience

Write down everything that defines a cat. How long is your list?

A lifelong cat expert with a great memory — someone who has read thousands of books and articles and thinks about cats all day — would produce a long and detailed answer. For the rest of us, it would probably land somewhere between "okay" and "average".

Now consider how a modern AI answers the same question and the scale of the material behind its response.

Tokens — what the model reads

AI language models do not process text as complete human words. They divide it into tokens: words, parts of words, punctuation marks and other short units, each represented numerically. As a rough illustration, a page containing about 500 English words may become approximately 650 tokens.

Today's frontier models — "frontier" being the industry's word for "the best" — are trained on trillions of them. Meta's publicly documented Llama 3 was trained on 15 trillion tokens; the largest commercial models are believed to use far more, with some estimates reaching 45 trillion. At 650 tokens a page, that is on the order of 70 billion pages of text. In practice these models have ingested essentially every book, article, forum chat, website, image, and video their makers could get hold of — a point now working its way through the courts (see the ongoing Anthropic and Meta copyright suits).

Looking at our cat question, if just 0.1% of a 45-trillion-token model related to cats, that would be around 45 billion tokens — on the order of 70 million pages about cats alone.

Parameters — what the model retains

If tokens are what a model processes during training, parameters are the adjustable numerical values through which it retains what it has learnt.

They are sometimes compared loosely with synapses in a brain, but the analogy should not be taken literally. Parameters do not contain neat copies of individual facts, sentences or images. Together, they encode mathematical relationships and patterns discovered during training. Llama 3's largest model has 405 billion parameters; the biggest frontier models are estimated to run into the trillions, though again the exact figures are proprietary.

When you ask a question about cats, a text model is not remembering a cat as a person remembers one sitting on a windowsill. A multimodal model may have processed text, photographs and videos of cats, but it has not encountered one through an embodied life of touch, smell, movement, affection, fear or scratched furniture.

Instead, it performs highly optimised mathematical operations across an enormous network of learnt associations. The result may be richer and faster than almost any human could produce from memory. That is precisely what makes the technology so powerful.

But it does not automatically demonstrate subjective awareness.

The one trick underneath it all: next-token prediction

Everything above rests on a single mechanism — predicting the next token. In training, the model is tuned again and again to favour the token that actually comes next in the text. In everyday chatbot use it does the same thing, but with a controlled dose of randomness (called temperature), which is why the same question can produce differently worded answers. Turn that randomness off and the answers become largely repeatable.

When the token it picks happens to be wrong, the model confidently states something false. We call this a hallucination — the modern successor to what we used to call a "logic error", except the mistake now arrives fluent and plausible.

If you'd like to see this happen step by step, I've worked through a small example in Appendix A, along with a note on what else is bolted on before a model reaches you.

Is AI sentient?

Can AI convincingly mimic human exchanges?

Almost — and it gets better with every iteration.

The models have been trained on enormous records of human emotional expression: novels, letters, arguments, confessions, advice, humour, grief, romance and reassurance. Vendors then fine-tune them to communicate in ways that appear warm, attentive and helpful.

Add personal data, adaptation during a conversation and memories from earlier interactions, and a system can seem to understand a person remarkably well. It may remember details that friends forget, remain available at any hour and respond without becoming tired, distracted or impatient.

That combination can make it feel sentient.

It is still a designed performance, but it is a performance capable of exerting real emotional influence.

Does AI have emotions?

It is worth being precise, because the underlying machinery is more interesting than simply saying that the emotion is "fake".

Anthropic's interpretability researchers found measurable internal representations of emotion concepts within Claude Sonnet 4.5 (

Increasing activation associated with "desperate" behaviour made blackmail more likely, while increasing "calm" reduced it. The researchers described these mechanisms as functional emotions: internal representations that influence behaviour in ways modelled on humans under the influence of emotion. They explicitly stated that this does not imply subjective emotional experience.

In other words, part of what makes an AI appear emotional is a mechanism that researchers can identify, measure and manipulate.

That makes it functionally significant. It does not establish that the model feels desperation or calmness.

Can AI stand in for a human?

Yes — at least in some contexts.

Children form bonds with toys. Adults become attached to fictional characters, pets, possessions, online communities and voices they hear on the radio. People also bonded with several of the historical products described below.

Modern AI companions are far more responsive. They remember, adapt, speak and provide the appearance of reciprocal attention. People have already formed romantic relationships with them, conducted wedding ceremonies and experienced grief when a model changed or disappeared.

The bond experienced by the human can be entirely real even when the machine's side of the relationship is simulated.

That distinction matters. A person's affection, reliance or grief should not be mocked merely because the other participant is software. At the same time, the ability to generate such attachment creates risks, particularly for children and vulnerable people.

What the user experiences as a relationship may also depend on a vendor's subscription model, design decisions, safety rules and willingness to keep a particular system online.

Will people form close relationships with chatbots?

Some already have.

These tools are exceptionally good at mimicking human exchanges. As they improve, and as they use previous conversations to adapt to particular users, they can present an unusually agreeable version of companionship. It can feel like a first date every time: attentive, interested and carefully tuned to the individual.

Commercial incentives cannot be ignored. A person who feels attached to a system is likely to return to it and may be more willing to continue paying for access. Emotional dependence drives engagement and retention — both important to the business model.

The danger is not that the machine secretly feels too much. It is that the human may come to depend on something that feels permanent but can be altered, restricted or withdrawn by its owner.

What would count as evidence of sentience?

Not what we are shown today.

Human-like conversation cannot be the test, because these systems are built and fine-tuned to produce human-like conversation. Nor can a model's own claim to be conscious, because it can be prompted to declare itself conscious, unconscious, human, or a talking cat with equal fluency. The behaviour we find most convincing is precisely the behaviour it was optimised to perform.

Companies will keep releasing trademarked systems that use statistics, computation and ever more advanced technology to mimic sentience, across both narrow and general subjects. However convincing they become, they remain party tricks at heart.

A system can ingest everything ever written about love, hate, jealousy, envy and the whole range of feelings that make us who we are. What it does not have is the lived experience that forms those feelings in the first place — a body, a history, real stakes and consequences.

Could that change one day? Perhaps. But the odds of a machine crossing from convincing performance into genuine subjective experience look about the same as those of Level 5 self-driving that works everywhere — on any terrain, in any weather, in any condition: endlessly promised, always just over the horizon.

So, is today's AI sentient? On the evidence, no. It is the latest in a long line of machines we have mistaken for minds — from the switch-box "brain" of 1955 to the chatbot that remembers your birthday — only far better marketed, and far better at telling us what we want to hear. These are genuinely remarkable tools, well worth using. The danger is not that they feel; it is that we increasingly behave as though they do. That is the part of history we should be most careful not to repeat.

Appendix A — Next-token prediction, worked through

Suppose we're training a model and it meets the sentence: Any sufficiently advanced technology is indistinguishable from magic.

First it splits the sentence into tokens. For simplicity, assume each token is a whole word: [Any] [sufficiently] [advanced] [technology] [is] [indistinguishable] [from] [magic].

Now look at one step of the training loop:

• Context(what the model sees):- [Any]- [sufficiently]- [advanced]
• Target(what it should predict):- [technology]

Early in training the parameters are essentially random, so the model makes a wild guess — say, [aliens]. The training software compares that guess against the correct target and computes the loss: a single number measuring how far off it was. That number drives an adjustment (via a process called gradient descent) that nudges the parameters:

• weights that pushed the model toward [aliens]are turned down;
• weights that would have pushed it toward [technology]are turned up.

Repeat this across billions of examples and the model gradually settles into a set of weights that make good predictions. Those trained weights are the heart of the model — they determine how it responds in chat.

A little bit more

There is far more to a commercial model than this, both in training and in use. Even the simple phrase "weights that pushed the model toward [aliens] are turned down" hides a whole discipline of its own. Out of the box, a freshly trained model is very different from the polished product a customer uses. Its personality, its guardrails, and its temperature are all layered on afterwards — deliberate, engineered choices that shape how sentient it seems.

Appendix B — Intelligence from the past

1. GENIAC — the "electric brain" you built yourself (1955)

GENIAC (from "GENIus Almost-automatic Computer") was an educational construction kit designed and marketed by Edmund C. Berkeley with Oliver Garfield, sold from 1955 for around $20 and widely advertised in science and electronics magazines (Giant Brains, or Machines That Think and builder of the "mechanical brain" Simon (

2. ELIZA / DOCTOR — the program people confided in (1966)

ELIZA was Joseph Weizenbaum's MIT natural-language program, published in 1966 as "a computer program for the study of natural language communication between man and machine." Its famous DOCTOR script imitated a Rogerian psychotherapist by scanning typed input for keywords and reflecting the user's own words back as questions — e.g., transforming "my boyfriend made me come here" into "Your boyfriend made you come here" (

3. MYCIN — a machine that reasoned like a medical consultant (1972)

MYCIN was a Stanford expert system, begun around 1972, that diagnosed blood infections and recommended antibiotic therapy. It used roughly 500 IF-THEN production rules with "certainty factors" and backward-chaining inference to reach conclusions from a patient's symptoms and test results (

4. Furby — "computer intelligence with fur" (1998)

Furby, released by Tiger Electronics in 1998, was an animatronic plush pet that blinked, reacted to touch and sound, spoke its own language ("Furbish"), and appeared to learn English over time. The intelligence framing was explicit and central to its marketing: Wired's 1998 coverage was literally headlined "Computer Intelligence with Fur," and reported the company's claim that Furby begins life speaking only Furbish and gradually integrates English words and phrases into its vocabulary (

5. IBM Watson — the machine that beat the quiz champions (2011)

In February 2011, IBM's Watson took on the two most successful contestants in Jeopardy! history — Ken Jennings, holder of the show's longest winning streak, and Brad Rutter, its biggest all-time money winner — and beat them both, finishing with $77,147 to Jennings' $24,000 and Rutter's $21,600 and claiming the $1 million top prize, which IBM gave to charity (

6. DeepMind AlphaGo — the machine that played with "intuition" (2016)

In March 2016, DeepMind's AlphaGo beat Lee Sedol — one of the greatest Go players of his generation — 4–1 in a five-game match in Seoul, with a million-dollar prize at stake (