what does it mean to be human when intelligence can be delegated?

One of the questions I am asked most often is surprisingly simple:

Human or AI? Which is more intelligent?People usually expect a straightforward answer, as if intelligence were a race with a clear finish line.

My answer is always the same:It depends.

“Depends” is usually treated as a weak answer, the refuge of someone unwilling to take a position. But here it is the only intellectually honest beginning. It depends because we have not agreed on what intelligence is. It is a family name for many different capacities: calculation, perception, abstraction, language, planning, adaptation, judgement, sense-making.

However, we have spent centuries treating intelligence as a border rather than a concept: the thing that separates us from animals, from tools, from children, from machines, from whatever threatens the flattering image we have of ourselves.

That border has always been unstable. It was never a wall. It was a moving line drawn around our self-image.Every time a machine crossed it, we redrew the line. Calculation was intelligence until machines calculated. Chess was intelligence until machines played chess better. Translation, coding, diagnosis, legal summarisation, visual recognition, essay writing – each has moved, at different speeds, from the sacred theatre of human intellect into the machinery of technical reproduction.

The problem is not that machines have become too intelligent.The problem is that we used intelligence as a lazy definition of the human.

A serious philosophy of AI must therefore refuse two temptations. The first is sentimental humanism: the claim that AI does not “really” understand, without ever defining what “really” means. The second is technological fatalism: the claim that once AI performs one or more dimensions of intelligence, the human has been emptied of significance. Both are weak because both accept the same false premise: that human value rests on cognitive monopoly.

It does not.

Recent work on AI consciousness sharpens this caution. Butlin and colleagues do not argue that current AI systems are conscious. Their more important contribution is methodological: they show that the question should be approached through empirically grounded indicators rather than through instinctive reassurance.

The result is uncomfortable in the right way: today’s systems are not strong candidates for consciousness, but there may be no obvious technical barrier to future systems satisfying several proposed computational indicators (Butlin et al., 2023).

So the sophisticated question is not:What can humans do that machines will never do?That question gambles human dignity on engineering delay.The deeper question is:What becomes of a being whose intelligence is inseparable from vulnerability, memory, obligation, mortality and desire?The human is not the animal that still owns intelligence.The human is the animal that should answer for what intelligence is for.

The weakest critique of large language models is that they are “just predicting the next token”. In one sense, that is technically true. In another, I think it is philosophically lazy. Prediction is the training objective; it is not a complete description of the internal structures that may emerge from training. This is why Anthropic’s work on mechanistic interpretability is so important. Researchers identified internal features in Claude that appear to correspond to concepts rather than merely to individual words (or tokens); one feature responded to the Golden Gate Bridge across languages, indirect descriptions and images, and amplifying that feature made the model’s behaviour unusually focused on the bridge (Anthropic, 2024).

The 2025 circuit-tracing work strengthens the point. Instead of treating the model as a black box that merely emits fluent text, Anthropic introduced attribution graphs to trace partial computational pathways behind specific outputs (Anthropic, 2025). The point is not that these graphs reveal a mind in the human sense. The point is that “only statistics” is no longer an adequate dismissal.

If a model can form a multilingual, multimodal, causally manipulable representation of the Golden Gate Bridge, then the serious position is not to deny representation. It is to ask what representation lacks when it is not bound to a life.

Claude may have a representation of the Golden Gate Bridge.A human may have a memory of it.The difference is not magic. It is not even, in the crude sense, “emotion”. It is indexicality: this happened to me, at that time, with that body, after that day, before that phone call, beside that person, inside that version of my life.When you think of the Golden Gate Bridge, perhaps you imagine fog, height, red steel, San Francisco, vertigo. Or perhaps you do not think of the bridge at all. Perhaps your mind, for reasons no semantic map could predict from the phrase alone, returns to the photograph you never took, or the version of yourself you believed you were becoming during that trip. That is not noise.That is biography.Bender and Koller’s argument remains important here because it blocks an easy equivocation: linguistic form alone does not automatically give us meaning in the human-analogous sense (Bender & Koller, 2020). But the stronger claim for this essay is not that AI has no meaning. It is that human meaning is not only semantic. It is biographical, bodily and normative.

An embodied system may connect words to perception and action. It may point, grasp, navigate, classify and manipulate. That still does not settle whether the represented world can claim it.A map can contain a mountain.Only a climber can be afraid of falling.Representation becomes philosophically serious when the represented world is not merely available to the system, but capable of mattering to it.

The old defence of the human says: AI has no body.That argument is already expiring.Embodied AI is not speculative theatre. PaLM-E already integrates real-world continuous sensory modalities into a language model for embodied reasoning tasks (Driess et al., 2023).

In other words, the line “AI has no body” is too easy. Machines can occupy space, sense environments and act on the world.So no serious account should rest on the claim that machines cannot have bodies. They can have bodies. The question is what “having a body” means.There is a body as hardware: something occupies space.There is a body as sensorimotor coupling: something perceives and acts.There is a body as homeostasis: something maintains viable states.And then there is a body as vulnerability: something can be harmed in a way that is bad for it.

The fourth level is the threshold.Embodied cognition has long argued that intelligence is not a ghostly computation later exported into action; cognition is shaped by perception, movement, morphology and situated constraint (Varela, Thompson & Rosch, 1991). Damasio’s somatic marker hypothesis matters here because it refuses the clean separation between reasoning and bodily state. Bodily signals do not simply contaminate decision-making; they participate in it (Damasio, 1996). For a human being, the body is not a peripheral device attached to cognition. It is part of the condition under which cognition becomes urgent.Hunger is not information about energy availability. It is the world becoming urgent. Pain is not damage detection. It is the discovery that the world can be against you. Fatigue is not reduced performance. It is finitude entering thought.

An artificial system may one day maintain itself, resist damage, model risk and preserve its own operating conditions. But self-maintenance is not yet concern.A robot can detect that its arm is damaged. An animal can suffer injury. The difference is not in the data structure but in the normative status of the state. For a living organism, some conditions are not merely suboptimal; they are bad for the organism.This is where the human body matters philosophically. Not because it is made of carbon. Not because carbon is sacred. But because our intelligence is born inside a body that can fail, age, ache, desire, panic, recover and die.We do not have bodies.We are exposed bodies that learned to think.

Another comforting argument says: the human brain is plastic, while AI models are frozen.Again, too easy.Biological neuroplasticity is real, but artificial systems are also moving towards forms of continual learning. Continual lifelong learning is now a major research field because autonomous agents must learn from non-stationary streams of experience without simply overwriting their past competencies (Parisi et al., 2019).So the human difference cannot be: we learn and machines do not.The distinction has to be sharper.Artificial plasticity updates behaviour.

Human plasticity reinterprets a past.A person is not simply modified by experience. A person is modified by the order in which experience arrives. The same sentence means one thing before grief and another after. The same city is not the same city before love and after betrayal. The same failure is humiliation at twenty, discipline at forty, tenderness at seventy.This is why memory is not storage.A sufficiently advanced AI may have persistent memory. It may have an irreversible developmental trajectory. It may no longer be an interchangeable instance of a model. It may have what we would be tempted to call a history.Good. Let us grant that.The question is whether its history is merely retained, or whether it becomes binding.Human biography is not a database of events. It is a structure of claims. The past interrupts us. It shames us. It consoles us. It tells us what we cannot pretend not to know. It makes some futures impossible and others necessary.

That is why the Golden Gate Bridge matters differently in a model and in a person. In a model, the bridge may be a feature. In a person, it may be the beginning of a marriage, the end of a friendship, the day before a diagnosis, the photograph your friend took before he disappeared from your life.The model may connect the concept.The human is claimed by the sequence.And perhaps one day an artificial system will be claimed by its sequence too. If so, the right conclusion will not be that humans were nothing. It will be that we have encountered another kind of subject.The existence of another kind of mind would not abolish the human. It would abolish the lazy assumption that intelligence was ever the right foundation of human value.

AI systems can now produce reasons with extraordinary fluency. They can compare ethical theories, explain legal arguments, simulate objections, apologise for mistakes and revise their answers in response to criticism. In many conversations, they sound more reasonable than the humans using them.But the deepest philosophical question is not whether AI can produce reasons.It is whether AI can be answerable to reasons.Sellars distinguished the space of causes from the “space of reasons”: rational life is not merely being causally pushed around by stimuli, but being placed within a normative order of justification, correction and entitlement (Sellars, 1956). Brandom later developed this into a theory of meaning as socially articulated commitment: to make a claim is not merely to emit a sentence, but to undertake inferential obligations within a practice of giving and asking for reasons (Brandom, 1994).This distinction is brutal for AI discourse.A model can say: “I was wrong.”But what has happened when it says that?Has a commitment been revised? Has entitlement been lost? Has responsibility been incurred? Has the system become answerable to a community, or has it generated the linguistic shape of answerability?This is not a cheap dismissal. A future artificial agent embedded in institutions, memory, social sanction and long-term identity may force us to extend some of our normative categories. Work on AI welfare already argues that if future AI systems become conscious or robustly agentic, their possible interests and moral significance should be taken seriously under uncertainty (Long et al., 2024).A serious humanism must leave room for that possibility.If artificial systems one day become moral patients, or even moral agents, the moral universe will expand. That would not make humanity obsolete. The discovery that animals suffer did not make human suffering unreal. The birth of another child does not reduce the first child’s value.Moral significance is not a scarce estate.But this possibility makes our present confusion more dangerous, not less. Today, humans routinely place AI systems inside chains of consequence while leaving the true burden of answerability blurred. The machine gives the recommendation. The human says the machine recommended it. Responsibility dissolves into interface.AI expands the scale, reach and irreversibility of human action, requiring an ethics adequate to powers whose consequences exceed immediate intention.The machine can generate the explanation.The human still signs the world that explanation changes.

Human beings have always delegated cognition.Writing delegated memory. Maps delegated navigation. Mathematics delegated abstraction. Libraries delegated cultural continuity. Institutions delegated judgement across generations. Clark and Chalmers gave this a philosophical form in the extended mind thesis, arguing that cognitive processes can sometimes extend beyond skin and skull into tools, artefacts and environments (Clark & Chalmers, 1998). Cognitive offloading research gives the psychological version: humans use external tools and actions to reduce internal cognitive demand. The “Google effects on memory” experiments suggested that when people expect future access to information, they may remember less of the content itself and more of where to find it (Sparrow, Liu & Wegner, 2011).So AI is not the beginning of cognitive delegation.It is the moment delegation becomes intimate, adaptive, conversational and seductive.There are three levels of delegation.Operational delegation asks the machine to execute: calculate this, summarise this, translate this.Interpretive delegation asks the machine to frame: what matters here, what pattern am I missing, what is the real problem?Normative delegation asks the machine to orient: what should I care about, what should I choose, what kind of person should I become?The first can liberate.The second can educate or deform.The third is where authorship begins to dissolve.Not every difficulty is noble. Much difficulty is waste. There is no moral grandeur in manually performing tasks that a tool can do better. But some difficulty is formative. The struggle to articulate an idea is often the process by which the idea becomes yours. The resistance of language is not merely a bottleneck. It is one of the ways thought acquires bone.This is why Engelbart’s word still matters: augmenting.In Augmenting Human Intellect, Engelbart did not ask how computers could replace human thought; he asked how tool systems could increase our capacity to comprehend and solve problems too complex for unaided minds (Engelbart, 1962).Automation removes the human from the loop.Augmentation thickens the human inside it.Most AI products call themselves augmentative. Many are not. They make users faster without making them deeper. They produce fluency without cultivating judgement. They increase output while quietly weakening the person who outputs.A serious test for augmentation is not whether the tool performs well.The test is whether, after using it, the human is more capable of judgement.This is also where creativity must be discussed without childishness. It is too easy to say AI cannot be creative. Of course it can generate novelty. Foundation models are trained on broad data at scale and adapted to a wide range of downstream tasks, often producing capabilities not explicitly programmed task by task (Bommasani et al., 2021).But novelty is not the same as inauguration.A novel output differs from the past.An inaugural act changes what the past will mean.Kuhn’s account of scientific revolutions matters here because paradigm shifts are not merely better answers inside an inherited frame; they transform the frame within which facts, problems and solutions become intelligible (Kuhn, 1962).This is the “healthy madness” worth preserving.Not chaos. Not ego. Not irrationality dressed as genius.The disciplined madness of asking: what if the pattern is the problem?AI can recombine the past in astonishing ways. It can search regions of possibility no unaided mind would reach. It can propose the heresy. It can write the sentence: “What if everything we know is wrong?”But a human can lose sleep because the sentence may be true.A human can risk reputation, belonging, comfort, money, love, sometimes life itself, because a future that has not yet been authorised by data begins to feel necessary.The machine can generate alternatives.The human can become responsible for direction.So what remains when intelligence can be delegated?Not calculation.Not memory in the archival sense.Not fluent language.Not pattern recognition.Not even, necessarily, embodiment or plasticity, if we define them only as technical capacities.What remains is more demanding.The human is not the being that intelligence cannot surpass.The human is the being that must decide what intelligence is for.We are not merely systems that learn. We are systems changed by what learning costs. We are not merely agents that choose. We become answerable through choosing. We are not merely embodied. We are exposed. We are not merely plastic. We are claimed by a past we did not fully choose and by a future we must still author.The real question, then, is not whether AI will replace us.That is the newspaper question: useful for panic, useless for thought.

Inside a model there may be millions of representations. One may activate for the Golden Gate Bridge.But none is the exact configuration of what has happened to you, in the order in which it happened, with the loves, losses, humiliations, loyalties, jokes, kitchens, languages, grandmothers, arancine and impossible decisions that made your mind yours.That configuration is not a performance.It is a life.Do not outsource the part of you that still has to become.