What is Intelligence | The Electrification of Matter

When a child asks whether the light is thinking

Imagine a child called Noor sitting at a dining table while her aunt Mei prepares a lesson. There is a smooth stone beside a notebook, a reading lamp above them and a laptop waiting for a password. These are fictional people in an ordinary room. Nothing about the question they are about to ask is fictional.

Mei presses the switch. The lamp comes on.

“It responded,” Noor says. “Does that mean it knows?”

We could settle the conversation quickly by telling her that a lamp is not intelligent. Most of us would be comfortable with that answer. But Noor has asked a better question than the answer acknowledges. Why does the lamp fail to qualify? It received something. Its state changed. It produced a useful result. If responding to the world makes something intelligent, the lamp seems to have a reasonable application.

Mei points to the stone. It, too, responds. Warm it in your hand and its temperature changes. Put it on the edge of a sloping surface and it may move. Matter is not generally unresponsive until intelligence arrives. The physical world was busy producing consequences long before anyone began explaining them.

The laptop complicates the afternoon. It can multiply numbers, sort photographs and produce paragraphs. Noor can ask it a question in a way she cannot ask the stone. Yet a paragraph may contain an error, and a calculator can outperform her at arithmetic without knowing why she is upset about tomorrow’s examination. One thing is extraordinarily capable at a narrow task. Another appears broadly conversational. Neither observation, by itself, tells us everything we want to know.

Then Mei asks Noor why the plant beside the window bends towards the light. Noor pauses, remembers a lesson, reaches for a word and changes her first explanation. The change is small, but it is revealing. She does not merely produce an output. She notices a difficulty in her own account and tries to repair it.

The room now contains several kinds of responsiveness. The stone changes physically. The lamp follows a switching arrangement. The computer performs organised operations. Noor can reconsider what a question means, search her experience, ask for help and attempt an explanation. Calling all four “intelligent” without further distinctions would make the word very generous and not very useful.

Calling only the child intelligent would also leave work undone. What, precisely, is she doing that matters? Would a person unable to speak show no intelligence? Would a bird solving an unfamiliar problem fail because it could not describe its reasoning? Would a machine’s correct answer become irrelevant simply because its internal method differed from ours?

The answer requires a journey through mechanisms and evidence, not a vote on which objects feel most like us. The existing How Intelligence Works begins with a noticed distinction and follows it towards organised thought. Here we begin one step closer to the material world. Before asking how large a mind’s map can become, we ask what makes anything capable of using a map at all.

What the title can honestly mean

“The electrification of matter” is an arresting phrase because it seems to describe a threshold. On one side lies ordinary material. On the other lies something that can recognise a face, imagine a bridge or wonder what intelligence means. It is tempting to place a spark between the two and let the spark carry the explanation.

It cannot carry that much.

Electric charge, electromagnetic interactions and the movement of charged particles belong to the physical world quite independently of thought. A lightning discharge is an electrical event of immense power. That power does not provide evidence that the lightning understands the landscape. A wire conducting current is not thereby considering where it should go. We have identified a physical process, not established a cognitive capacity.

In a nervous system, electrical signalling occurs within a highly organised living system. Cells maintain conditions across their membranes; signals interact with other signals; connections have particular properties; the organism has a developmental and evolutionary history. In a computer, electronic components participate in an engineered architecture that carries out operations. Organisation makes the activity capable of doing particular work. Merely increasing the electricity does not supply the missing organisation.

The distinction resembles the difference between ink and a sentence. Ink makes many printed sentences physically possible. More ink does not necessarily produce a better sentence, or a sentence at all. Meaning depends on arrangements, conventions, contexts and readers. The analogy is limited, but it catches the error: a material ingredient should not be promoted into a complete explanation of the capacity it helps support.

We should also resist the suggestion that matter starts without physical structure and receives intelligence from somewhere outside nature. A living brain is material throughout its operation. Explaining its activity through biology and physics does not diminish the reality of learning or love. It changes the level at which we are asking the question. A description of the paper fibres in a letter and a description of the promise written on it can both be true.

The eduKate account of physics is a useful companion here because it asks for models that connect quantities, mechanisms and observations. Biology adds the organisation of living systems, including metabolism, development and inheritance. A serious account of intelligence needs these neighbouring explanations without pretending that one short phrase has completed them.

So the title will remain a metaphor with a physical foothold. It points towards matter arranged so that signals can contribute to perception, learning and flexible action. It does not claim that every electrical event is a thought, that a thought is simply a unit of electricity, or that every possible intelligent system must use the same mechanism as a human brain.

This leaves the wonder intact. The remarkable thing is not that an unexplained spark has entered the room. It is that some arrangements of the world’s ordinary constituents can become capable of investigating the room, correcting an explanation of it and teaching that explanation to someone else.

Four questions hidden inside one question

When Noor asks what intelligence is, she might be asking four different questions. They overlap, but answering one does not automatically answer the others.

The first concerns the material mechanism. How do cells, molecules, circuits or other physical processes make the relevant activity possible? Here we ask about signals, connections, energy, memory and the operations a system can implement. A good answer explains how something happens in a particular kind of system.

The second concerns the capacity. What can the system do? Can it distinguish situations, infer a missing relationship, learn a useful regularity, solve an unfamiliar problem or revise a failing response? Here we describe intelligence functionally. Two systems can share a capacity while implementing it differently, just as different instruments can measure the same distance through different physical arrangements.

The third concerns evidence. How do we know the capacity is present? A correct answer may arise from understanding, memorisation, a hidden helper, a fortunate guess or a shortcut the examiner did not notice. The measurement problem is therefore more demanding than collecting an impressive example. We must decide which alternative explanations our evidence can exclude.

The fourth concerns experience. Is there something it is like to be the system? Does it feel, notice its own existence or undergo a subjective experience? This is the question of consciousness. Intelligent performance and subjective experience are related topics in the study of minds, but they are not interchangeable definitions. A result on a problem-solving task does not, on its own, settle the experience question.

Imagine that someone opens the laptop and explains its electronic components beautifully. Noor may learn a great deal about how the device works. She still does not know whether its translation is accurate. Now imagine that someone shows a flawless translation. She has evidence of a successful output, but has not yet learned whether the system can translate unfamiliar dialects or recognise when a sentence is ambiguous. Each answer leaves a different question open.

Confusion often begins when the speaker changes questions without announcing the change. “It is only computation” is offered as if a physical implementation automatically disproved a capacity. “It solved the puzzle” is offered as if performance automatically established consciousness. “Its brain is small” is offered as if a size measurement settled everything about behaviour.

Exactness begins with refusing these substitutions. The eduKate guide to evidence gives this conversation its discipline: ask what a piece of evidence supports, and how far that support can travel. We can investigate physical mechanisms, recognise useful capacities, design demanding tests and remain uncertain about experience at the same time. Those positions are compatible. Keeping them separate makes the conversation more productive, and makes the next question possible: what definition of intelligence can we actually use?

A definition we can actually use

For this essay, intelligence is the capacity to acquire and use information to learn, infer, solve problems and adapt effectively when relevant conditions change. The definition is deliberately a working one. It gathers a family of capacities rather than declaring that every intelligent act must display every member of the family.

A person can reason about a new problem without visibly learning during that particular minute. A trained system can use knowledge acquired earlier even when its stored parameters are no longer changing. An animal can show a specialised competence supported partly by inherited organisation. Requiring a complete new learning episode every time intelligence appears would misdescribe all three.

The important word is capacity. We are interested in what the system can reliably do across appropriate opportunities, not only what happened once. “Acquire and use” also matters. Possessing a large collection of records is different from bringing relevant information to bear on a situation. “Effectively” asks whether the response serves the task or problem being considered. “Relevant conditions” prevents us from demanding impossible success across every conceivable world.

These choices fit a broad tradition without resolving every disagreement within it. The American Psychological Association’s overview of intelligence includes learning from experience, reasoning and adaptation. In machine intelligence, Legg and Hutter developed a formal account centred on an agent’s performance across environments. Chollet later emphasised how efficiently a system acquires skills, given its prior knowledge and experience. These are different research approaches, not three versions of one uncontested measuring instrument. See Legg and Hutter and Chollet.

What does this definition help us notice in Noor? Suppose she has learned to divide a rectangular sheet into equal parts. A differently coloured rectangle should not destroy her method. A triangle should make her reconsider which parts of the old method still apply. If she simply repeats the last procedure because the teacher has not supplied another, we have learned something about the limits of her current understanding.

The definition also makes room for different profiles. Someone may be quick to detect visual relationships and slow to interpret a difficult paragraph. Another may organise a team well and struggle with an unfamiliar symbolic notation. A single umbrella term can describe these capacities without requiring them to rise and fall together in every person or task.

There is an important restraint here. We are not defining intelligence as success at any cost. A system may pursue a specified objective effectively while causing harm. Whether the objective is worth pursuing is a further question of values and judgement. Nor are we defining intelligence as constant novelty. Using a familiar method when it fits can be exactly the intelligent response.

This is why How Intelligence Works connects noticing, organised knowledge, correction and shared capability. The new definition gives us a way to examine those connections. It does not turn them into a compulsory checklist through which every living creature, machine and human being must pass in the same way.

The wire, the thermostat and the chess player

Mei changes the question. Instead of asking Noor which object has intelligence, she asks what evidence would persuade her that an object had more flexible competence than it first appeared to have.

The wire conducts. Its behaviour depends on physical conditions, but it does not follow that it is interpreting those conditions as a problem to solve. The lamp produces light when its circuit permits it. The thermostat adds a more interesting arrangement: it senses a temperature and participates in a feedback process that helps regulate a target. The response changes according to a measured difference.

That is a genuine functional advance. It is still a narrow one. A basic thermostat does not need to understand comfort, discover why a room is losing heat or decide whether a different target is appropriate. Calling it a simple controller tells us more than treating the word “smart” on a box as an explanation.

The chess program offers another advance. It can select among alternatives in a structured problem space and, depending on its design, use learned evaluations, search or other methods to guide its choices. That can justify a claim about chess competence. It does not establish that the program understands a tenancy agreement or can teach Noor fractions.

There is no need to turn this table into a ladder of moral importance. It compares claims about capacities. It does not rank the worth of lives, and it does not imply that every lower row contains all the capacities of the rows above it.

Some researchers use intelligence very broadly and may describe elementary control as a minimal case. Others reserve the term for richer learning and flexible problem-solving. The boundary depends partly on the purpose of the definition. We can acknowledge that disagreement while remaining precise about the thermostat’s mechanism and limitations.

This is more useful than arguing over a badge. If a manufacturer says that a device learns, ask what changes through experience. If it says the device adapts, ask which changes in the environment it can handle. If it says the device reasons, ask what relationships it can use and which counterexamples expose its limits. Each claim should arrive with an appropriate burden of evidence.

Noor now sees why movement was too easy a criterion. A falling stone moves very efficiently. She also sees why complexity was too easy. A tangled arrangement can be complicated without solving anything. Intelligence is not established by activity, complexity or impressive speed alone. We need a connection between the information available, the problem faced and the competence demonstrated under conditions that make the demonstration meaningful.

How matter carries a signal

The next step is physical, and it deserves more care than a picture of a brain with lightning drawn across it. A signal is a physical change that can make a difference elsewhere in a system. It may be a change in voltage, a pattern of light, a concentration of molecules or another detectable variation. Whether it functions as a signal depends on the arrangement in which it occurs.

In many neurons, an action potential is a brief change in voltage across the cell membrane that propagates along an axon. Ion channels and differences in ion concentration are central to this process. Inputs can increase or decrease the likelihood that a neuron will produce an action potential. The Queensland Brain Institute’s explanation of action potentials provides an accessible account of these events.

Communication between neurons is often chemical as well as electrical. At a chemical synapse, activity in one neuron can lead to the release of neurotransmitter, which acts on receptors associated with another cell. The effect depends on the relevant receptors and cellular conditions; “a chemical signal” is not a synonym for a simple instruction to fire. The neuroscience account of chemical synapses explains the machinery involved.

There are also electrical synapses, where current can pass between coupled cells through gap junctions. Their existence is another reason not to draw all neural communication as one universal sequence. The nervous system contains multiple mechanisms and many interacting scales. See the corresponding account of electrical synapses.

Notice what these descriptions have and have not done. They explain some of the mechanisms by which activity is transmitted and combined. They do not tell us that a particular spike means “umbrella,” that every neuron stores one thought, or that adding spikes adds intelligence in a fixed proportion. To explain a capacity, we need the organisation and activity of the relevant system, together with its relation to a task and environment.

The comparison with computing is useful only if we preserve the differences. Electronic circuits can implement logical and numerical operations through controlled physical states. A biological neuron is not simply a transistor with a different appearance. An artificial neural network’s numerical units are mathematical components of a model; the name does not make them miniature biological cells.

We can nevertheless see a shared requirement. Useful distinctions must be made physically available to whatever process will use them. A light-sensitive component may respond differently to different illumination. A sensor reading may enter a controller. A spoken question may initiate activity in a nervous system. If all relevant differences are erased before they reach the next stage, the system cannot use those differences to choose a response.

The physical story therefore gives us a necessary discipline: follow what can affect what, through which medium, under which conditions. But the presence of a signal is still an opening, not a complete mind. Noor’s next question is sharper. “How does a change in something become information about something?” That is where the room begins to connect with representation.

How a signal becomes about something

A mark on the page is physically present whether Noor understands it or not. Its usefulness changes when it enters a relationship she can use. The numeral 4 might indicate four objects, the fourth floor, an answer to a calculation or a bus route. Recognising the shape is not enough to determine which interpretation is appropriate.

Information becomes useful to an intelligent system through such relationships. A detectable difference can help distinguish possible states of the world. A representation preserves some of those differences in a form the system can use. A map retains relationships among locations while leaving out most of the appearance, sound and smell of the places represented.

Noor draws the room. Her drawing is imperfect, but it puts the window beside the plant and the table beneath the lamp. If Mei asks where to place a second chair without blocking the doorway, the drawing might help them decide. Its value is not that it resembles every part of the room. Its value is that it preserves something relevant to their problem.

Now Noor rotates the page. If she confuses the top of the paper with a fixed direction in the room, her representation can mislead her. A useful map requires more than marks. It requires a relation between the marks, their interpretation and the world in which they will be used. The estate’s How Information Works and How Models Work provide two routes into that distinction.

We should not assume that every intelligent system needs a little picture inside it. A representation may be distributed across many changing states; some successful control can be explained through close interaction with the environment without positing a detailed internal map. The drawing is an example that makes the relationship visible, not a claim about the format of every thought.

Nor should we equate technical measures of information with understanding. A string can contain many unpredictable symbols without giving Noor a usable explanation of anything. A short formula can be highly useful because she knows how its terms connect to the task. Length, surprise and semantic usefulness are different properties. Treating them as one quantity would lose the very distinctions we are trying to preserve.

The practical question is whether the information can change an interpretation or response in an appropriate way. If the door moves in the drawing but remains fixed in the real room, a good planner should not blindly trust the drawing. If the room changes but the drawing does not, the representation needs revision. An intelligent system must somehow remain answerable to what its representations are supposed to help it do.

This gives us a first form of exactness: relevance. The system need not record everything. It must retain enough of what matters. That requirement is demanding because what matters depends on the question. A floor plan may be excellent for moving furniture and useless for identifying a gas leak. Intelligence includes finding out which description the present problem needs.

Memory is a changed possibility

Mei erases the drawing. Noor can still point to where she put the doorway. Something has survived the disappearance of the marks. We call that memory, but the word can make the process sound more uniform than it is.

A notebook preserves marks outside the person. A computer can retain data in a storage medium. A nervous system can change through experience in ways that influence later activity. These are different mechanisms of persistence. They become comparable at a functional level when a past event makes a relevant difference to a later response.

The biological story is not that an electrical impulse circles forever, carrying a complete memory like a train carrying a passenger. Research on learning and memory includes changes in synaptic function and, in some forms of longer-lasting change, molecular and structural processes. Eric Kandel’s work on the sea slug Aplysia helped establish mechanisms linking learning to changes in synaptic transmission. His Nobel lecture on memory storage is a primary account of that research programme.

That research should not be stretched into the claim that every human memory is a single modified connection. Remembering a person, executing a practised movement and holding a number briefly while calculating are not interchangeable jobs. A short essay cannot turn one important mechanism into a complete account of all of them.

For Noor, the educational question is often more immediate. Can the previous lesson change what she can do today? She may recognise a solution while looking at it and be unable to reconstruct it when the page is closed. She may recall a formula but fail to recognise the question that needs it. She may remember the method clearly after a hint but not know how to begin independently.

These are different access conditions, not proof that nothing was learned. They matter because a claim about usable knowledge should specify when the knowledge is available. How Memory Works in Education follows that practical distinction between what has been encountered and what can later be retrieved and used.

Memory supports intelligence because the system does not have to begin again from the entire uncertainty of its first encounter. Yet memory can also preserve a mistake. If Noor repeatedly applies an inappropriate rule, its familiarity may make it feel more secure. Persistence alone is therefore not progress. We need to ask what survives, how it is accessed and whether it remains open to correction.

There is a future-facing side as well. Noor intends to bring the notebook to school tomorrow. Remembering the intention at the right time is different from recalling its existence after arriving without it. The estate’s study of prospective memory examines this bridge between a present intention and a later opportunity to act.

Memory, then, is not merely the storage room behind intelligence. It changes the possibilities available to the next moment. The important question is whether the right part of the past can become useful when the future finally arrives.

The narrow doorway of attention

The room has more information than Noor can use at once. There is the sentence she is reading, the sound of traffic, a movement at the window, the position of her pencil and the remembered worry about tomorrow. The problem is not simply a shortage of input. It is the need to select and organise input while doing a particular job.

Attention is one name for the processes involved in that selection. In the language of everyday experience, something becomes foreground while other things recede. The Attention Gate article makes this selection a central part of the intelligence story. Its gate is a useful metaphor, provided we do not mistake it for one literal doorway in the brain.

Mei gives Noor a word problem. It contains two prices, a discount, an irrelevant description of the shop and a question about the final amount paid. Noor can read every word correctly and still give the wrong detail too much importance. Good performance requires identifying which relationships constrain the answer.

Now Mei changes the task. Instead of asking for the price, she asks whether the description makes the shop sound welcoming. The formerly irrelevant wording becomes relevant evidence. Attention cannot be judged independently of the task. A detail that distracts from one question may answer another.

The selected pieces also have to remain available long enough to be combined. In a multistep problem, Noor may need to preserve an intermediate result while deciding the next operation. The study of working memory concerns this active use of information. It should not be reduced to a fixed number of slots that applies identically to every person, material and circumstance.

Writing down the intermediate result can help. The paper takes over part of the burden of persistence, allowing Noor to attend to the relationship she is trying to understand. The improvement does not make her unaided memory larger by definition. It changes the arrangement through which the task is being performed. That distinction will matter when we compare people, tools and teams later.

Selection also contains a value judgement. When several demands compete, which one deserves time? Checking the unit in an answer may be more useful than making the handwriting decorative. Revisiting a misunderstood instruction may be more useful than completing five more questions under the same misunderstanding. The estate’s discussion of prioritisation asks how limited resources should be allocated among competing claims.

This is why intelligence cannot sensibly mean maximum processing of everything. A system that treats every input as equally urgent may be less effective than one that makes a careful selection. The difficulty lies in choosing without becoming blind. Noor must concentrate enough to solve the present problem while remaining able to notice that she has misunderstood it.

The next advance is therefore not merely a brighter spotlight. It is a better relationship between selection and revision: attend to what appears relevant, use it, and remain capable of changing the selection when the world gives a reason.

Pattern recognition and its impostors

Mei writes 2, 4, 6 and leaves a space. Noor supplies 8. It is a reasonable answer to a familiar kind of classroom question. It is not a logically unavoidable consequence of those three numbers alone.

Many rules can agree with a short sequence and disagree about what follows. Perhaps the numbers describe an increasing sequence of even numbers. Perhaps they are labels selected from a much larger list. Without assumptions about the task, the observations do not uniquely determine the continuation. Noor has made an inference using both the visible pattern and her knowledge of the situation.

This does not make pattern recognition worthless. It makes its conditions visible. Intelligence often has to act before every ambiguity has been removed. The aim is to find useful regularities while retaining enough awareness of uncertainty to avoid treating a plausible pattern as an established law.

The Pattern Recognition study follows the discovery of recurring structure. The neighbouring Rule Induction study asks what happens when observed cases suggest a more general rule. The difference matters: noticing repetition and justifying a rule that extends beyond the observations are related but distinct achievements.

Suppose Noor notices that she performs badly on a worksheet whenever she uses a blue pen. She might decide that blue pens cause poor thinking. Mei asks what else changes on those days. Perhaps the blue pen is the one she keeps in a bag used for a difficult afternoon class. Perhaps she remembers the disappointing blue-pen days more vividly than the successful ones. Perhaps there are too few examples to support any conclusion.

The point is not that children make mistakes adults have escaped. A grown-up can create an equally fragile theory about a market, a colleague or a machine. An apparent pattern becomes especially seductive when it explains an anxiety or confirms a story we already want to believe.

A stronger approach asks what observations would distinguish the candidate explanations. Does the pattern survive new cases? Does it remain when an irrelevant feature changes? Does it predict something we have not already used to invent the rule? Can a simpler explanation account for the result? These questions turn recognition into investigation.

They also make mistakes informative. If Noor predicts 8 and the next item is 10, she does not need to conclude that reasoning is useless. She needs to discover which assumption failed. The sequence might have been constructed differently; the question might have omitted essential context; the displayed item might itself contain an error.

Intelligence is visible in how the system responds to this mismatch. A brittle system repeats its preferred pattern regardless. A more flexible system searches for a better account while preserving what remains valid. The first success showed that Noor could see a regularity. The surprise begins to reveal whether she can examine it.

Concepts make the world portable

Noor has seen many chairs. They differ in colour, size, material and design. Some have arms; some fold; some are built for a child. The word “chair” allows her to treat these varied encounters as related without carrying a separate full description of each one into every conversation.

A concept makes some similarities usable and some differences secondary. That is a considerable economy. If every object were entirely new, experience would travel poorly. If every object were treated as identical, experience would travel carelessly. Useful concepts help a system preserve distinctions at the scale the problem requires.

This is not always a matter of finding one visible feature shared by every example. Everyday categories can depend on function, context, history and overlapping similarities. Formal concepts may have sharper definitions. A mathematical square can be defined through precise relationships; the boundary of “a good place to sit” depends on the person and situation.

The estate’s Concept Formation article is useful because it keeps the boundary in view. A category should help us handle new cases, including cases that reveal weaknesses in the category. The related account of discrimination asks which differences matter to the task and which can safely be ignored.

Mei gives Noor two containers. Both look large. One is wide and shallow; the other is narrow and tall. If the question concerns fitting a long ruler, the relevant relationship differs from the relationship needed to compare liquid capacity. “Bigger” has become too vague. Noor needs a more exact concept for the particular comparison.

This is one reason vocabulary can help thought. A useful word can stabilise a distinction and make it easier to recall, discuss and refine. But the word is not a magic package containing understanding. A learner can say “volume” confidently while still comparing containers only by height. We need evidence that the distinction can guide interpretation and action.

Abstraction takes the portability further. Noor can recognise a relationship across objects that look quite different. Sharing twelve counters among three groups and dividing twelve minutes among three activities may use the same numerical structure. The Abstraction study examines the selection of structure that can survive changes in surface detail.

There is a cost to this economy. Every abstraction omits something. Twelve minutes cannot always be divided among tasks as freely as twelve counters can be placed in piles; some tasks have setup time or cannot be interrupted. Transfer becomes intelligent when the preserved relationship fits the new problem and the omitted details do not secretly determine the outcome.

The companion piece on cognitive compression gives this balance a useful form. We want a manageable structure that retains the distinctions needed for use. A beautifully simple idea that erases the decisive difference is not a triumph of intelligence. It is a compact way to be wrong.

Reasoning keeps a relationship intact

Noor now has information, memories and concepts. Reasoning concerns what she can do with the relationships among them. Can a conclusion be supported by the premises? Can a proposed explanation account for the observations? Can she detect a step that changes the meaning of the problem without permission?

Mei offers a simple rule: if the afternoon activity is cancelled, the teacher will send a message. A message arrives. Noor concludes that the activity has been cancelled. That conclusion might be true, but it does not follow from the rule alone. The teacher might send a message for another reason. The rule promised a message after cancellation; it did not say cancellation was the only possible cause of a message.

This is a small example of a large difficulty. The mind can recognise familiar ingredients while reversing the direction of a relationship. A plausible story then hides the invalid step. Exact reasoning asks what actually follows, not merely whether the conclusion sounds compatible with the situation.

How Reasoning Works follows the route from premises to conclusions. Its relevance extends well beyond formal logic. When a person interprets a contract, evaluates a scientific claim or reads a graph, a relationship must remain intact through the movement from evidence to judgement.

Different forms of reasoning have different standards. A valid deductive argument preserves truth from true premises to its conclusion. An inductive inference extends beyond the observations and therefore carries uncertainty. An explanatory inference compares accounts of why something happened. Confusing these standards can make an uncertain conclusion look certain or make a useful probabilistic judgement look defective simply because it is not a proof.

Consider a damaged plant. Dry soil may support an explanation involving insufficient water, but it may not uniquely establish the whole cause of the damage. A careful reasoner can say that an explanation is plausible while remaining open to another factor. That qualification is not evasiveness. It states the strength of the relationship between evidence and conclusion.

Reasoning also makes checking possible. If Noor writes only a final answer, Mei can see whether it matches the expected result. If Noor shows the intermediate relationships, Mei may discover that a correct answer was produced by two errors that happened to cancel. The visible route provides evidence the result alone concealed.

This does not mean that every intelligent system must explain itself in fluent language. A nonverbal task can reveal sensitivity to relationships, and a fluent explanation can itself be misleading. The form of evidence must suit the claim. For a learner, showing working is often useful because it makes a particular process inspectable, not because sentences are the only medium of intelligence.

The principle is simple enough to carry into the next section. A relationship should survive the journey from what is given to what is concluded. When it changes, the change needs a reason. Intelligence becomes more dependable when it can notice that requirement before an attractive answer persuades it to stop looking.

Intelligence asks what would change the result

Noor has noticed that the room becomes brighter when Mei opens the curtain. She has a pattern. To understand the relationship more deeply, she can ask what would happen if the curtain stayed closed, if the sky became dark or if the lamp were the main source of light instead.

These questions move beyond describing what occurred together. They concern dependence: which changes would make a difference to the result, under which conditions? The estate’s Causal Reasoning study examines this movement from association towards explanations of how outcomes are produced.

The distinction has practical consequences. Suppose a student begins attending a new class and improves during the following month. The improvement matters, but the timing alone does not establish how much the class caused it. The student may also have practised differently, recovered from illness or received help elsewhere. A useful explanation has to consider plausible alternatives and the evidence that could separate them.

We do not need to pretend that every everyday decision can wait for a perfectly controlled experiment. We do need to distinguish a reasonable working explanation from a conclusion stronger than the evidence permits. Where a decision matters, improving the comparison can be more valuable than making the story sound more confident.

Counterfactual thinking gives this investigation a special reach. We imagine changing one relevant condition while keeping the rest of the assumed situation sufficiently stable. If the student had received the same practice without the new class, what might have happened? We cannot directly observe both histories for the same student at the same time. That limitation is part of the problem, not something imagination removes.

The Counterfactual Simulation article makes imagined alternatives useful while keeping them connected to assumptions. A simulated future can help compare options, but its output inherits the strengths and weaknesses of the model that produced it. A vivid imagined outcome is not automatically a likely one.

Mei asks Noor to plan where the class should display its projects. Noor imagines the visitors arriving, a queue forming and someone trying to pass through the doorway. She can revise the arrangement before moving the tables. Here intelligence gains a practical advantage: some consequences can be considered before they are incurred.

The advantage grows when Noor asks a better question. “Will this work?” is broad. “Can two people pass while someone is reading this display?” identifies a specific constraint. The study of question generation belongs here because an intelligently chosen question can change the efficiency of the entire search.

The room has not become a laboratory in which certainty is guaranteed. It has become a place where assumptions can be made visible and tested against consequences. That is enough to transform the conversation. Noor is no longer merely reporting that things happened together. She is beginning to ask what makes the difference, what remains uncertain and which observation would most improve her account.

A model has to earn its place

Mei asks how long it will take to reach the exhibition. Noor names a distance. The number is relevant, but it does not answer the question. Travel time also depends on the route, the mode of travel and conditions along the way. She has selected a description that leaves out part of the mechanism.

This is a common failure of capable thinking. The calculation inside a model may be correct while the model is inappropriate for the job. A neat answer can therefore conceal an earlier mistake in framing. The estate’s Model Selection article addresses this choice of a workable representation before action.

Suppose Noor treats the journey as distance divided by a constant speed. That may be a useful first approximation for some purposes. It may be poor for a trip with a long transfer wait. Adding more decimal places to the division will not repair the missing wait. Numerical precision and explanatory adequacy are different achievements.

A model earns its place by preserving enough of the relevant structure to support its intended use. It should also make its assumptions visible enough to reveal when the use becomes inappropriate. A rough sketch can outperform a detailed table if the sketch shows the blocked doorway and the table records only furniture dimensions.

This does not require every model to contain everything. Completeness would often defeat usability. Instead, intelligence has to manage a changing balance: enough detail to protect the important relationship, enough simplicity to make the relationship usable within available time and resources.

Now consider confidence. Noor says the journey will take exactly twenty minutes because the arithmetic produced twenty. Mei asks how uncertain the waiting time is. Noor realises that the certainty of the calculation has been borrowed for the uncertainty of the world. She can improve the answer by giving a range or identifying the condition under which twenty minutes is plausible.

The studies of calibration and uncertainty connect directly to this moment. Confidence should track the evidence and the limits of the model. “I am uncertain” becomes useful when it identifies what is uncertain and how that uncertainty affects the decision.

Evidence must also be weighted. Three people repeating the same unverified message do not necessarily provide three independent confirmations. A recent direct observation may answer a question that an older general description cannot. The Evidence Weighting article keeps these differences from disappearing inside a simple count of supporting statements.

The intelligent response may be to continue with a rough estimate, obtain one missing observation or change the model altogether. There is no single ritual that fits every case. The achievement is a disciplined fit between representation, evidence, confidence and use. Exactness lives in that fit. It does not require pretending that every useful answer can be exact to the minute.

Learning is a change that can travel

By the next lesson, Noor can solve the example she practised. Mei is pleased, but she does not stop there. She changes the numbers, then the wording, then the order in which the information is presented. Finally, she gives a problem in which the familiar method is tempting but inappropriate.

These changes ask different questions. Can Noor execute the procedure? Can she recognise the relationship under a different surface? Can she select the method without a direct cue? Can she withhold it when the necessary conditions are absent? A collection of correct answers becomes more informative when the differences among the tasks are deliberate.

The existing Mechanics of Learning provides a route from exposure towards usable capability. The focused article on Transfer and Recomposition asks how knowledge remains useful when the problem changes. This essay borrows that question, not a promise that all learning automatically transfers everywhere.

Learning research gives a reason for caution. The National Academies’ account of the How People Learn studies emphasises the importance of understanding underlying principles and recognising their application across contexts. It also treats transfer as something that depends on how knowledge is organised and learned. The relevant discussion appears in How People Learn II.

Our example can make the practical issue concrete. If every division problem appears immediately after a heading that says “Division,” the heading does some of the selection work. Remove the heading and mix problem types, and a different part of the learner’s capability becomes visible. A lower score under the new conditions may reveal a missing selection skill rather than the disappearance of all previous learning.

There is also a question of distance. Changing the colour of a diagram is a smaller departure than applying a relationship from a worksheet to an unfamiliar physical situation. Neither “transfer happened” nor “transfer failed” is complete without a description of the change. What survived, across which difference, with how much help?

Noor’s error can be useful if the next lesson repairs the relevant weakness. Repeating the identical explanation may help when she missed a step. It may not help when she can execute every step but cannot decide when the procedure applies. The response should match the failure being observed.

Learning also includes becoming less dependent on a particular support. Mei may first model a method, then provide a partial prompt, then ask Noor to choose and check it independently. The point is not to remove every tool. It is to know which part of the competence now belongs to the learner and which part still depends on the learning arrangement.

This is one of the clearest ways intelligence becomes visible in education. Experience changes what the learner can do, and the change remains useful beyond one rehearsed encounter. Yet the result is always bounded. Noor has gained a capability with a certain reach. The next intelligent step is to discover that reach accurately and extend it where doing so matters.

The same answer can hide different capacities

Imagine that Noor and another learner both answer a question correctly. Noor has seen the exact question ten times. The other learner has encountered the underlying idea but has never seen this arrangement. The two answers may deserve the same mark on that question. They do not provide identical evidence about how the result was achieved.

This distinction is especially important when evaluating artificial intelligence. A system may have encountered a benchmark, close variants of its questions or extensive task-specific training. A high score still describes something real about the tested performance. Interpreting it as evidence of unfamiliar problem-solving requires knowing more about exposure and preparation.

Chollet’s On the Measure of Intelligence argues that skill alone is insufficient for measuring intelligence because knowledge, experience and training can strongly influence the skill displayed. His proposed approach foregrounds skill acquisition and generalisation under specified prior conditions. It is a research proposal with particular assumptions, not a universally adopted replacement for every existing test.

We can make the measurement problem visible with invented numbers. Suppose two systems face a familiar set and a changed set, each containing one hundred questions. They receive the same allowed tools and time within each set. These are illustrative results, not findings about actual products or students.

If we care only about the familiar set, A has the higher observed score. If we care about the changed set, B has the higher observed score. A broad declaration that one system is “more intelligent” would require more than selecting the column that supports our preferred conclusion.

We would want to know how the changed set was constructed. Did it preserve the underlying task while altering irrelevant details, or did it introduce a genuinely new skill? Were the two systems’ prior opportunities comparable? Are these results stable across further samples? Do some mistakes matter more than others? Were the allowed tools actually used in the same way?

The questions do not make evaluation impossible. They make it interpretable. A benchmark is a lens constructed for a purpose. It can reveal a useful difference while leaving other differences unmeasured. The danger begins when the purpose is forgotten and the score is treated as a property without conditions.

The same caution applies to Noor. We should not dismiss a practised skill because it is familiar. Fluency is useful. We should also avoid using fluency on rehearsed material as the sole evidence for broad reasoning. Good teaching needs both a dependable repertoire and the capacity to select, adapt and extend it.

The exactness we want is therefore not “one answer, one quantity of intelligence.” It is an account of the task, the preparation, the resources, the performance and the change the performance survived. That account is less convenient than a single badge, but it is much closer to the phenomenon we are trying to understand.

How exact can measurement become?

Exact arithmetic can describe an inexact interpretation. That sentence is worth keeping beside any intelligence score.

Suppose an invented question-answering system receives one hundred questions. It answers ninety, gets eighty-one of those right and declines ten. Its accuracy among answered questions is 81 divided by 90, or 90 per cent. Its coverage is 90 per cent. It correctly answers 81 per cent of all questions presented. Every one of those calculations is correct, and each describes something different.

A second system answers all one hundred questions and gets eighty-four right. Its accuracy and coverage are 84 per cent and 100 per cent respectively. Which system is preferable? The arithmetic cannot answer until the use and the cost of mistakes are specified.

If every correct answer earns one point and everything else earns zero, the second system scores eighty-four and the first eighty-one. If a correct answer earns one point, an incorrect answer loses two and a declined question scores zero, the first scores sixty-three and the second fifty-two. The ranking changes because the evaluation question changes.

These invented scoring rules are not recommendations for examinations or real deployments. They demonstrate why a metric must be connected to a purpose. A system that refuses every difficult case can make its accuracy look impressive while being unhelpful. A system that answers everything may hide an unacceptable number of consequential mistakes inside a high volume of useful work.

A meaningful evaluation therefore specifies at least the task, the input conditions, the allowed support, the success criteria and the resources available. It should examine variation across relevant kinds of cases, rather than allowing an average to conceal a serious weakness. Where a sample is used to infer future performance, sampling uncertainty matters too.

This is the measurement counterpart of How Assessment Works. An assessment collects evidence to support a decision. If the decision concerns independent reading, extensive reading assistance changes what the result can establish. If the decision concerns productive use of tools, banning every tool may remove the very capability being evaluated.

Formal theories can make their quantities mathematically exact within stated assumptions. Legg and Hutter’s universal intelligence proposal, for example, specifies a theoretical measure over environments, while also involving computational limitations that prevent it from functioning as an ordinary direct test of every real system. Mathematical definition and practical measurement are distinct achievements.

For an actual learner, animal or machine, we usually need a profile: what succeeds, where it fails, what it costs, what support it uses and how it responds to change. Exactness means specifying these conditions honestly. It does not mean that one number has finally extracted an essence from the material world.

IQ is a useful instrument, not a measure of human worth

Noor has heard people call someone “high IQ” as though the phrase settled a person’s entire place in the world. It does not. It refers, properly, to a particular tradition of assessing cognitive performance through standardised instruments and comparison with an appropriate reference population.

Well-constructed cognitive tests can provide useful evidence. Performance across different cognitive tasks often shows positive relationships, and psychometric research examines both broad and more specific abilities. The APA task-force report Intelligence: Knowns and Unknowns describes these patterns and the predictive usefulness, as well as the conceptual limits, of test scores. It is a historical foundation from 1996, not a claim to summarise every subsequent study.

The reason to resist overstatement is not that measurement has no value. It is that an instrument acquires its value through the question it can answer. A cognitive assessment may help identify a pattern of strengths and difficulties. A school examination measures performance against its own curriculum and task demands. The two should not be casually treated as interchangeable, and neither is a complete biography.

Imagine two descriptions of Noor. The first says that she scored poorly on a task requiring rapid manipulation of unfamiliar symbols. The second says that she is incapable of understanding difficult ideas. The first may be a faithful report of an observation. The second travels much further. To justify that journey, we would need evidence the first statement does not contain.

Now imagine the reverse. Noor receives a very strong result and begins assuming that her first impression must be correct in every disagreement. The score has become a shield against correction. Whatever the strength it measured, using it this way makes her thinking less dependable in the situation before her.

A humane account of intelligence can acknowledge differences in cognitive ability without converting those differences into a hierarchy of dignity. A person’s right to care, respect and participation does not depend on winning a reasoning contest. Someone who requires substantial support remains a person whose experiences and interests matter.

This is also an educational distinction. A teacher needs an accurate description of the learner’s present work in order to teach well. Labels that are either falsely flattering or unnecessarily final can interfere with that description. “Show me where the problem became difficult” often opens more useful work than “You are clever” or “You are not clever.”

The next task is concrete: identify what the learner can do, what support changes the result, which difficulty is obstructing progress and what improvement can actually be demonstrated. None of this requires promising that everyone will develop the same profile. It requires taking the individual seriously enough to investigate rather than assume.

Intelligence can be studied without being worshipped. Scores can inform decisions without becoming identities. When those boundaries are preserved, measurement can serve learning instead of closing the conversation at precisely the point where a careful teacher should begin asking questions.

Creativity finds a possibility that can survive contact

Noor wants the exhibition display to do something unexpected. She suggests arranging the projects as a maze. The idea is new to the group, and for a moment that is enough to make everyone excited. Then Mei asks how someone using a mobility aid would move through it, how visitors would find the exit and whether the labels would remain readable.

The questions do not extinguish creativity. They give the idea a world in which it has to work.

Novelty alone is easy to produce. A random combination of objects can be unusual. A creative contribution usually requires some further relation to a purpose: an illuminating explanation, a useful design, an expressive work or a solution that opens a worthwhile possibility. The criteria differ across domains, but surprise by itself does not settle the question.

Intelligence contributes to creativity by making relationships available for recombination and by evaluating what those combinations accomplish. Noor might borrow an idea from a walking trail: several possible routes, clear landmarks and places to pause. She is not copying a forest into the classroom. She is transferring a pattern of navigation into a different setting.

The estate’s account of analogy helps explain both the power and the risk. A useful analogy preserves a relationship that matters. It becomes misleading when resemblance in one respect is used to claim equivalence in all respects. A classroom can borrow the idea of a trail without acquiring the physical space or ecology of a woodland.

There is also a search problem. The first workable idea may not be the best available idea, but searching indefinitely has a cost. Noor has a deadline and limited materials. The study of Search and Exploration asks how a system discovers useful possibilities beyond the ones it already knows.

Exploration becomes more effective when alternatives are meaningfully different. Producing ten minor variations of the same blocked layout may be less useful than considering three distinct arrangements. Constraints can help here. If the group must keep the doorway clear and allow visitors to read at different heights, those requirements can direct invention rather than merely restrict it.

Evaluation then returns the imagined possibility to the world. The group sketches the arrangement, walks through it and notices that a display board hides the next sign. Noor changes the position. The idea becomes less dramatic in one respect and more usable in another. That revision is part of the creative achievement.

The Synthesis article belongs at this stage: several partial ideas have to become one coherent working arrangement without erasing the constraints each contributed. A beautiful entrance, readable labels and accessible movement are not three independent successes if they cannot coexist.

Creativity therefore enlarges our account of intelligence. The system is not confined to selecting from a fixed list of responses. It can help construct a better list. But the same discipline remains: what is being proposed, which relationships support it, how will it be tested and what would make revision necessary?

Why stopping can be an intelligent action

Noor begins answering the next question before Mei finishes reading it. The opening words resemble a familiar problem, and the familiar method arrives quickly. Then the final sentence changes the task. Noor has to interrupt a response that was already under way.

Speed is useful only in relation to the right work. An immediate answer to the wrong question is not improved by arriving even sooner. The estate’s study of inhibitory control examines the ability to withhold or interrupt a response when it no longer fits the goal or situation.

Stopping is not the same as becoming inactive. It may be the action that makes a better action possible. Noor pauses, rereads the condition and discovers that the question asks for what remains, not what was used. The pause protects a distinction that haste would have erased.

There is an opposite problem as well. A learner can reconsider so often that no approach receives enough time to work. Flexible thinking requires a balance between persistence and revision. We need some stability to pursue a difficult task and some responsiveness to abandon a route that is demonstrably failing.

The Cognitive Flexibility study concerns changing the frame or strategy while preserving the problem that made the change necessary. The neighbouring account of goal maintenance concerns keeping the relevant objective available through interruption and delay. These capacities can support one another: remember what matters, then change the method when the method stops serving it.

Mei asks Noor to prepare an explanation for younger children. Noor spends twenty minutes choosing a decorative heading and has not yet decided what the children need to understand. The activity is related to the project, but it has displaced its main purpose. Recovering the goal is an intelligent correction, even though nothing visible has broken.

This is also a form of monitoring. What am I trying to do? Is the current activity helping? Do I understand this step, or does it merely look familiar? The estate calls this territory metacognition: the monitoring and regulation of thinking. It does not imply that a mind has perfect access to all of its own processes.

Self-explanations can be mistaken. Confidence can be poorly calibrated. Someone can describe a careful procedure and fail to follow it. Monitoring therefore needs contact with evidence: a worked example, a checkable result, feedback from another person or a test that exposes the claimed understanding.

For machines, an analogous functional question can be asked without attributing human self-awareness. Can the system detect a relevant uncertainty, stop an inappropriate operation, seek missing information or route the problem to a more suitable process? Such behaviours may improve practical reliability. They do not by themselves establish a conscious inner observer.

Intelligence is often imagined as an engine that produces more. This section adds another possibility. Sometimes the most useful contribution is a well-placed interruption that prevents the wrong process from consuming the rest of the available time.

The mind is a body in a place

Noor’s intelligence is not encountered as a collection of answers floating in empty space. She has eyes and ears, hands that move objects, a body that occupies a position and a history of acting in particular surroundings. The questions available to her are partly shaped by that situation.

Consider fitting the display boards into the room. Noor can calculate dimensions on paper. She can also stand beside a board, turn her body and notice that an arrangement looks navigable from above but is awkward at eye level. Each method makes different information available. Intelligent work can combine symbolic description with direct interaction.

The study of spatial reasoning connects position, orientation, distance and movement. It reminds us that intelligence need not appear as a stream of words. A person may solve a spatial problem through a sketch, a gesture or the controlled movement of an object before giving a verbal explanation.

This matters for fairness. If a task requires a particular motor response, a person unable to make that response may be prevented from displaying the understanding we intended to assess. We have measured performance through an interface. Before turning the result into a judgement about the whole person, we should ask what the interface allowed or obstructed.

The environment can also hold part of the work. A ruler makes a length easier to compare. A written list preserves an intention. A labelled tray reduces the need to search for materials. These supports do not erase the distinction between unaided and supported performance. They show why that distinction should be specified rather than assumed.

An intelligent arrangement may reduce unnecessary cognitive demand instead of asking the person to overcome it repeatedly. If the display materials are consistently labelled, Noor can spend more effort on explaining ideas and less on locating tape. The improvement belongs partly to the organisation of the task environment.

This is a different claim from saying that every useful object literally becomes part of one conscious mind. We can describe the contribution of a notebook or a label without resolving the philosophical boundary of the mind. The practical question is what the person-plus-support arrangement can accomplish, and which elements are responsible for which parts of the result.

Bodies also require maintenance, and real performances occur under real conditions. Noor may approach a task differently after a long, tiring day than during a quiet morning. The example cautions against treating one moment as an exhaustive statement of capacity. It does not offer a diagnosis or a simple formula linking a bodily state to an intelligence score.

The physical grounding of intelligence is therefore broader than the electrical signalling inside a head. It includes the processes that sustain the organism, the channels through which information arrives, the actions available and the surroundings that make some problems easier or harder.

Once we see this, the question about other intelligent beings changes. We should not begin by asking whether they can imitate our preferred performance format. We should first ask what problems their bodies and environments present, what information they can use and what evidence reveals flexibility within that relationship.

Small brains and distributed lives

A bee visits the flowers outside the window. Noor has spent the afternoon discussing intelligence as if its natural home were a person sitting at a desk. The bee offers a useful interruption. A creature need not solve our worksheet to face problems involving navigation, learning and changing conditions.

The eduKateSingapore series on why insect intelligence is not human intelligence provides a careful entry. The comparison should neither dismiss insects as unvarying mechanisms nor turn them into tiny humans with miniature versions of every human capacity.

Experimental evidence can identify particular abilities. In a 2017 study, Loukola and colleagues trained bumblebees on a task involving moving a ball to a location associated with reward. The researchers compared different demonstration conditions and examined flexibility in the bees’ performance. The original study supports specific claims about learning and behaviour under its experimental conditions. It does not establish that bees possess human-like understanding of every aspect of the task.

That is the right shape of an inference: describe the species, the task, the preparation and the behaviour, then make a claim proportionate to the evidence. A surprising success deserves attention without becoming permission to attribute an entire imagined inner life.

The colony introduces another scale. A collection of insects can produce coordinated behaviour through interactions among individuals and their surroundings. The Insect World as Distributed Intelligence and the study of local rules and collective behaviour explore this organisation. Group-level outcomes need not require one individual to hold a complete plan of the group.

Here, too, the boundary matters. A coordinated colony is not automatically one conscious subject. A successful collective pattern does not tell us that every member understands the overall result. We should identify the level at which the relevant information is stored, transmitted and used, and the level at which success is being assessed.

What about organisms without nervous systems? Living systems can sense conditions, regulate activity and change through biological processes. Whether particular forms of non-neural responsiveness should be called intelligence depends on definitions and on evidence of the capacities being claimed. Growth towards a resource, by itself, does not prove deliberation or a human-like representation of the future.

Evolution also requires a separate description. Populations can become adapted across generations through processes that do not involve a population consciously choosing its design. An inherited capacity can support an organism’s effective behaviour without having been learned from scratch during that organism’s lifetime. Development, individual learning and evolutionary change operate on different timescales and through different mechanisms.

These distinctions make the material question more interesting. Nature contains many forms of regulation and information use, and some support impressive flexibility. We do not need to flatten them into one universal mind or draw an arbitrary border around human language. We can investigate each system with suitable questions.

Noor watches the bee leave. The most intelligent response available to her may be to admit that she knows less about its competence than she assumed. The creature has not become human. The question has become better: what can this particular organisation of living matter learn and do in the world it inhabits?

What artificial intelligence actually makes artificial

The laptop returns us to an engineered system. Here, “artificial” refers to the construction of the system and its mechanisms. It does not mean that every demonstrated capability is illusory. An artificial light really illuminates. An engineered system can really classify an image, play a game or generate a useful explanation, while remaining very different from a biological mind.

The field of AI contains multiple approaches. Some systems rely heavily on explicitly represented rules and search. Others learn numerical relationships from data. Many combine learned components with other procedures, tools and constraints. There is no single internal mechanism shared by everything marketed as artificial intelligence.

The estate’s How AI Works gives the broader route from data and algorithms to outputs and practical use. For this conversation, the key point is that competence has to be explained through an actual arrangement. The label “AI” does not supply the explanation on its own.

In a learned model, training adjusts parameters through an optimisation process guided by an objective and data. During use, inputs interact with the resulting model to produce outputs. Different designs and training procedures produce different capabilities and limitations. The representation, training history and evaluation conditions therefore matter at least as much as the fact that the machine uses electricity.

The Transformer architecture is an important example of a particular design, introduced by Vaswani and colleagues in Attention Is All You Need. Its attention mechanisms are mathematical operations that relate parts of an input representation. The technical term “attention” does not imply human conscious attention, just as an artificial “neuron” does not imply a biological cell.

Large language-model research offers a further example. Brown and colleagues’ Language Models are Few-Shot Learners investigated a trained model performing tasks from instructions and examples supplied in text, without updating its parameters for those test tasks. That distinction is useful: adapting an output to the current context is not necessarily the same process as permanently changing the model through training.

We should avoid describing such systems as either omniscient minds or mere shelves of copied sentences. Both descriptions can obscure the work. A learned model can produce a novel combination and succeed on tasks beyond an exact stored example. It can also fail on a simple-looking variation, generate an unsupported claim or respond inconsistently. The appropriate conclusion depends on observed performance and its limits.

Noor asks the laptop to explain the exhibition layout. A useful explanation may save time and reveal an option the group missed. It still needs to be compared with the actual room. The system may not have been given the width of the doorway, the height of the boards or the visitors’ needs. An elegant answer cannot restore information that was absent from the task.

This brings the title back into focus. Electrical hardware makes the computation physically possible. The capability depends on how that hardware, software, training and context are organised. The resulting system should be credited for what it can demonstrate, examined where it fails and understood without assigning it properties the evidence has not established.

Learning and following rules are not opposites

“But the computer is following instructions,” Noor says. “Doesn’t that mean it cannot be intelligent?”

The objection sounds stronger than it is because “following instructions” can refer to several different levels. A program executes through an implemented procedure. That procedure may perform a fixed calculation, search through alternatives or update a model from experience. Describing the implementation as a procedure does not tell us which of those capacities it supports.

A learning algorithm is still an algorithm. Its operation can change how later inputs are handled. The designer need not have written a separate answer for every future situation. Conversely, a system can be highly complicated and still fail to learn anything useful. Complexity, programmability and learning are different properties.

Human learning also involves physical processes. If we discovered a more detailed account of the neural events supporting Noor’s reasoning, that discovery would not make the reasoning disappear. A capacity does not cease to exist when its mechanism becomes better understood. The task is to explain the capacity through the mechanism, while checking that the claimed capacity is actually present.

There remains an important difference between possessing a procedure and using it flexibly. A simple script might produce the expected sentence whenever it detects a keyword. A more capable system might use context to distinguish several meanings, combine relevant information and handle a changed request. Both run through physical operations. The difference we care about is what the organised system can reliably do.

Nor must every useful system modify itself continuously. A trained model may remain fixed during a particular use while demonstrating generalisation from prior training. A human can apply a well-learned skill without forming a new long-term memory of every application. When a claim specifically concerns ongoing learning, however, we should test whether experience produces a relevant persistent change.

The testing problem returns. Has the system learned a relationship that survives a change, or has it learned a shortcut tied to the training environment? Can it use a new instruction appropriately, or only reproduce a familiar response associated with familiar wording? What happens when the two possibilities lead to different answers?

This is where cognitive routing becomes a useful bridge from the human learning estate. A capability can exist but fail to reach the problem that needs it. In an AI system, the analogous practical issue may involve selecting the right tool, using the right context or invoking an appropriate checking process. The mechanisms differ; the functional question is still recognisable.

We should also distinguish an explanation generated by a system from a verified account of its internal causal process. A persuasive account of “why I answered this way” may be useful as an argument to inspect, but its fluency does not prove that it faithfully reports every mechanism that produced the answer. We can check the stated reasoning against the task without assuming transparent self-knowledge.

The useful question is therefore not whether matter obeys laws or whether software executes operations. It is what those operations make possible, which evidence supports the capability and how far the capability travels. Repeating that the system is physical does not settle the intelligence question. It tells us where an explanation must ultimately be grounded.

Words, understanding and the world they refer to

Noor can repeat the phrase “the electrification of matter” now. Repetition is not the same as understanding it. Mei asks her to explain why the lamp does not become intelligent simply by using electricity. Noor has to connect the words to distinctions they have examined together.

This is a useful way to think about understanding in a specified domain. Can the learner explain a relationship, recognise an example, reject a misleading example, apply the idea under a change and repair an inconsistency? These are forms of evidence for a usable grasp of the subject. They do not produce a complete philosophical definition of understanding for every possible system.

Language helps because a distinction can become shareable. Once Noor and Mei agree on what they mean by a signal, they can discuss several examples without rebuilding the entire explanation from the beginning. How Language Works connects shared patterns of expression with the reconstruction of meaning in context.

The context matters. “That is light” might describe illumination or the weight of an object. The words alone do not determine the intended meaning. Noor uses the situation, the object being discussed and the speaker’s purpose to select an interpretation. If those cues are missing, an intelligent response may be to ask which meaning is intended.

Words can also conceal a gap. A learner may use “energy,” “system” or “intelligence” as though the name itself supplied a mechanism. The cure is often an example that forces the distinction into action. Which part of this arrangement stores information? What changes when the input changes? Why does this case fit the concept while that one does not?

The eduKateSingapore discussion of vocabulary and thought asks whether a thought requires an existing word. We should resist a simple equation. People can recognise, imagine and solve some problems without first producing a verbal label. Language can then help stabilise, refine and communicate the distinction. It is a powerful resource for thought, not evidence that every thought must already be a sentence.

This has consequences for AI evaluation. A fluent explanation is evidence of a linguistic performance. To assess the understanding claimed for that explanation, we can examine whether its relationships remain coherent across questions, examples and changes, and whether its factual claims are supported. A system may be useful within those tests and still have limits that a persuasive speaking style conceals.

We should not define understanding so that only humans can possess it by stipulation, then present that definition as an experimental finding. We should also avoid granting every form of human understanding to a machine because it can discuss the word. The productive route is to specify the capacity being claimed and investigate it.

For Noor, the distinction is mercifully concrete. She understands today’s idea better when she can explain a counterexample and recognise where her explanation stops. She does not need to sound more elaborate. She needs the words to connect more reliably with the relationships they are meant to express.

The sentence becomes intelligent work when it helps a reader see, infer or do something more accurately. Its polish may help that work travel. Its polish cannot substitute for the work itself.

Consciousness remains another question

At some point Noor asks the question that has been waiting behind the laptop: “Does it know that it is answering?”

We should take the question seriously enough not to answer it with a demonstration that addresses something else. Producing an answer, monitoring an error and reporting uncertainty are observable or testable functions. Whether a system has subjective experience is a further question. The word consciousness often concerns that experience: whether there is something it is like to be the system.

Human beings encounter their own experiences directly and study the experiences of others through behaviour, reports and biological evidence. The scientific challenge is to connect such evidence to mechanisms and explanations. A description of neural activity is essential to many research approaches, but a measurement of electrical activity is not, by itself, a complete theory of experience.

There are competing theories. The Cogitate Consortium’s 2025 adversarial comparison of two theories of consciousness tested predictions associated with integrated information theory and global neuronal workspace theory. Its results supported some predictions while challenging important aspects of both. The study illustrates a research problem being tested, not a final instrument that settles consciousness in every organism or machine.

That distinction protects the rest of this essay. We can investigate whether a system learns or solves problems without pretending that we have thereby measured subjective experience. We can also treat questions about experience as significant without making every practical capability wait for a final philosophical settlement.

Nor should awareness be confused with moral excellence. A conscious person can make a poor judgement. A highly skilled problem-solver can pursue an objectionable purpose. Experience, competence and values are dimensions that interact, but one does not define all the others.

For an engineered system, a first-person sentence should be handled with particular care. The ability to generate “I feel confused” does not by itself establish a felt state. The sentence may be part of a learned linguistic pattern or a designed reporting behaviour. Equally, a system’s inability to produce our preferred verbal report would not, on its own, settle every possible question about experience in every kind of being.

The reasonable position is an evidence boundary. We should say what is observed, what the observation supports and what remains unsettled. That is more exact than either casually declaring a machine conscious or treating an absence of proof as a completed theory of why experience is impossible.

Mei can therefore give Noor a useful answer without inventing certainty. The laptop’s abilities can be tested in specific ways. Its outputs do not automatically establish that it has an inner life. The scientific and philosophical questions about consciousness require their own arguments and evidence.

The title “the electrification of matter” might tempt us to fuse all these questions into one dramatic awakening. We should keep the drama from doing the work of a theory. Intelligence concerns capacities we can examine. Consciousness concerns experience we are trying to understand. A careful account can leave the second question open while making real progress on the first.

Another person is more than a problem to solve

Noor’s friend arrives and looks at the unfinished display. “It’s fine,” the friend says, but begins quietly rearranging a label. Noor could treat the words as the entire message. She could also infer that something may be wrong. The inference is plausible, not certain.

Human intelligence often operates in this uncertain social space. We interpret actions, language and context to form provisional accounts of what other people know, intend or need. The study of social inference examines this work without granting us direct access to another person’s mind.

The intelligent next step may be a question: “Which part would you like to change?” That question can obtain information while respecting the friend’s ability to describe her own concern. Noor does not have to promote her first interpretation into a fact about someone else’s feelings.

This is a place where emotion belongs inside the account rather than being dismissed as interference. A friend’s disappointment may reveal a neglected commitment. Anxiety may identify a task that feels uncertain. Enthusiasm may sustain difficult work. None of these experiences guarantees a correct judgement, but excluding them from the information considered would also lose something relevant.

The challenge is to interpret emotional information without allowing it to dictate every conclusion. Feeling sure does not establish that a claim is true. Feeling threatened does not establish that another person intended harm. A careful response can recognise the experience and still investigate the explanation.

Social competence also has an ethical boundary. A person can be skilled at predicting reactions and use that skill selfishly. The capacity to influence someone does not establish a right to do so, and effectiveness at an objective does not establish that the objective is good. Intelligence and wisdom must remain distinguishable if we are to criticise capable wrongdoing accurately.

Wisdom, in this conversation, concerns judgement about what deserves pursuit, what consequences matter and how capability should be used under uncertainty. It draws on knowledge and reasoning, but also on values, experience and responsibility. We need not invent a universal wisdom score to recognise that solving the assigned problem may leave the most important question unanswered: should this be the problem we are trying to solve?

The estate’s How Trust Works gives social intelligence a practical horizon. People need reasons to rely on one another. Competence is one reason, but integrity and fair treatment matter too. An excellent explanation from a person who habitually misleads us does not create the same basis for cooperation as an equally excellent explanation from someone accountable for its accuracy.

Noor and her friend eventually discover that they had different audiences in mind. One designed the labels for younger children; the other imagined parents reading them. The disagreement was not evidence that one of them lacked intelligence. It revealed a hidden difference in the problem they were solving.

Making that difference visible allows the group to improve the work. They can choose an audience, provide two levels of explanation or redesign the display. Intelligence has helped the relationship because it has become a means of understanding and coordinating with another person. The person has remained a participant, not merely a variable to be controlled.

The classroom is where precision becomes care

A teacher looking at a wrong answer has several possible stories available. The learner may not know a word, may have lost track of the sentence, may have chosen the wrong relationship, may have made an arithmetic error or may not have understood what the question requested. The visible result does not identify its own cause.

If the teacher chooses the wrong story, effort can be directed at the wrong repair. More calculation practice will not necessarily resolve a language misunderstanding. More vocabulary work will not necessarily repair an invalid inference. Encouragement may be welcome, but encouragement alone does not identify the missing step.

This is where an exact account of intelligence becomes humane. It asks for a sufficiently detailed description of the difficulty to make help relevant. It replaces a broad judgement about the learner with a question about a process that can be examined.

The Punggol guide on Primary 3 reading and comprehension distinguishes difficulties that can look similar in a finished answer. A child may struggle to decode words, understand vocabulary, maintain passage meaning or express an inference. These possibilities lead to different teaching decisions.

Mei tries the principle with Noor. She asks Noor to explain the problem in her own words before calculating. The explanation reveals that “remaining” was interpreted as “used.” Mei can now teach the distinction and check it in several contexts. Calling the original answer careless would have named the disappointment without locating the misunderstanding.

On another day, Noor explains the problem accurately but loses an intermediate value. Writing a clear line of working changes the result. On a third day, she can perform the calculation but cannot decide which operation to use. Mixing problem types becomes useful because the missing work concerns selection. The teaching changes because the evidence changes.

The eduKate Sengkang account of a Tutorial treats the lesson as a structured encounter among a learner, material, teaching, practice and feedback. That is a useful counterpart to this essay: intelligence develops and becomes visible within arrangements that can either reveal or obscure the present difficulty.

A good lesson also checks what happens after support changes. Noor may succeed while Mei points to the decisive sentence. Can she find that sentence later without the gesture? She may explain immediately after hearing an explanation. Can she reconstruct the relationship after a delay? These are useful questions about the reach of the learning, not attempts to catch the child failing.

The Sengkang study-skills guide connects preparation, retrieval, practice, checking and transfer. Its relevance is practical: studying contains several decisions, and a learner can improve by discovering which decision is currently weak. Time spent working is real effort, but its effect depends on what that effort is doing.

Precision also protects strengths. If Noor understands a concept well but struggles with its written expression, the teacher should preserve the conceptual achievement while improving the expression. If she is fluent with a method but inflexible in selecting it, the fluency remains an asset. Diagnosis should not erase everything that already works.

The classroom therefore offers a daily answer to “what makes anything intelligent?” We can watch a learner acquire a distinction, use it, encounter a limit and revise. We can also watch teaching make that sequence more likely. The work is modest in appearance: a question clarified, a relationship restored, a check performed independently. Across years, those modest changes become a person’s expanding ability to meet the world.

A pencil can expose what a fluent answer hides

Mathematics gives the conversation a particularly clear testing ground because certain relationships can be stated exactly. If the assumptions and domain are fixed, a transformation can preserve the problem or change it. A correct-looking line is not enough; we can ask what licenses the step.

Mei writes x² = x. Noor divides both sides by x and obtains x = 1. The answer is a solution, but it is not the complete solution over the real numbers. Dividing by x assumed that x was nonzero. The value zero also satisfies the original equation.

They can instead write x² − x = 0, then x(x − 1) = 0. The possibilities are x = 0 or x = 1. Substituting each value into the original equation confirms it. The exercise makes a hidden assumption visible, and the check returns the transformed statement to the problem it was meant to solve.

This is not a claim that a learner who misses the zero lacks intelligence. It identifies a specific vulnerability in the present reasoning. Noor may know how division works and still need to learn how a condition on division affects a solution set. The useful response is precise teaching, followed by problems that reveal whether the condition has become part of her usable understanding.

The Bukit Timah Tutor article How Additional Mathematics Works follows the preservation of relationships through changes in representation. Words can become equations; equations can become graphs; graphs can support an interpretation. Each change has to preserve the meaning relevant to the question.

Now Mei draws a rectangle two units by three units. Its area is six square units and its perimeter is ten units. Double both side lengths and the rectangle becomes four by six: its area is twenty-four square units and its perimeter twenty units. “Everything doubles” was too broad. Length and area change differently under the same enlargement.

Noor can remember these numbers and still miss the relationship. A stronger explanation shows why doubling one side doubles the area, and doubling the other side doubles it again. The fourfold area is connected to two multiplicative changes. A new rectangle can then test whether the relationship has travelled beyond the first example.

Mathematics also distinguishes exact symbolic reasoning from approximate measurement. A drawn side may be measured with limited precision, while a stated algebraic relationship may be exact under the problem’s assumptions. A decimal expansion with many digits does not make an uncertain measurement equally certain. The earlier discussion of journey time has now acquired a formal companion.

These examples explain why working matters. A final result can hide a lost solution, an unjustified transformation or a lucky cancellation. An inspectable route lets the learner and teacher locate the first step that needs repair. It also lets them recognise a sound alternative method instead of assuming that only one memorised route can be valid.

The same discipline reaches beyond mathematics. What did we assume? Which operation preserved the relevant relationship? Did a change of representation remove an important possibility? Does the conclusion survive a return to the original situation? These are mathematical questions with a wider intellectual life.

A pencil has no intelligence of its own in this account. Yet a person using one can make reasoning more visible, preserve intermediate work and discover an error. The object supports a capacity without becoming a person. Matter has been arranged to help thought examine itself, one line at a time.

A group can think badly with brilliant people

Noor’s group now contains enough knowledge to complete the display. One person understands the science, another writes clearly, another sees the spatial problem and another remembers the deadline. The project can still fail. Relevant knowledge being present somewhere in the group is different from that knowledge reaching the decision that needs it.

Suppose the student who notices the blocked doorway says nothing because the group has already praised the layout. Suppose the best explanation is stored in a file nobody can find. Suppose everyone assumes someone else has checked the dimensions. No individual needs to be incapable for the combined result to be poor.

Collective intelligence concerns what a group can accomplish through its organisation and interactions. It should not be treated as the automatic sum of members’ test scores. The original research by Woolley and colleagues on a collective intelligence factor in human groups investigated performance across several group tasks. Evidence from such studies concerns the groups and tasks examined; it does not make a city or civilisation a single measured brain.

The practical lesson can be explored without extending that empirical claim. A group needs ways to contribute information, combine it, resolve relevant disagreements and connect decisions to consequences. The estate’s How Cooperation Works examines the dependencies that allow separate contributions to become a shared result.

Mei asks each student to identify one thing the others might not know. The question changes the meeting. Instead of repeating their agreement that the display looks good, they reveal the different pieces of the problem. One has measured the doorway. Another has spoken to the younger class. Another knows that the available boards are shorter than the sketch assumes.

The group now needs to integrate the information. Agreement achieved by ignoring the measurement would not be a successful synthesis. Nor would giving every opinion equal weight regardless of what supports it. Respecting people and evaluating claims are compatible. A directly relevant measurement can deserve more weight for a dimensional question while everyone remains entitled to be heard.

The group also needs to distinguish independent evidence from repetition. If four students copied the same incorrect source, their agreement does not amount to four independent checks. A second method of measurement or a return to the original document may add more information than another confident repetition.

Responsibility matters at the handoff. “Someone should check the final version” is an intention without an owner. “Leila will check the dimensions on the version we print” makes the task more definite. A deadline and a visible result make it easier to discover whether the check actually happened.

This is not a recipe for turning every conversation into administration. It is an explanation of why coordination can determine whether existing intelligence becomes useful. A group that makes relevant concerns safe to raise, preserves a shared version and checks consequential assumptions can outperform its own more chaotic arrangement.

The collective achievement remains dependent on people and tools with limited knowledge. It can improve or deteriorate as its organisation changes. That observation prepares us for the next scale. A civilisation can contain enormous expertise and still fail to connect the right observation, the right authority and the right action in time.

How a city holds conclusions

Leave the room and the scale changes. Noor steps onto a pavement, reads a sign, crosses at a marked place and enters a building whose stairs have already been designed. She is using decisions made by other people before she arrived. Some of those people are no longer alive. Much of their work remains available without requiring their presence.

A city can be read as a material record of selected knowledge. A drainage channel embodies judgements about moving water. A timetable records an intended coordination of journeys. A standardised sign allows a traveller to use a distinction learned elsewhere. None of these objects is thinking merely because it carries the result of thought. They make earlier thought usable within present activity.

This is where the Intelligence of Civilisation becomes a companion to the individual account. A society needs more than people who know things. It needs routes by which observations, expertise and decisions can meet. The right answer in the wrong office may have little effect on the problem outside.

Imagine a maintenance worker noticing that a sign is repeatedly misunderstood. The observation has potential value. To become an improvement, it must reach someone who can assess the problem, compare alternatives, approve a change and check whether the change helps. The intelligence of the arrangement depends partly on whether that route exists and remains usable.

Now imagine a city with a large database but poor information quality. The records are abundant, but dates are missing, definitions differ and outdated entries remain indistinguishable from current ones. More storage has not automatically produced more usable knowledge. It may have increased the work required to discover what can be trusted.

The estate’s study of distributed memory concerns a related problem: groups do not need every member to know everything, but they do need ways to locate relevant knowledge. Knowing who knows, where the record is and how to assess its reliability becomes part of the collective capability.

This makes institutions important without making them infallible. Procedures can preserve hard-won lessons. They can also preserve an obsolete assumption. Expertise can guide a decision. It can also be isolated from local information. A shared standard can improve coordination while requiring revision when the conditions it assumed no longer hold.

Calling a civilisation intelligent is therefore a functional description at a particular scale. It asks whether the organised system can notice, interpret, coordinate, correct and preserve relevant knowledge. It does not claim that the civilisation has one nervous system, one intention or one stream of consciousness.

The broader Civilisation essay follows the organised work that makes ordinary life possible. Intelligence belongs inside that account as a capacity that must be connected to materials, energy, roles, trust and maintenance. A brilliant plan cannot repair a pipe without the people, tools and permissions needed to act.

Noor’s walk becomes more interesting when she sees these dependencies. The signs and structures are not simply background scenery. They are part of an inherited arrangement that changes what she can know and do. The city has held some conclusions in matter. Its continuing intelligence depends on whether living people can understand those conclusions, discover their limits and revise them when necessary.

Intelligence that crosses generations

An idea can outlive the person who first worked it out, but survival is not automatic. Someone must preserve a usable trace, someone else must be able to interpret it, and the conditions for applying it must still be available or recoverable. A page that nobody can read is materially present and functionally limited.

This is why the earlier essay What is a Museum | The Idea belongs beside the present question. A museum can preserve an object and evidence of its history, while interpretation helps later people ask what the object can tell them. The object does not transmit a complete past simply by remaining intact.

Consider an old instrument. A future visitor may see its shape without knowing what it measured, how it was calibrated or which errors its users had to avoid. A label may restore some of that context. A manual, a surviving workshop practice and a trained demonstrator may restore more. Preserving the object and preserving the capability are related tasks with different requirements.

Education performs another part of the handover. Noor inherits language, notation, procedures and examples she did not invent. Her teachers help her turn those inheritances into capacities she can exercise. If she later explains the corrected algebra problem to a younger student, she participates in the same process on a small scale.

The 1000-Year Civilisation Test makes the dependence vivid by asking how much of modern life a person could reproduce without its surrounding tools and supply chains. The thought experiment separates knowing that a capability exists from being able to rebuild the arrangements that produce it. Individual intelligence can remain present while access to accumulated capability disappears.

Time also creates obligations. A generation may preserve knowledge while leaving the next generation expensive failures, neglected maintenance or commitments it cannot easily change. The Finance essay, How Civilisations Borrow from the Future, examines how present decisions rely on expectations about future resources and work. Intelligence enters through the quality of those expectations and the institutions that examine them.

Noor’s display has a smaller version of this problem. The group wants next year’s class to reuse it. They can leave a photograph, but a photograph alone may not reveal which materials were borrowed, which dimensions mattered or which arrangement blocked the doorway before it was corrected. A useful handover includes the reason for the successful arrangement and the conditions that made it work.

The estate’s Temporal Reasoning article is relevant because consequences do not all arrive at once. Some decisions work today and create a problem later. Some investments impose present effort to produce a later capability. Judgement across time requires keeping delays, dependencies and changing conditions in view.

The future should inherit room to disagree. A preserved explanation is more valuable when its evidence, assumptions and limitations remain visible. Later people may have better instruments or face different problems. Respecting the past does not require freezing every conclusion it reached.

This is a further meaning of intelligence in matter. A book, instrument, drawing or building can carry selected results of earlier intelligence into a later world. Living people then have to reactivate, test and extend them. Civilisation enlarges what can pass between generations when it preserves both useful answers and the means of correcting them.

How capable systems lose their way

It would be comforting if more intelligence automatically produced fewer mistakes. In practice, a capable system can use its abilities to pursue the wrong objective, protect a mistaken belief or produce an impressive explanation of an inadequate result. The capacity to construct a persuasive argument is not identical to the discipline of testing it.

Noor experiences a small version when she becomes attached to the maze layout. She can now produce several reasons why it should work. The reasons become more elaborate as the evidence against the blocked route becomes clearer. Her reasoning ability has not vanished. It has been recruited to defend the preferred outcome.

One failure begins with the objective. If the group rewards the number of labels printed, it may produce many labels that nobody can understand. If it rewards a display’s first impression, it may neglect what visitors can learn after entering. A metric can be met while the purpose behind it remains poorly served.

Another failure begins with the evidence channel. People may stop reporting problems when reporting them brings embarrassment or punishment. An organisation can then look increasingly successful because the information reaching its decision-makers has become increasingly selective. Silence is misread as confirmation.

A third failure begins with the model’s age. A method that worked under one set of conditions continues to be applied after a relevant condition changes. Past success supplies confidence, but the system fails to check whether the success depended on assumptions that no longer hold.

A fourth failure concerns accumulated error. A small mistake enters a record, is copied into a summary and later appears as an established fact in a new explanation. The number of repetitions grows while the independent evidence remains unchanged. The result can look authoritative precisely because its original weakness has become difficult to locate.

None of these failures requires a mysterious disappearance of intelligence. They concern how capability is directed, informed and corrected. The estate’s Error-Correction study makes the decisive distinction: an error becomes useful only when its evidence reaches the process that can change the next response.

That return can fail in a classroom, a machine or an institution. Noor sees a red mark but never discovers the misunderstanding. A model produces an error but its user never checks the result. An organisation receives a complaint but routes it somewhere unable to alter the process. The information exists; the repair does not follow.

Scientific inquiry offers an organised response to this problem. Its methods aim to make claims, observations and reasoning available for scrutiny and correction. The estate’s How Scientific Research Works follows this connection between an individual result and a larger process of investigation. The institution remains a human undertaking with limitations, not a guarantee that every published claim is correct.

The practical lesson is to protect the route by which reality can object. Keep the original question visible. Preserve evidence that challenges the preferred answer. Make consequential assumptions inspectable. Allow a changed observation to change the decision. These habits do not replace intelligence; they help its capacities remain connected to the world they are supposed to serve.

Noor finally moves the board. The display becomes less like the picture she first imagined and more like a place visitors can actually use. The correction is not a defeat of her intelligence. It is evidence that the relevant part of the intelligence has returned to work.

A practical examination of anything called intelligent

We can now return to Noor’s first question with a more useful set of instruments. They are questions for examining a claim, not a validated universal intelligence test. No short checklist can assign a final intelligence score to a child, a bee, a language model and a city on one uncontested scale.

Begin with the capacity being claimed. “Intelligent” is too broad to evaluate without further description. Does the claim concern route-finding, learning a category, solving algebra, interpreting language, coordinating a team or something else? A claim that names its domain can be tested more fairly than a claim that expands whenever a limitation appears.

Next, identify the information available. What could the system observe or retrieve? Was the decisive fact in its input? Did a person supply a hint? Was the answer present in its preparation? A failure caused by missing information differs from a failure to use information that was available. A success supported by a hidden answer differs from one produced through the claimed capability.

Then change a relevant condition. Keep the intended problem sufficiently clear while varying something that can distinguish competing explanations of success. Change the wording, alter a surface feature, introduce an exception or require the same relationship in another representation. The purpose is to investigate reach, not to construct an arbitrary impossible task.

Compare with simpler explanations and suitable baselines. Would a fixed rule, a lookup table or a lucky guess account for the result? Sometimes a simple procedure is exactly what the task needs. If a broad intelligence claim is being made, however, the evidence should demonstrate more than the simpler explanation can support.

Record the support and cost. How much time, training, computation, instruction or external help was needed? A capable arrangement may be worthwhile despite substantial support. The support must remain visible so that its contribution is not silently attributed to one component. A learner using a calculator and a learner calculating mentally are performing differently specified tasks.

Examine uncertainty and correction. What happens when the answer is ambiguous, information is missing or a previous response is contradicted? Can the system detect the relevant difficulty, revise appropriately or obtain help? Not every intelligent act displays all these capacities, but claims of broad dependable competence should explain their limits.

Finally, separate capability from permission and purpose. Even a very effective system does not thereby establish which goals should be pursued or who should bear their consequences. Evaluation of intelligence informs a judgement. It does not replace ethical or institutional responsibility for the judgement.

Apply these questions to the room. The lamp’s useful behaviour is explained by its circuit and switch. The basic thermostat adds a narrow feedback arrangement. The learner shows capacities across explanation, changed examples and correction. The laptop requires examination of the particular system, its task, its training and its available context. The group requires examination of its interactions as well as its members.

The result is not one dramatic line separating all matter into the intelligent and the unintelligent. It is a set of more defensible claims about different systems. Some boundaries remain matters of definition. Many individual capacities can nevertheless be investigated with considerable precision.

That is a worthwhile form of exactness. Say what you mean, construct evidence that addresses it, test plausible alternatives and state the limits. The word intelligence becomes more useful when it carries an accountable claim instead of doing the work of praise, fear or marketing.

Questions behind the question

What is intelligence in one sentence?

Intelligence is the capacity to acquire and use information to learn, infer, solve problems and adapt effectively when relevant conditions change. This is the essay’s working definition. It describes a family of capacities whose evidence must be specified by task and context.

Is intelligence literally electricity?

No. Electrical and electrochemical processes support important implementations, including nervous systems and electronic computers. Intelligence concerns capacities of the organised system. Electrical activity alone is insufficient evidence, and the general concept does not require every possible implementation to share one substrate.

What makes something intelligent rather than merely responsive?

We look for the information-using capacities being claimed: for example, learning a useful relationship, drawing an appropriate inference or handling relevant changes beyond a fixed response. The threshold for using the word varies across theories. Describe the mechanism and demonstrated flexibility instead of relying only on the label.

Is knowledge the same as intelligence?

Knowledge concerns what is retained and organised; intelligence concerns capacities that acquire and use information and knowledge. They are closely connected. A system can contain extensive records yet be poor at selecting or applying them. A person can reason well while lacking a fact essential to a particular answer.

Must an intelligent system be learning every moment?

No. A capacity can be exercised using earlier learning, and some competence depends partly on inherited or engineered organisation. A claim of ongoing learning requires evidence of relevant change through experience. A temporary absence of performance also does not, by itself, establish the absence of an underlying capacity.

Can a machine be intelligent without being conscious?

Intelligent performance and consciousness concern different claims. We can evaluate a machine’s learning, inference and problem-solving without thereby establishing subjective experience. Whether any particular system has experience requires further arguments and evidence; a fluent first-person report does not settle it.

Can an intelligence score be exact?

The calculation of a score can be exact within a stated scoring system. Its interpretation depends on what the tasks sample, how they were administered and what the score is being used to infer. No score removes the need to describe scope, support, uncertainty and relevant limitations.

Can intelligence improve?

People can improve many specific skills, knowledge structures, strategies and habits of checking. Teams can improve how they share and use information. Such improvements should be demonstrated in the capacities claimed. They do not justify a promise that every person’s general cognitive profile will change identically or without limit.

Is civilisation one giant mind?

Civilisation can organise distributed knowledge and preserve capabilities across time. Calling that arrangement intelligent is a functional description of collective performance. It does not imply one brain, one intention, one consciousness or agreement among all the people within it.

Why does this matter to a teacher or parent?

Because useful help depends on identifying the actual difficulty. A mistaken answer may involve language, memory, selection, reasoning or execution. Investigating the process makes teaching more relevant and protects the learner from a sweeping judgement based on evidence too narrow to support it.

The light is still on

The exhibition is finished. Noor returns to the dining table, where the stone still rests beside the notebook. Mei switches on the lamp. It responds exactly as it did before their conversation.

Noor’s account of the response has changed.

She can now distinguish physical activity from the capacities that activity might support. She can explain why a signal is not automatically a thought, why a remembered answer is not the whole of understanding and why a fluent voice is not a complete theory of a mind. She can also recognise that intelligence need not appear only in her own preferred form.

The bee, the computer and the group have become more interesting without becoming interchangeable. Each invites questions about organisation, information and competence. Each requires an appropriate kind of evidence. The unanswered questions have acquired clearer shapes.

The conversation has changed something else. Noor can look at her own earlier explanation without needing to defend it. “I thought responding was enough,” she says. “Now I would ask what the response can do when the situation changes.” She has not solved every dispute in the science of intelligence. She has improved a distinction and made it available for future use.

That modest event contains much of what we have been examining. An experience entered a living system. It was connected with other experiences, represented in language, tested through examples and revised through disagreement. The resulting understanding can now be used in another conversation. It can travel beyond the room.

The title points towards this possibility. Matter can be organised so that the world makes a difference within it, and those differences can contribute to learning, inference and action. In humans, that organisation belongs to a living, embodied history. In machines, it belongs to particular engineered systems. Between people, tools and institutions, it can support capabilities no individual holds alone.

Electricity is part of some of these physical stories. It is not the missing definition. The exactness of intelligence lies in describing the capacity, finding its mechanism where we can, testing its reach and refusing to claim more than the evidence supports.

The lamp continues to illuminate the page. On the page, Noor writes a better question.

Reading the evidence and continuing the conversation

The room, its characters and the numerical comparisons are illustrative constructions. The working definition and the connections across learning, AI and civilisation are this essay’s synthesis. The electrical title is a metaphor with an explicitly limited physical meaning.

The public eduKate links throughout the article lead to the existing intelligence, learning, language, mathematics, insect-world and civilisation studies. Begin with How Intelligence Works for the larger conceptual map, or return to the particular study beside the question that interests you.

The scientific foundations include the following primary research and institutional explanations. Each supports its associated claim; none is presented as endorsing the essay’s complete synthesis.

• Neural signalling: the Queensland Brain Institute, alongside the neuroscience chapters on chemical and electrical synapses.
• Memory mechanisms: Eric Kandel’s Nobel lecture, drawing on his research into learning and synaptic change.
• Definitions and measurement: the APA overview, Neisser and colleagues’ 1996 report, Legg and Hutter’s formal proposal, and Chollet’s evaluation proposal.
• Learning and transfer: the National Academies’ How People Learn II.
• Insect cognition: Loukola and colleagues’ bumblebee experiment, published in 2017.
• Machine learning: Vaswani and colleagues on Transformers and Brown and colleagues on few-shot language-model performance.
• Group performance: Woolley and colleagues’ collective-intelligence study, published in 2010.
• Consciousness research: the Cogitate Consortium’s 2025 adversarial study, presented here as evidence of an active research debate.