Chapter 13: Machine Intelligence
Ananya orders a school bag from an online shop. It arrives with a broken zip. She opens the shop’s help page and types, “My bag has a broken zip. Can I return it?” A reply appears in about two seconds. It is polite, it is correct, and it even mentions her order number.
She types again, this time asking about a different bag she liked. The reply is helpful once more. Then she asks whether the shop stays open on Sunday, because her exams begin on Monday. This time the reply simply repeats the return policy.
That is when Ananya stops and wonders: have I been talking to a person all this while, or to a program? She cannot tell by looking. There is no face and no voice, only words on a screen. Her friend Rehan is sure it was a person, because the answers sounded so natural. Ananya is not so sure. The speed of that first reply, and the odd answer at the end, both point the other way (see Fig. 13.0-1).
Diagram
Fig. 13.0-1 – a girl at a laptop looking at a chat window; two speech bubbles come out of the screen, one leading to a question mark shaped like a human head and the other to a question mark shaped like a computer; label the girl “user”, label the two possibilities “a person?” and “a program?”
This chapter will answer each of these five questions.
13.1 What Is Intelligence?
Think about an ordinary morning. You look at the sky, decide it might rain, and put a raincoat in your bag. On the way to school you meet a friend who looks upset, so you speak gently instead of joking. In class you meet a maths problem you have never seen before, and you work it out using a method you learnt last month for something else.
Nobody gave you a fixed rule for any of those three moments. You used what you already knew, judged a new situation, and acted. That ability is what we call intelligence.
Intelligence is the ability to acquire, understand, and apply knowledge and skills — to solve problems, adapt to new situations, think critically, and learn from experience. The important word here is new. Repeating something you have done a hundred times needs memory. Handling something you have never faced before needs intelligence.
13.1.1 The Abilities We Call Intelligence
Intelligence is not one single skill. It is a bundle of abilities that work together, and different tasks call on different ones. Five of these abilities come up again and again (see Fig. 13.1-1).
Diagram
Fig. 13.1-1 – a central circle labelled “Intelligence” with five arrows going out to five labelled boxes: “Reasoning”, “Problem-solving”, “Abstract thinking”, “Comprehension”, “Learning”
Reasoning is drawing a sensible conclusion from what you already know. The floor is wet and the window is open, so it probably rained in the night.
Problem-solving is finding a way from where you are to where you want to be. Your bicycle chain slips, and you work out what to check first.
Abstract thinking is handling ideas that you cannot touch. The number seven, the idea of fairness, and the plan of a route on a map are all abstract.
Comprehension is understanding the meaning behind words or events, not just the words themselves. When a friend says “I am fine” in a flat voice, you understand that they may not be fine at all.
Learning is changing what you do because of what happened last time. You burnt the milk once by leaving the flame high, so now you lower it.
13.1.2 Intelligence Is More Than Marks
It is easy to think intelligence means only school work — solving equations, remembering dates, writing neat answers. That is too narrow a view.
Intelligence also covers emotional and social understanding: knowing when a friend needs to be left alone, or how to settle an argument between two teammates. It covers creativity, which is producing something that did not exist before, such as a story or a new way of arranging a shop’s shelves.
It covers practical skills too — the mechanic who listens to an engine and knows which part is loose is doing something difficult, even though it is not written work. And it covers adaptability: the ability to cope when the situation changes and the old method stops working.
So intelligence is broad. Some of these abilities can be measured in a test, and many cannot. Keep this full picture in mind, because the rest of the chapter compares this broad human ability with what a machine can do — and the comparison only makes sense if you remember how much “intelligence” really covers.
13.2 Human Intelligence and Machine Intelligence
Give a class of forty students a list of ten thousand numbers and ask them to find the largest one. It will take a long time, and somebody will make a mistake. Give the same list to a computer and the answer comes back at once, correct every time.
Now change the task. Ask the same class to look at a photograph of a crowded railway platform and say which person looks worried. Every student will manage it in a second, without being taught how.
Two different kinds of ability are at work here. The first is machine intelligence — the ability of a computer or a machine to carry out tasks that would need intelligence if a person did them. It is the practical side of artificial intelligence (AI), the branch of computer science that builds such systems.
Machine intelligence is not born and it does not grow by itself. It is built. Somebody writes the algorithm — a fixed, step-by-step method for doing a job — and somebody supplies the data, the recorded facts, numbers, words or images the system works on. Human intelligence, by contrast, is biological. It is shaped by evolution and by everything you have lived through, and it develops over years of learning and interacting with the world around you.
13.2.1 Where Human Intelligence Is Strong
Human intelligence is strongest exactly where the situation is messy and the rules are unwritten.
You understand context. If a friend says “nice haircut” while smiling in a particular way, you know whether it is a compliment or a joke. The words alone do not carry the answer; the situation does. You also handle ambiguity — a sentence that could mean two things — by picking the meaning that makes sense here, not the one that is more common in general.
You learn from very little. A small child who has seen two or three dogs can recognise a breed of dog it has never seen. You can be told a rule once and apply it correctly the next day to a completely different case.
You are adaptable. When a road is closed you find another route; when a method fails you try a new one. And your decisions are tied to consciousness, self-awareness and emotion. You know that it is you who is deciding, and how you feel affects what you choose. That is not always a weakness — caring about the result is often what makes a person check their work.
13.2.2 Where Machine Intelligence Is Strong
Machine intelligence is strongest exactly where a human gets slow, bored or tired.
A computer can go through a very large amount of data quickly and find patterns in it — a repeating shape, a repeating word, a repeating relationship between numbers. It performs a narrow task the same way the ten-thousandth time as the first time. It does not get tired, and it does not lose concentration at the end of a long shift.
This is why machine intelligence is used for jobs such as image recognition (deciding what is present in a picture), natural language processing (working with human language as text or speech), and playing games with very large numbers of possible moves.
That last point matters. A machine does not step outside its task. It can adapt to new data of the kind it was designed for, but a situation of a completely new kind usually produces a confident, wrong answer rather than the words “I do not know”.
A machine also has no consciousness, self-awareness or emotion. A system can be built to sort faces in photographs into groups labelled “smiling” and “not smiling”, but it is matching pixel patterns against labelled examples. Nothing in it feels anything about the result.
13.2.3 Human Intelligence versus Machine Intelligence
Put the two side by side and the pattern is clear: they are strong in opposite places. Table 13.1 sets out the main differences.
Neither column is “better”. They answer different questions. If the job is to check ten lakh bank entries for an unusual pattern, the second column wins easily. If the job is to decide whether a particular customer deserves a second chance, the first column is the one that matters.
13.3 The Turing Test
At the start of this chapter, Ananya was stuck on a question she could not settle: was the help page being answered by a person or by a program? She had no way to look behind the screen. All she had were the replies themselves.
That is exactly the situation a British mathematician and computer scientist named Alan Turing thought about, long before online shopping existed.
Diagram
Fig. 13.3-1 – photograph of Alan Turing
13.3.1 The Question Alan Turing Asked
In 1950, Turing published a paper called “Computing Machinery and Intelligence”. He began with the question “Can machines think?” and then set it aside, because it leads nowhere. To answer it you would first have to define “think”, and nobody agrees on that definition even for humans.
So he replaced it with a question that can actually be settled. Instead of asking what is going on inside the machine, ask what the machine can do. If a machine can hold a conversation so well that a person cannot tell it apart from another person, what more would we demand before calling its behaviour intelligent?
This idea became known as the Turing Test.
Notice what the test measures. It does not look inside the machine at all. It judges only the behaviour that comes out.
13.3.2 How the Turing Test Is Carried Out
The set-up is simple, and every part of it has a reason (see Fig. 13.3-2).
Diagram
Fig. 13.3-2 – a human evaluator at a keyboard on the left, behind a wall; two doors in the wall lead to a person on one side and a computer on the other; arrows in both directions labelled “typed messages only”; the evaluator has a thought bubble reading “which one is the machine?”
A human evaluator communicates with two hidden entities. One is a machine, the other is a human. The evaluator can ask anything and can go on for as long as allowed. All communication happens as typed text — no faces, no voices, no photographs. That rule is there to stop the evaluator from deciding on the wrong evidence. A metallic voice or a mechanical body would give the answer away without saying anything about the quality of the replies.
At the end, the evaluator says which of the two is the machine. If the evaluator cannot reliably distinguish between them from the responses alone, the machine has passed.
Step 6 matters. One evaluator fooled once proves very little, because a single lucky exchange can go either way. The test is meaningful only when many evaluators, over many conversations, still cannot do better than guessing.
13.3.3 The Role of the Turing Test in the Development of AI
The Turing Test was never a machine anyone built. It was an idea — but it was one of the earliest and most influential ideas in the development of AI, and it shaped the field in four ways.
It gave researchers a target
Before Turing, “build an intelligent machine” was too vague to work on. The test turned it into something concrete: build a machine whose written replies cannot be told apart from a person’s. A goal you can aim at is far more useful than an ambition you cannot measure.
It pushed work on language
Because the test is conducted entirely in conversation, it directed attention towards making machines handle human language — understanding a typed question and producing a sensible reply. Natural language processing became a central part of AI research, and it still is.
It started a debate that is still going on
The test raised questions that go well beyond engineering. If a machine behaves exactly like a person, does the difference matter? What is intelligence, and is consciousness a separate thing from it? These arguments shaped the ethical side of AI — including the question of what should happen when machines can convincingly imitate people.
It showed its own limits
As AI systems improved, researchers found that the test rewarded imitation rather than ability. That pushed the field towards more detailed ways of measuring performance: how accurately a system labels images, how well it translates a passage, how often it answers factual questions correctly. Modern AI research reports these specific measures rather than a single pass or fail.
13.3.4 What the Turing Test Cannot Tell Us
Passing the Turing Test means one thing only: in a text conversation, this machine was not reliably distinguishable from a person. Several things do not follow from it.
It does not show consciousness or self-awareness. The test never examines what is inside the system, so it cannot report on something it does not look at.
It does not show understanding. A program can be built with a large stock of prepared replies and rules for matching them to what the user types. Such a program may hold up for a while in a short conversation, but it is matching patterns, not grasping meaning. Fooling an evaluator and understanding the conversation are two different achievements.
It also says nothing about abilities outside conversation — recognising objects, moving safely across a room, or reasoning about numbers. A system can be excellent at chat and useless at everything else.
13.4 Connectivity Between Human Intelligence and Machine Intelligence
So far this chapter has set the two kinds of intelligence against each other. That is only half the picture, and it is the less interesting half.
The two are also closely connected, for a simple reason: machine intelligence was designed by people who were copying what people do. When engineers wanted a machine to sort photographs, they thought about how a person sorts photographs. When they wanted a machine to improve with practice, they thought about how a person improves with practice. So the parallels between the two are not an accident — they were built in on purpose.
This section looks at six places where human and machine intelligence meet.
13.4.1 Learning and Adaptation
You learn from experience. Burn the milk once and you turn the flame down next time. Nobody gives you a rule; the experience itself changes what you do.
Machines were given a version of the same idea, and it is called machine learning (ML) — a way of building systems that improve at a task by processing examples, instead of being given a complete rule for every case. An algorithm goes through a large collection of examples, finds patterns in them, and adjusts its own internal settings until its answers match the correct ones closely enough. The set of adjusted settings that results is called the model, and the model is what makes predictions on new data afterwards.
Think of an email service separating unwanted advertisements from ordinary mail. Nobody could write a rule for every trick an advertiser might use. Instead the system is shown a very large number of emails already marked as wanted or unwanted, and it learns which patterns of words tend to go with which label.
The parallel is real, but it is not exact. You needed to burn the milk once. The system needs thousands of marked examples. And when something genuinely new appears — a kind of message unlike anything in its examples — the system has nothing to fall back on, while you would simply use common sense.
13.4.2 Reasoning and Problem-Solving
People reason in many ways. You work forward from facts to a conclusion, backward from a goal to the steps needed, and sideways by noticing that a new problem resembles an old one.
Machines were given algorithms and models to do particular kinds of reasoning. Logical reasoning is one: given a set of conditions, work out what must follow. Optimisation is another: out of a very large number of possible choices, find the best one according to some measure. A delivery app working out the shortest route through eight addresses is solving an optimisation problem, and it will beat any human at it.
But the machine is reasoning inside a box that somebody drew for it. It has the general adaptability and the common-sense reasoning of a human nowhere in its design. The route planner will happily send a truck down a road that a local driver knows is flooded every monsoon, because “everyone knows about that road” is not in its data.
13.4.3 Human-AI Collaboration
Because the strengths sit in opposite places, the sensible arrangement is not one replacing the other. It is the two working together, each doing the part it is good at.
An AI system can take on work that is repetitive, data-heavy or simply too large for a person — going through thousands of records, scanning every photograph, checking every transaction. What comes out is a short list: the cases worth a human’s attention, with the reasons attached.
The person then does the part the machine cannot. They bring context the data does not contain, they weigh up what is fair, and they take responsibility for the final decision. Figure 13.4-1 shows the loop.
Diagram
Fig. 13.4-1 – a four-stage loop: box 1 “Data” → box 2 “AI system: sorts, spots patterns, suggests” → box 3 “Person: adds context, judges, decides” → box 4 “Decision and outcome” → arrow back to box 1 labelled “feedback becomes new data”
In a hospital, for example, a system may go through a large number of scans and mark the ones that show an unusual pattern. The doctor examines those scans, considers the patient’s history and symptoms, and makes the diagnosis. The system narrowed the search. The doctor decided.
13.4.4 Human Feedback and Supervision
The loop in Fig. 13.4-1 only works because people are involved at every stage — not just at the end.
People collect and prepare the data, and mark what the correct answers are. People choose and design the algorithm. People look at the results and judge whether they make sense, because a number that is statistically correct can still be practically absurd. And people are responsible for what the system does once it is running.
That last responsibility exists because a learning system copies whatever is in its data, including the mistakes. If the recorded examples came mostly from one kind of situation, the model will work poorly on every other kind. This is called bias — a systematic tilt in the data or the method that makes the results consistently unfair or wrong for some group of cases. A machine has no way to notice that its own data was one-sided. Only a person can ask where the data came from and who is missing from it.
13.4.5 Neural Networks: An Idea Borrowed from the Brain
The clearest case of an idea taken from human biology is the artificial neural network.
Your brain contains a very large number of nerve cells, each connected to many others. A cell receives signals from the cells connected to it and passes on a signal of its own. No single cell recognises your friend’s face; the recognition comes out of the pattern of activity across an enormous number of them.
An artificial neural network borrows that arrangement (see Fig. 13.4-2). It is built from simple units arranged in layers. Each unit receives numbers from the units before it, combines them, and passes a number on to the units after it. Every connection carries a weight — a number that decides how much that particular input matters — and learning consists of adjusting those weights until the network’s output matches the correct answers on the training examples.
Diagram
Fig. 13.4-2 – a simple network of circles in three columns; left column labelled “input layer” with three circles, middle column labelled “hidden layer” with four circles, right column labelled “output layer” with two circles; every circle in one column joined by a line to every circle in the next; one line labelled “each connection has a weight”
13.4.6 AI and Human Values
The last connection is the most important one, and it is not technical.
Every AI system is put to work in a human situation, so its design and its deployment need careful thought about fairness, privacy and accountability. Fairness asks whether the system treats similar cases alike. Privacy asks whether the data it uses was collected with permission and is kept safe. Accountability asks a plain question: when the system gets something wrong, who answers for it?
A machine can be given a target to optimise. It cannot be given values. It has no way to weigh what is right against what is efficient, because it does not compute anything of that kind. That weighing has to stay with people — which is why human oversight and control over AI systems is not an optional extra but part of the design.
The aim is straightforward: AI systems should act in ways that are beneficial and aligned with human values, and there should always be a person able to inspect what the system did and stop it if it is doing harm.
13.5 Summary
Intelligence is the ability to gain knowledge and skills and apply them — to solve problems, adapt to new situations, think critically, and learn from experience.
The abilities that make up intelligence include reasoning, problem-solving, abstract thinking, comprehension and learning, along with emotional and social understanding, creativity, practical skill and adaptability.
Machine intelligence is the ability of a machine to carry out tasks that would need intelligence if a person did them. It is built by people using programs, algorithms and data.
Human intelligence is biological and develops over a lifetime; machine intelligence is artificial and is created through design, programming and large amounts of data.
Human intelligence is stronger where the situation is unfamiliar or unwritten — understanding context, handling ambiguity, using common sense, being creative, and learning from very few examples.
Machine intelligence is stronger where the work is repetitive and data-heavy — processing large amounts of data quickly, recognising patterns, and giving the same result every time without tiredness.
A machine has no consciousness, self-awareness or emotion. It can be built to detect and label emotional signals in data, but nothing in it feels anything about the result.
The Turing Test was proposed by Alan Turing in his 1950 paper “Computing Machinery and Intelligence”. A human evaluator holds typed conversations with a hidden human and a hidden machine and tries to say which is which.
If the evaluator cannot reliably tell the machine from the human on the basis of the replies, the machine is said to have passed the Turing Test.
The conversation is text-only so that appearance and voice cannot influence the evaluator’s judgement.
The Turing Test shaped AI by turning a vague ambition into a testable goal, by directing research towards human language, by starting lasting debates about intelligence and ethics, and by revealing its own limits, which led to more specific ways of measuring performance.
Passing the test shows only that the machine’s conversational behaviour was indistinguishable from a person’s. It does not show consciousness, understanding, or any ability outside conversation.
Human and machine intelligence are connected because AI was designed to copy aspects of human thinking: learning from examples, reasoning towards a goal, and recognising patterns.
Machine learning is a way of building systems that improve at a task by processing examples rather than by being given a rule for every case. The set of internal settings it arrives at is called the model.
In human-AI collaboration, the system narrows large amounts of work down to the cases worth attention, and the person adds context, judges fairness, and takes the final decision.
Human feedback and supervision run through every stage: people collect and mark the data, design the algorithms, interpret the results, and answer for what the system does.
A learning system copies whatever is in its data, including a one-sided view, which produces bias. Only people can ask where the data came from and who is missing from it.
An artificial neural network borrows one idea from the brain: many simple units, richly connected in layers, producing a result together. Each connection carries a weight, and learning means adjusting those weights.
The design and deployment of AI systems need careful attention to fairness, privacy and accountability, and human oversight must remain in place so that these systems act in ways that are beneficial and aligned with human values.
13.6 Key Terms
Algorithm – A fixed, step-by-step method for carrying out a job.
Artificial intelligence (AI) – The branch of computer science that builds systems able to carry out tasks that would need intelligence if a person did them.
Artificial neural network – A machine learning method built from many simple connected units arranged in layers, loosely modelled on the connected nerve cells of the brain.
Bias – A systematic tilt in the data or the method that makes a system’s results consistently unfair or wrong for some group of cases.
Data – The recorded facts, numbers, words or images that a system works on.
Intelligence – The ability to gain knowledge and skills, understand them, and apply them to solve problems, adapt to new situations, and learn from experience.
Machine intelligence – The ability of a machine to perform tasks that would need intelligence if done by a person, created by people using programs, algorithms and data.
Machine learning (ML) – A way of building systems that improve at a task by processing examples, instead of being given a complete rule for every case.
Model – The set of internal settings a machine learning system arrives at after processing its examples, which it then uses to make predictions on new data.
Turing Test – A test in which a human evaluator holds a text conversation with a machine and with a human and tries to tell which is which; if the evaluator cannot reliably tell them apart, the machine has passed.
13.7 Objective Questions
Fill in the blank: The test in which a human evaluator holds typed conversations with a hidden person and a hidden machine and tries to tell them apart is called the __________.
Machine intelligence is produced mainly through which of the following?
- Instinct and reflexes
- Programs, algorithms and data
- Emotional reasoning
- Random guessing
State whether true or false: A machine that passes the Turing Test has been shown to be self-aware.
In the Turing Test, what is the human evaluator expected to do?
- Write the program that the machine will run
- Teach the machine new skills during the conversation
- Decide which of the two hidden participants is the machine
- Time how fast the machine replies
Which of these tasks is machine intelligence better suited to than human intelligence?
- Settling an argument between two classmates
- Checking ten lakh electricity bills for an unusual pattern
- Deciding whether a joke is in poor taste
- Learning a new rule from a single example
Match the term in Column A with its meaning in Column B.
| Column A | Column B |
|---|---|
| a. Algorithm | i. A systematic tilt in the data or the method that makes results consistently unfair for some cases |
| b. Model | ii. A fixed, step-by-step method for carrying out a job |
| c. Bias | iii. Simple units arranged in layers, with a weight on every connection |
| d. Artificial neural network | iv. The set of internal settings a system arrives at after processing its examples |
- Name the layer of an artificial neural network at which numbers first enter the network.
13.8 Concept Questions
Define machine intelligence, and state one way in which it differs from human intelligence.
Why is the Turing Test carried out using typed text alone, rather than speech or a face-to-face meeting? Give the reason.
A child who has seen three dogs can recognise a dog it has never met before, while an image recognition system needs thousands of pictures. Explain what this difference tells us about how humans and machines learn.
A route-planning program sends a delivery truck down a road that every local driver knows is flooded during the monsoon. Explain what this shows about the reasoning done by a machine.
What would happen if a system that marks unusual bank transactions were trained only on records from one small town and then used across the whole country? Name the problem and explain how a person could catch it.
Explain how an artificial neural network borrows an idea from the human brain, and state two ways in which it is not the same as a brain.
13.9 Case-Based Questions
- A city bus depot uses a computer system that studies ticket records from past months and suggests how many buses to run on each route in each hour of the day. Every morning the depot manager reads the suggestions, compares them with what she knows about the day ahead, and then fixes the final schedule.
- Which part of this arrangement uses machine intelligence and which part uses human intelligence? Give a reason for each.
- The system was given ticket records only from weeks when schools were open. Explain what could go wrong when it is used during the holidays, and name the problem this is an example of.
- Give one reason why the final schedule should be fixed by the manager rather than by the system.
- The help page of an online shop answers typed questions within a few seconds. The replies are polite and use correct grammar, and most customers assume they are chatting with a shop employee. In fact, the replies are produced by a program.
- Which test from this chapter does this situation resemble, and what exactly does that test check?
- Most customers cannot tell that the replies come from a program. Can we conclude from this that the program understands their questions? Give a reason for your answer.
- Suggest one step the shop should take so that customers are not misled, and explain why it matters.