Minds and Machines: The Science of Intelligence

We use the word intelligence constantly, yet scientists still cannot fully agree on what it is. Is it problem-solving, learning, language, self-awareness, or some blend of all of these? Studying intelligence forces us to turn the tools of science on the thing doing the science, and that is one of the deepest frontiers we can explore.

Building machines that perform tasks once thought to need a mind gives us a new way to probe the question. Each time a machine matches or surpasses human performance, it sharpens rather than settles the puzzle, revealing which parts of thinking we understood and which we had merely taken for granted.

What do we mean by intelligence?

Intelligence is not a single measurable substance. It shows up as a family of abilities: perceiving, reasoning, learning from experience, planning, and adapting to new situations. Because these can come apart, a system can be brilliant at one and helpless at another, which is why any tidy one-line definition tends to leak.

Can machines think? The imitation game

In 1950, mathematician Alan Turing sidestepped the hard question of what thinking is. In his paper Computing Machinery and Intelligence, he proposed a practical test he called the imitation game, now known as the Turing test. If a human judge, communicating only through typed messages, cannot reliably tell a machine from a person, Turing argued, we have no clear basis for denying that the machine thinks. It replaces an unanswerable definition with an observable behaviour.

Narrow skill versus general understanding

A crucial distinction runs through the whole field. Narrow intelligence excels at one well-defined task; general intelligence flexibly handles many. A system can convincingly imitate conversation, or beat any human at a game, while having no understanding of the wider world. Impressive performance is not proof of a mind.

Imagine a judge chatting by text with two hidden partners, one human and one machine, trying to spot which is which. The machine bluffs by making a small arithmetic slip and pausing before replying, mimicking human imperfection. If the judge is fooled, the Turing test has been passed for that exchange. Notice what this shows and does not show: it measures whether behaviour is indistinguishable, not whether anything is genuinely felt or understood inside.

As early as the 1960s, a simple program called ELIZA imitated a therapist by rephrasing users' sentences as questions. Many people felt they were understood, yet ELIZA followed a handful of text rules and grasped nothing. It shows that appearing intelligent and being intelligent can come apart completely, and that our readiness to see minds where there are none is itself part of the science.

For about fifty years, biologists faced a stubborn challenge: predicting the three-dimensional shape a protein folds into from its chain of amino acids, a shape that governs what the protein does. In 2020, at the community assessment known as CASP14, DeepMind's AlphaFold system predicted many structures with accuracy competitive with slow, costly laboratory methods, a result experts called a genuine solution to a decades-old problem. AlphaFold learned from databases of known structures rather than being told the physics by hand. The work was recognised when Demis Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry, alongside David Baker. Yet AlphaFold predicts shapes without understanding biology the way a scientist does. It is a powerful narrow tool, and a striking example of intelligence as a scientific frontier rather than a settled question.

• That passing the Turing test proves a machine is conscious, when it only shows its behaviour is hard to distinguish from a human's.
• That being superb at one task, such as chess or protein prediction, means a system is generally intelligent.
• That AlphaFold understands biology, when it statistically predicts structures without insight into what proteins mean.
• That because we can build systems that act intelligently, we have therefore explained human intelligence; the two remain open questions.

Pick an ability you think requires real intelligence, then design a test a machine would have to pass to demonstrate it. Ask what a clever system could do to fake success, and what the loophole reveals about the ability you were trying to measure.

Think Like a Maester: When a machine impresses you, ask whether it truly understands the task or has merely found a shortcut that mimics understanding.

1. Why is intelligence difficult to capture in a single definition?
2. What did Alan Turing propose in his 1950 paper, and what does the test actually measure?
3. What is the difference between narrow and general intelligence?
4. What long-standing scientific problem did AlphaFold address, and what was notable about its 2020 result?
5. Why does AlphaFold's success not settle the question of what understanding is?

Intelligence is one of science's live frontiers: we use the word daily but cannot fully define what we mean. Alan Turing's 1950 imitation game reframed can machines think as a question about observable behaviour, while examples from ELIZA to modern systems show that acting intelligent and being intelligent can diverge. AlphaFold's 2020 breakthrough in protein folding, later honoured with a Nobel Prize, proves that machines can solve deep scientific problems, even as the nature of understanding, in minds or machines, remains genuinely open.

• How Science Changes Its Mind
• The Biggest Unsolved Mysteries