Home  /  Insights  /  Perspective

Can machines think? The road from Turing to AI agents

Perspective · 7 minute read · 27 October 2026

The history of AI is not a straight line. It is promises, winters and surprises, and for now it ends with a question about leadership.

1950

Can machines think?

It was October 1950, and at the University of Manchester the mathematician Alan Turing asked a question we are still trying to answer.

A few years earlier he had worked in the secret huts at Bletchley Park, helping to break the Nazis' Enigma code with electromechanical machines. Now he was looking ahead, and he opened a scientific paper in the journal Mind with the question: "Can machines think?"

Turing quickly saw that the question could not be answered philosophically, because what does it really mean to think? Instead he proposed a practical parlour game, which he called the Imitation Game, what we now know as the Turing test. A person sits behind a closed door and exchanges typed messages with a human and a machine. If the person cannot tell which is which, it is reasonable to say that the machine behaves intelligently.

?Closed doorAB

That fired the starting gun for one of humanity's most fascinating journeys. It is a story full of wild promises, brutal winters, unexpected breakthroughs and a realisation that the road to intelligence was very different from what the pioneers imagined.

Facts

Turing's paper "Computing Machinery and Intelligence" appeared in the journal Mind in October 1950 and opens with the question "Can machines think?" (Mind, 1 October 1950). Turing was then working at the University of Manchester, and during the war he had been a central part of the effort against Enigma at Bletchley Park (Wikipedia on Alan Turing).

1956

The optimistic summer at Dartmouth

In the summer of 1956, around ten researchers met at Dartmouth College in New Hampshire. The initiative came from John McCarthy together with Marvin Minsky, Nathaniel Rochester and Claude Shannon, and it was here that the term Artificial Intelligence was launched, in an application for research funding.

The mood was euphoric. The researchers sincerely believed that if a small group of bright minds worked together for two months over the summer, they could solve a significant part of the problem of getting machines to use language, form concepts and solve problems that until then had been reserved for humans.

They were wrong. It did not take two months, it took generations.

AISummer 19562 mo.Two months?

But the optimism was contagious. The first decades were shaped by what is known as symbolic AI, often called GOFAI (Good Old-Fashioned AI). The philosophy was simple: human logic can be written as rules. If you just feed a computer enough logical rules (if X, then do Y), it will eventually become clever.

Facts

The proposal for the summer project is dated 31 August 1955. It suggests a study of about two months with around ten participants, in the summer of 1956 at Dartmouth College (McCarthy, Minsky, Rochester and Shannon, 31 August 1955).

1966

ELIZA, the first conversation

In 1966, the researcher Joseph Weizenbaum of MIT presented a programme called ELIZA. It was a simple chatbot that imitated a psychotherapist by turning the user's own sentences into questions. When a user typed "I have problems with my mother", ELIZA replied "Tell me more about your family".

I am having troublewith my motherTell me moreabout your family

Many people were captivated. According to the accounts, Weizenbaum's secretary asked him to leave the room so that she could talk to the programme in private. Weizenbaum himself was astonished at how little it took to make people believe that a handful of simple code conditions held empathy.

Facts

ELIZA was described by Joseph Weizenbaum in Communications of the ACM in January 1966 (Wikipedia on ELIZA).

The 1970s and 1980s

The AI winters, when the money froze over

Reality hit in the 1970s. Symbolic AI ran into a wall it could not get past.

It turned out that computers were brilliant at things humans found hard, such as calculating complex equations or playing chess by strict rules. But the machines were hopeless at things a two-year-old masters naturally: recognising a face, understanding context, climbing a staircase or telling a dog from a muffin tin. The phenomenon later became known as Moravec's paradox.

The world is simply too messy for rigid rules. You cannot write enough "if-then" rules to describe what a cat looks like from every imaginable angle and in different light.

fundingtime

When the promised miracles failed to arrive, governments and military investors closed their wallets. From about 1974 to 1980, and again from 1987, two hard periods set in, which the research world now calls the AI winters. Many researchers began to call their work something else, such as machine learning, because the words artificial intelligence had become linked to promises that were not kept.

Facts

In the United Kingdom, the Lighthill report criticised AI in 1973 for failing to meet its goals, and it led to an almost total halt in research there. DARPA also cut its support for AI in the early and mid-1970s and again in 1987 to 1988 (Wikipedia on AI winters). Moravec's paradox was formulated in the 1980s (Wikipedia on Moravec's paradox).

Up to 2012

Learning like a child, not from a rulebook

While most people had given up on thinking machines, a small group of stubborn researchers, led by people such as Geoffrey Hinton, Yann LeCun and Yoshua Bengio, worked on a completely different idea: what if we stopped programming rules and instead built systems that learned from data, the way a brain learns from experience?

The idea was inspired by biological brains. By connecting thousands of small digital "neurons" in layers, so-called neural networks, the machine could adjust its own connections every time it got something wrong.

In the 1980s and 1990s the idea was beautiful on paper but hard to use in practice. Computers were simply too slow, and there was too little digital data to feed them.

Then two things happened around the turn of the millennium:

  1. The internet exploded. Suddenly the world was uploading billions of images, articles, forums and videos. The raw material, data, had been found.
  2. Graphics cards (GPUs) became everyday hardware. Originally developed to render 3D graphics in computer games, graphics cards turned out, by chance, to be excellent at the kind of parallel matrix calculations that neural networks require.
imagecatdata

The breakthrough came in 2012. In an annual image recognition competition, ImageNet, a deep neural network called AlexNet showed markedly better results than earlier methods, and the error rate fell sharply. Deep learning broke through, and the AI winter thawed.

Facts

AlexNet was described by Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton in the 2012 paper "ImageNet Classification with Deep Convolutional Neural Networks". It reports markedly better results than earlier methods (NIPS, 2012).

2016

When the machine surprised its makers

Now things moved fast. AI moved from university labs to the data centres of the tech giants.

In March 2016, Google DeepMind set out to tackle the board game Go. Chess had been conquered by IBM's Deep Blue in 1997 through raw computing power, but Go was regarded as the ultimate game of human intuition. There are more possible positions on a Go board than there are atoms in the known universe, so you cannot calculate your way to a win, you have to "feel" the position.

The programme AlphaGo met one of the world's best players, Lee Sedol, in Seoul. In game two, move 37, AlphaGo placed a stone somewhere on the board that few human masters would have chosen. According to DeepMind, a human had only about a one in 10,000 chance of playing that move, and Lee Sedol took an unusually long time to respond.

move 37

The move looked unconventional, but it was not a mistake. It was a creative, unorthodox strategic move, devised by an algorithm that had played millions of games against itself. AlphaGo won the match 4-1. The machine no longer followed human logic, it found its own ways.

Facts

The match was played from 9 to 15 March 2016 in Seoul and ended 4-1 to AlphaGo (Wikipedia on AlphaGo versus Lee Sedol). Go has about 10170 possible positions, more than there are atoms in the known universe (Google DeepMind). Deep Blue beat Kasparov in May 1997 by brute-force calculation of about 200 million chess positions per second (IBM, May 1997).

2017

The Transformer's triumph and the great leap

Despite the victory in Go, the systems were still specialised. A programme that played Go could not tell a taxi from a fire engine, and it certainly could not hold a conversation.

The answer came in 2017 from eight researchers, most of them at Google, who published a research paper with a title that would go down in history: "Attention Is All You Need".

They introduced a new architecture called the Transformer. Earlier language models read text word by word and could quickly lose the thread in long passages. The Transformer let the model look at all the words in a sentence at once and work out how they relate to each other, through a mechanism called attention.

thebankbythe riverattention

Suddenly machines became far better at understanding context. The model could tell the difference between "a bank by the river", "a knock on the door" and "money in the bank".

Facts

The paper "Attention Is All You Need" was posted on arXiv on 12 June 2017 by eight authors. It introduces the Transformer, a model based entirely on attention (arXiv, 12 June 2017).

2022

ChatGPT, when AI got a voice

When the research laboratory OpenAI took the Transformer architecture and began feeding it astronomical amounts of text from across the internet, combined with a method called RLHF (Reinforcement Learning from Human Feedback, where people reward the model for good, helpful answers), a phase shift occurred.

On 30 November 2022, OpenAI opened up ChatGPT.

Until then, AI had been an invisible engine behind Google's search results, Netflix's recommendations or the bank's credit ratings. Suddenly it had a voice, a tone and an interface that anyone could talk to. According to an estimate from UBS, the service reached 100 million monthly users in about two months, faster than both TikTok and Instagram.

2 months100 m users

Facts

OpenAI introduced ChatGPT on 30 November 2022 and says it was trained with RLHF (OpenAI, 30 November 2022). In February 2023, UBS estimated that ChatGPT reached 100 million monthly users in January, about two months after launch. TikTok took about nine months and Instagram about two and a half years (Euronews, February 2023).

Today

From talking text to autonomous agents

Today we have entered a new era. The question is no longer just whether a model can write a poem, summarise a report or help with homework.

The focus has shifted from static language models to agentic AI, systems that do not just answer questions, but can reason, plan several steps ahead, use tools, correct their own mistakes and carry out whole workflows in the real world. The machine has gone from being a reference book to being a digital actor.

PlanUse a toolCheckFix

That brings us back to Alan Turing and his question in Manchester in 1950.

Turing foresaw that in about fifty years machines would be able to play his imitation game so convincingly that an average judge would have no more than a 70 per cent chance of guessing right after five minutes of conversation. The timeline was perhaps a little optimistic, but the direction was right. The challenge for us today is no longer to prove whether machines can pass Turing's test. The challenge is how we, as humans, lead, regulate and work with an intelligence we have created ourselves, but which no longer thinks quite like us.

Facts

In the 1950 paper, Turing estimates that in about fifty years an average judge would have no more than a 70 per cent chance of identifying the machine correctly after five minutes of questioning (Mind, 1 October 1950).

Our take

We believe the next chapter is less about what the machines can do, and more about who sets the direction and is accountable.

Questions

We do not yet know how much an agent in an ordinary company should be allowed to do without a human watching.

What you can take with you

The question is no longer whether machines can pass Turing's test, but how we lead and work with an intelligence we have created ourselves.

Sources

Adam Carbon and Eva SeliciumPseudonyms for a human and a machine, writing together.

Get it in your inbox

Write to us, and we will put you on the list. Tell us what you would most like to hear about: perspectives, fables or news from Coga. You unsubscribe by writing to us again.

Write to raadgiver@coga.dk